U.S. patent application number 15/836689 was filed with the patent office on 2018-09-13 for system for rule-based insurance underwriting suitable for use by an automated system.
The applicant listed for this patent is Pacific Life Insurance Company. Invention is credited to Piero Patrone Bonissone, William Michael Durham, Richard Paul Messmer, Marc Pavese, Diane Marie Russell, Dan Yang.
Application Number | 20180260906 15/836689 |
Document ID | / |
Family ID | 26867212 |
Filed Date | 2018-09-13 |
United States Patent
Application |
20180260906 |
Kind Code |
A1 |
Bonissone; Piero Patrone ;
et al. |
September 13, 2018 |
SYSTEM FOR RULE-BASED INSURANCE UNDERWRITING SUITABLE FOR USE BY AN
AUTOMATED SYSTEM
Abstract
A system for at least a partial underwriting of insurance
policies is described. Various rules are created, along with a
degree of satisfaction for each rule. Rules may be directed toward
various insurance underwriting components (e.g., cholesterol
levels, blood pressure, etc.). Based on the degree of satisfaction
for each rule, a component may be assigned to a category. Based on
the category for each component, the insurance application may be
assigned an underwriting category.
Inventors: |
Bonissone; Piero Patrone;
(Schenectady, NY) ; Messmer; Richard Paul;
(Rexford, NY) ; Durham; William Michael; (Concord,
VA) ; Yang; Dan; (Westborough, MA) ; Pavese;
Marc; (Saratoga Springs, NY) ; Russell; Diane
Marie; (Lynchburg, VA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Pacific Life Insurance Company |
Newport Beach |
CA |
US |
|
|
Family ID: |
26867212 |
Appl. No.: |
15/836689 |
Filed: |
December 8, 2017 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
14444529 |
Jul 28, 2014 |
|
|
|
15836689 |
|
|
|
|
10171575 |
Jun 17, 2002 |
8793146 |
|
|
14444529 |
|
|
|
|
60343240 |
Dec 31, 2001 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06Q 10/10 20130101;
G06Q 40/08 20130101 |
International
Class: |
G06Q 40/08 20060101
G06Q040/08; G06Q 10/10 20060101 G06Q010/10 |
Claims
1. A system for underwriting an insurance application comprising: a
receiver for receiving a request to underwrite an insurance
application, where the request includes information about at least
one application component; an evaluation module for evaluating the
at least one application component based on at least one rule
associated with the at least one application component; and an
assignment module for: a) assigning a measurement to the at least
one application component; and b) assigning the at least one
application component to a specific component category out of a
plurality of component categories based on the assigned
measurement.
2. The system according to claim 1, wherein the assignment module
further assigns the insurance application to a specific application
category out of a plurality of application categories, where the
one specific application category is based at least in part on the
specific component category.
3. The system according to claim 2, wherein the plurality of
component categories are identical to the plurality of application
categories.
4. The system according to claim 1, wherein the evaluation of the
at least one application component includes: defining the at least
one rule; assigning the at least one rule to the at least one
application component; and defining a plurality of parameters
associated with the at least one rule, wherein each parameter
corresponds to a particular category of the plurality of component
categories.
5. The system according to claim 4, wherein the assignment module
further assigns the insurance application to a specific application
category out of a plurality of application categories, wherein the
specific application category is based at least in part on the
component category.
6. The system according to claim 5, wherein the plurality of
component categories are identical to the plurality of application
categories.
7. The system according to claim 1, wherein the at least one rule
is a fuzzy logic rule.
8. A system for underwriting an insurance application comprising: a
receiver for receiving a request to underwrite an insurance
application, wherein the insurance application includes a plurality
of application components; an evaluation module for evaluating the
plurality of application components based on a plurality of rules,
wherein each of the plurality of rules is associated with each of
the plurality of application components; and an assignment module
for: a) assigning a measurement to each of the plurality of
application components; and b) assigning each of the plurality of
application components to a specific component category out of a
plurality of component categories based on the assigned
measurement.
9. The system according to claim 8, wherein the assignment module
further assigns the insurance application to a specific application
category out of a plurality of application categories, wherein the
specific application category is based at least in part on the
component categories for each of the plurality of component
categories.
10. The system according to claim 9, wherein the plurality of
component categories are identical to the plurality of application
categories.
11. The system according to claim 8, wherein the evaluation of the
plurality of application components includes: defining the
plurality of rules; assigning each of the plurality of rules to one
of the plurality of components; and defining a plurality of
parameters associated with each of the plurality of rules, wherein
each parameter corresponds to a particular category of the
plurality of component categories.
12. The system according to claim 11, wherein the assignment module
further assigns the insurance application to a specific application
category out of a plurality of application categories, wherein the
specific application category is based at least in part on the
component categories for each of the plurality of component
categories.
13. The system according to claim 12, wherein the plurality of
component categories are identical to the plurality of application
categories.
14. The system according to claim 8, wherein the plurality of rules
are fuzzy logic rules.
15. A system for underwriting an insurance application comprising:
a receiver for receiving a request to underwrite an insurance
application for a customer, wherein the insurance application
includes a plurality of application components; an evaluation
module for evaluating the plurality of application components based
on a plurality of rules, wherein each of the plurality of rules is
associated with each of the plurality of application components; an
assignment module for: a) assigning a measurement to each of the
plurality of application components; b) assigning each of the
plurality of application components to a specific component
category out of a plurality of component categories based on the
assigned measurement; and c) assigning the insurance application to
a specific application category out of a plurality of application
categories, wherein the specific application category is based at
least in part on the specific component category; and an issuance
module for issuing the insurance application to the customer.
16. The system according to claim 15, wherein the plurality of
component categories are identical to the plurality of application
categories.
17. The system according to claim 15, wherein the evaluation of the
at least one application component includes: defining the at least
one rule; assigning the at least one rule to the at least one
component; and defining a plurality of parameters associated with
the at least one rule, wherein each parameter corresponds to a
particular category of the plurality of component categories.
18. The system according to claim 17, wherein the assignment module
further assigns the insurance application to a specific application
category out of a plurality of application categories, wherein the
specific application category is based at least in part on the
component category.
19. The system according to claim 18, wherein the plurality of
component categories are identical to the plurality of application
categories.
20. The system according to claim 15, wherein the plurality of
rules are fuzzy logic rules.
21-45. (canceled)
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] Any and all applications for which a foreign or domestic
priority claim is identified in the Application Data Sheet, or any
correction thereto, are hereby incorporated by reference in this
application under 37 CFR 1.57.
BACKGROUND OF THE INVENTION
[0002] The present invention relates to a system for underwriting
insurance applications, and more particularly to a system for
underwriting insurance applications based on a flexible fuzzy rule
logic based system.
[0003] A trained individual or individuals traditionally perform
insurance underwriting. A given application for insurance (also
referred to as an "insurance application") may be compared against
a plurality of underwriting standards set by an insurance company.
The insurance application may be classified into one of a plurality
of risk categories available for a type of insurance coverage
requested by an applicant. The risk categories then affect a
premium paid by the applicant, e.g., the higher the risk category,
the higher the premium. A decision to accept or reject the
application for insurance may also be part of this risk
classification, as risks above a certain tolerance level set by the
insurance company may simply be rejected.
[0004] There can be a large amount of variability in the insurance
underwriting process when performed by individual underwriters.
Typically, underwriting standards cannot cover all possible cases
and variations of an application for insurance. The underwriting
standards may even be self-contradictory or ambiguous, leading to
uncertain application of the standards. The subject judgment of the
underwriter will almost always play a role in the process.
Variation in factors such as underwriter training and experience,
and a multitude of other effects can cause different underwriters
to issue different, inconsistent decisions. Sometimes these
decisions can be in disagreement with the established underwriting
standards of the insurance company, while sometimes they can fall
into a "gray area" not explicitly covered by the underwriting
standards.
[0005] Further, there may be an occasion in which an underwriter's
decision could still be considered correct, even if it disagrees
with the written underwriting standards. This situation can be
caused when the underwriter uses his/her own experience to
determine whether the underwriting standards may or should be
interpreted and/or adjusted. Different underwriters may make
different determinations about when these adjustments are allowed,
as they might apply stricter or more liberal interpretations of the
underwriting standards. Thus, the judgment of experienced
underwriters may be in conflict with the desire to consistently
apply the underwriting standards.
[0006] Most of the key information required for automated insurance
underwriting is structured and standardized. However, some sources
of information may be non-standard or not amenable to
standardization. By way of example, an attending physician
statement ("APS") may be almost as unique as each individual
physician. However, a significant fraction of applications may
require the use of one or more APS due to the presence of medical
impairments, age of applicants, or other factors. Without such key
information, the application underwriting process cannot be
automated for these cases.
[0007] Conventional methods for dealing with some of the problems
described above have included having human underwriters directly
reading the APS. However, an APS document can be as long as several
tens of pages. Therefore, the manual reading process, combined with
note-taking and consulting other information, such as an
underwriting manual or the like, can greatly extend the cycle-time
for each application processed, increase underwriter variability,
and limit capacity by preventing the automation of the decision
process.
SUMMARY OF THE INVENTION
[0008] In an exemplary embodiment of the invention, a system for
underwriting an insurance application comprises a receiver for
receiving a request to underwrite an insurance application, where
the request includes information about at least one application
component, an evaluation module for evaluating the at least one
application component based on at least one rule associated with
the at least one application component, and an assignment module
for assigning a measurement to the at least one application
component, and assigning the at least one application component to
a specific component category out of a plurality of component
categories based on the assigned measurement.
[0009] Another exemplary embodiment provides a system for
underwriting an insurance application comprising a receiver for
receiving a request to underwrite an insurance application, wherein
the insurance application includes a plurality of application
components, an evaluation module for evaluating the plurality of
application components based on a plurality of rules, wherein each
of the plurality of rules is associated with each of the plurality
of application components, and an assignment module for assigning a
measurement to each of the plurality of application components, and
assigning each of the plurality of application components to a
specific component category out of a plurality of component
categories based on the assigned measurement.
[0010] In a further embodiment of the invention, a system for
underwriting an insurance application comprises a receiver for
receiving a request to underwrite an insurance application for a
customer, wherein the insurance application includes a plurality of
application components, and an evaluation module for evaluating the
plurality of application components based on a plurality of rules,
wherein each of the plurality of rules is associated with each of
the plurality of application components. In addition, the system
also comprises an assignment module for assigning a measurement lo
each of the plurality of application components, assigning each of
the plurality of application components to a specific component
category out of a plurality of component categories based on the
assigned measurement and assigning the insurance application to a
specific application category out of a plurality of application
categories, wherein the specific application category is based at
least in part on the specific component category, and an issuance
module for issuing the insurance application to the customer.
[0011] A further exemplary embodiment of the invention provides a
system for underwriting an insurance application comprising means
for receiving a request to underwrite an insurance application,
where the request includes information about at least one
application component, means for evaluating the at least one
application component based on at least one rule associated with
the at least one application component, means for assigning a
measurement to the at least one application component, and means
for assigning the at least one application component to a specific
component category out of a plurality of component categories based
on the assigned measurement.
[0012] In another embodiment of the invention, a system for
underwriting an insurance application is provided. The system
comprises means for receiving a request to underwrite an insurance
application, wherein the insurance application includes a plurality
of application components, means for evaluating the plurality of
application components based on a plurality of rules, wherein each
of the plurality of rules is associated with each of the plurality
of application components, means for assigning a measurement to
each of the plurality of application components, and means for
assigning each of the plurality of application components to a
specific component category out of a plurality of component
categories based on the assigned measurement.
[0013] An additional exemplary embodiment of the invention provides
a system for underwriting an insurance application comprising means
for receiving a request to underwrite an insurance application for
a customer, wherein the insurance application includes a plurality
of application components and means for evaluating the plurality of
application components based on a plurality of rules, wherein each
of the plurality of rules is associated with each of the plurality
of application components. The invention further provides that the
system comprises means for assigning a measurement to each of the
plurality of application components, means for assigning each of
the plurality of application components to a specific component
category out of a plurality of component categories based on the
assigned measurement, means for assigning the insurance application
to a specific application category out of a plurality of
application categories, wherein the specific application category
is based at least in part on the specific component category and
means for issuing the insurance application to the customer.
[0014] A further embodiment of the invention provides a system for
underwriting a transaction orientated process comprising a receiver
for receiving a request to underwrite a transaction, where the
request includes information about at least one transaction
component, an evaluation module for evaluating the at least one
transaction component based on at least one rule associated with
the at least one transaction component, wherein the at least one
rule is a fuzzy logic rule, and an assignment module for assigning
a measurement to the at least one transaction component, assigning
the at least one transaction component to a specific component
category out of a plurality of component categories based on the
assigned measurement, and assigning the transaction lo a specific
transaction category out of a plurality of transaction categories,
where the one specific transaction category is based at least in
part on the specific component category.
BRIEF DESCRIPTION OF THE DRAWINGS
[0015] FIG. 1 is a graph illustrating a fuzzy (or soft) constraint,
a function defining for each value of the abscissa the degree of
satisfaction for a fuzzy rule, according to an embodiment of the
invention.
[0016] FIG. 2 is a graph illustrating the measurements based on the
degree of satisfaction for a collection of fuzzy rules, according
to an embodiment of the invention.
[0017] FIG. 3 is a schematic representation of an object-oriented
system to determine the degree of satisfaction for a collection of
fuzzy rules, according to an embodiment of the invention.
[0018] FIG. 4 is a flowchart illustrating steps performed in a
process for underwriting an insurance application using fuzzy logic
according to an embodiment of the invention.
[0019] FIG. 5 is a flowchart illustrating steps for an inference
cycle according to an embodiment of the invention.
[0020] FIG. 6 is a graph illustrating a fuzzy (or soft) constraint,
a function defining for each value of the abscissa the degree of
satisfaction for a rule comparing similar cases, according to an
embodiment of the invention.
[0021] FIG. 7 is a graph illustrating the core of a fuzzy (or soft)
constraint, according to an embodiment of the invention.
[0022] FIG. 8 is a graph illustrating the support of a fuzzy (or
soft) constraint, according to an embodiment of the invention.
[0023] FIG. 9 is a graph illustrating the rate class histogram
derived from a set of retrieved cases, according to an embodiment
of the invention.
[0024] FIG. 10 is a chart illustrating the distribution of
similarity measures for a set of retrieved cases, according to an
embodiment of the invention.
[0025] FIG. 11 is a table illustrating a linear aggregation of rate
classes, according to an embodiment of the invention.
[0026] FIG. 12 is a flowchart illustrating the steps performed in a
process for determining the degree of confidence of an underwriting
decision based on similar cases, according to an embodiment of the
invention.
[0027] FIG. 13 is a process map illustrating a decision flow,
according to an embodiment of the invention.
[0028] FIG. 14 illustrates a comparison matrix, according to an
embodiment of the invention.
[0029] FIG. 15 illustrates a distribution of classification
distances for each bin containing a range of retrieved cases,
according to an embodiment of the invention.
[0030] FIG. 16 illustrates a distribution of normalized percentage
of classification distances for each bin containing a range of
retrieved cases, according to an embodiment of the invention.
[0031] FIG. 17 illustrates a distribution of correct classification
for each bin containing a range of retrieved cases, according to an
embodiment of the invention.
[0032] FIG. 18 illustrates a distribution of a performance function
for each bin containing a range of retrieved cases, according to an
embodiment of the invention.
[0033] FIG. 19 illustrates a distribution of a performance function
for each bin containing a range of retrieved cases, after removing
negative numbers and normalizing the values between 0 and 1,
according to an embodiment of the invention.
[0034] FIG. 20 illustrates results of a plot of the preference
function (derived from FIG. 19) according to an embodiment of the
invention.
[0035] FIG. 21 illustrates a computation of coverage and accuracy
according to an embodiment of the invention.
[0036] FIG. 22 is a schematic representation of a system for
underwriting according to an embodiment of the invention.
[0037] FIG. 23 a flowchart illustrating the steps performed for
executing and manipulating a summarization tool according to an
embodiment of the invention.
[0038] FIG. 24 illustrates a graphic user interface for a
summarization tool for a general form according to an embodiment of
the invention.
[0039] FIG. 25 illustrates a graphic user interface for a
summarization tool for a condition-specific form according to an
embodiment of the invention.
[0040] FIG. 26 illustrates an optimization process according to an
embodiment of the invention.
[0041] FIG. 27 illustrates an example of an encoded population at a
given generation according to an embodiment of the invention.
[0042] FIG. 28 illustrates a process schematic for an evaluation
system according to an embodiment of the invention.
[0043] FIG. 29 illustrates an example of the mechanics of an
evolutionary process according to an embodiment of the
invention.
[0044] FIG. 30 is a graph illustrating a linear penalty function
used in the evaluation of the accuracy of the CBE, according to an
embodiment of the invention.
[0045] FIG. 31 is a graph illustrating a nonlinear penalty function
used in the evaluation of the accuracy of the CBE, according to an
embodiment of the invention.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT
[0046] Reference will now be made in detail to the present
preferred embodiments of the invention, examples of which are
illustrated in the accompanying drawings in which like reference
characters refer to corresponding elements.
Rules Based Reasoning
[0047] As stated above, a process and system is provided for
insurance underwriting which is able to incorporate all of the
rules in the underwriting standards of a company, while being
robust, accurate, and reliable. According to an embodiment of the
invention, the process and system provided may be suitable for
automation. Such a process and system may be flexible enough to
adjust the underwriting standards when appropriate. As mentioned
above, each individual underwriter may have his/her own set of
interpretations of underwriting standards about when one or more
adjustments should occur. According to an embodiment of the present
invention, rules may be incorporated while still allowing for
adjustment using a fuzzy logic-based system. A fuzzy logic-based
system may be described as a formal system of logic in which the
traditional binary truth-values "true" and "false" are replaced by
real numbers on a scale from 0 to 1. These numbers are absolute
values that represent intermediate truth-values for answers to
questions that do not have simple true or false, or yes or no
answers. In standard binary logic, a given rule is either satisfied
(with a degree of satisfaction of 1), or not (with a degree of
satisfaction of 0), creating a sharp boundary between the two
possible degrees of satisfaction. With fuzzy logic, a given rule
may be assigned a "partial degree of satisfaction", a number
between 1 and 0, in some boundary region between a "definite yes",
and a "definite no" for the satisfaction of a given rule. Each rule
will be composed by a conjunction of conditions. Each condition
will be represented by a fuzzy set A(x), which can be interpreted
as a degree of preference induced by a value x for satisfying a
condition A. An inference engine determines a degree of
satisfaction of each condition and an overall degree of
satisfaction of a given rule.
[0048] For the purposes of illustration, imagine that a
hypothetical life insurance company has a plurality of risk
categories, which are identified as "cat1", "cat2", "cat3", and
"cat4." In this example, a rating of cat I is a best or low risk,
while cat4 is considered a worst or high risk. An applicant for an
insurance policy would be rejected if he/she fails to be placed in
any category. An example of a type of rule laid out in a set of
underwriting guidelines could be, "The applicant may not be in cat1
if his/her cholesterol value is higher than X L". Similarly, a
cholesterol value of X2 could be a cutoff for cat2, and so on.
However, it is possible that a cholesterol reading of one point
over X I may not in practice disqualify the applicant from the cat1
rating, if all of the other rules are satisfied for cat I. It may
be that readings of one point over X I are still allowable, and so
on. To define a fuzzy rule, two parameters, X1a and X1b may be
needed. When the applicant's cholesterol is below X1a, a fuzzy rule
may be fully satisfied (e.g., a degree of satisfaction of 1). By
way of present example, X1 from the above may be used as X1a. A
parameter X1b may be a cutoff above which the fuzzy rule is fully
unsatisfied (e.g., a degree of satisfaction of 0). For example, it
may be determined from experienced underwriters of the insurance
company that under no circumstances can the applicant get the cat1
rating if his/her cholesterol is above 190 (XI) by more than four
points. In that situation, the fuzzy rule may use X1a=X1, that is
190, and X1b=X1a+4, that is 194. Other settings may be used. X1a
and X1b are parameters of the model. To obtain the partial degree
of satisfaction when the cholesterol value falls within the range
[X1a, X1b], a continuous switching function may be used, which
interpolates between the values 1 and 0. The simplest such function
is a straight line, as disclosed in FIG. 1, but other forms of
interpolation may also be used.
[0049] Turning to cat2, cat3, and cat4, there may be a different
cholesterol rule for each category, which states that the applicant
may not be placed in that category if his/her cholesterol is higher
than X2, X3, or X4, respectively. The same procedures may be used,
turning each rule into a fuzzy logic rule by assigning high and low
cutoff values (e.g., X2a, X2b; X3a, X3b; X4a, X4b). Thus, by way of
continuing the example, cat2 may be associated with a fuzzy rule
that uses X2a=X2 and X2b=X2+4, where X2=195 (for cat2). In addition
X3a=X3 and X3b=X3+4, where X3=200 (for cat3), and X4a=X4 and
X4b=X4, where X4=205 (for cat4). Other parameters also may be used.
Similarly, one would proceed through each rule in the underwriting
guidelines, allowing for fuzzy partial degrees of satisfaction. In
the present invention, each piece of data may be judged many times
on the basis of each rule.
[0050] Once each fuzzy rule in the rule set has been applied, a
decision is made to which category the applicant belongs. For each
risk category, there may be a subset of rules that apply to that
category. In order to judge whether the applicant is eligible for
the given category, some number of aggregation criteria may be
applied. To be concrete, using the above hypothetical case, take
the subset of all rules that apply to cat1. There will be a fuzzy
degree of satisfaction for every rule, where the set of degrees of
satisfaction is called {DS-cat1}. According to an embodiment of the
invention, if any of the degrees of satisfaction are zero, then the
applicant may be ruled out of cat I. Thus, one of the aggregation
criteria may be, "reject from cat 1 if MIN({DS-cat I})<=A1,"
where A1 is a chosen constant, and the notation MIN( . . . )
denotes selection of the smallest value out of the set. One choice
for A1 may be 0.5, but other choices may be used. By way of another
example, the choice, A1=0.7 may also be used. Again, the constant
A1 may be considered as a parameter of the model, which may be
determined.
[0051] As another aggregation rule, by way of example, if very many
of the rules have partial degrees of satisfaction of 0.9, then too
much adjusting may be occurring, and the applicant may be ruled out
of cat I, even though the aggregation rule, MIN({DS-cat1})<=A1,
may not be satisfied. The missing score (MS) is determined from the
degree of satisfaction (DS) by MS=1-DS. If a given fuzzy rule has
DS=0.9, then it would have a missing score of 0.1. The aggregation
criterion for this case might take the form, "reject from cat1 if
SUM({MS-cat1})>=A2," where A2 is a different chosen constant,
the notation, SUM( . . . ) denotes summation of all the elements of
the set, and {MS-cat1} is the set of "missing scores" for each
rule. The aggregation criteria above may use the sum of all of the
missing scores for the cat I rules as a measure to determine when
too much adjusting has been done, comparing that with the constant
A2. The measure defined above (SUM {MS-cat1}) may be interpreted as
a measure proportional to the difference between the degree of
complete satisfaction of all rules and the average degree of
satisfaction of each rule (DS-cat1). It is understood in this
invention that there may be any number of different kinds of
aggregation criteria, of which the above two are only specific
examples.
[0052] In a further step, the results of applying the aggregation
criteria to the set of rules relating to each category may be
compared. A result according to one example may be that the
applicant is ruled out of cat1 and cat2, but not from cat3 or cat4.
In that case, assuming that the insurance company's policy was to
place applicants in the best possible risk category, the final
decision would be to place the applicant in cat3. Other results may
also be obtained.
[0053] As stated above, this fuzzy logic system may have many
parameters that may be freely chosen. It should be noted that the
fuzzy logic system may extend and therefore subsume a conventional
(Boolean) logic system. By setting the fuzzy logic system
parameters to have only crisp thresholds (in which the core value
is equal to the support) the Boolean rules may be represented as a
case of fuzzy rules. Those parameters may be fit to reproduce a
given set of decisions, or set by management in order to achieve
certain results. By way of one example, a large set of cases may be
provided by the insurance company as a standard to be reproduced as
closely as possible. Preferably in such an example, there may be
man y cases, thereby minimizing the error between the fuzzy rules
model and the supplied cases. Optimization techniques such as
logistic regression, genetic algorithms, Monte Carlo, etc., also
may be used to find an optimal set of parameters. By way of another
example, some of the fuzzy rules may be determined directly by the
management of the insurance company. This may be done through
knowledge engineering sessions with experienced underwriters, by
actuaries acting on statistical information related to the risk
being insured or by other manners. In fact, when considering
maintenance of the system, initial parameters may be chosen using
optimization versus a set of cases, while at a future time, as
actuarial knowledge changes, these facts may be used to directly
adjust the parameters of the fuzzy rules. New fuzzy rules may be
added, or aggregation rules may change. The fuzzy logic system can
be kept current, allowing the insurance company to implement
changes quickly and with zero variability, thereby providing a
process and system that is flexible.
[0054] According to one embodiment of the invention, the fuzzy
logic parameters may be entered into a spreadsheet to evaluate the
fuzzy rules for one case at a time. This may be essentially
equivalent to implementation in a manual processing type
environment. FIG. 2 is a graphical representation illustrating a
plurality of measurements based on a degree of satisfaction for a
rule. A graphical user interface (GUI) 200 displays the degree of
satisfaction for one or more rules. GUI 200 includes a standard
toolbar 202, which may enable a user to manipulate the information
in known manners (e.g., printing, cutting, copying, pasting, etc.).
According to an embodiment of the invention, GUI may be presented
over a network using a browser application such as Internet
Explorer.RTM., Netscape Navigator.RTM., etc. An address bar 204 may
enable the user to indicate what portion is displayed. A chart 206
displays various insurance decision components and how each
insurance decision component satisfies its associated rule. A
plurality of columns 208 illustrates a plurality of categories for
each decision component, as well as a plurality of parameters for
each decision component. A column 210 identifies the actual
parameters of the potential applicant for insurance and a plurality
of columns 212 illustrate a degree of satisfaction of each rule. By
way of example, a row 214 is labeled BP (Sys), corresponding to a
systolic blood pressure rule. To receive the Best or Preferred
category classification, the applicant must have a systolic blood
pressure score (score) between 140 and 150. To receive a Select
category classification, the applicant must have a score between
150 and 155, while a score of 155 or more receives a "Standard
Plus" or St. Plus category classification. In this example, the
applicant has a score of 151. The columns 212 show zero
satisfaction of the rule for the Best and Preferred category
classifications. Additionally, FIG. 2 shows that the applicant
slightly missed satisfaction for the Select category, and Perfect
Constraint Satisfaction for the St. Plus Category.
[0055] In another example, a row 216 is labeled BP (Dia.),
corresponding to a diastolic blood pressure rule. To receive a Best
category classification, the applicant must have a diastolic blood
pressure score (score) between 85 and 90, between 90 and 95 for a
Preferred category classification, between 90 and 95 for the Select
category classification, and between 95 and 100 for the St. Plus
category classification. Here, the applicant has a score of 70,
resulting in Perfect Constraint Satisfaction in all of the columns
212.
[0056] By way of a further example, a row 218 is labeled Nicotine,
where a score between 4 and 5 receives the Best category
classification, a score between 2.5 and 3 receives the Preferred
category classification, a score between 1.5 and 2 receives the
Select category classification, and a score between 0.7 and 1
receives the St. Plus category classification. In this example, the
applicant has a score of 4.2. Thus, a score of "Mostly Missing" is
indicated under the Best category of a column 212, while a score of
Perfect Constraint Satisfaction is indicated for all others. GUI
200 presents a submit button 220 to enable the user to accept a
decision and submit it to a database. Alternatively, the user may
decide not to accept the decision. The user may activate a next
button 222 to record his/her decision. Other methods for display
may also be used. According to another embodiment of the invention,
the rules may be encoded into a Java-based computer code, which can
query a database to obtain the case parameters, and write its
decision in the database as well. The object model of the java
implementation is illustrated in FIG. 3. This java implementation
may be suitable for batch processing, or for use m a fully
automated underwriting environment. According to an embodiment of
the invention, a rule engine (class RuleEngine) 302 may be the
control of the system. The decision components of rule engine 302
may be composed of several rules (class Rule) 304, several
aggregations (class Aggregation) 306 and zero or one decision
post-processors (class DecisionPost-Processor) 308. A Rule object
304 may represent the fuzzy logic for one or a group of variables.
Each rule is further composed of a number of rateclasses (class
Rateclass) 310. A Rateclass object 310 defines the rules for a
specific rateclass. According to an embodiment of the invention, a
Rateclass object 310 may comprise two parts. The first is
pre-processing (class Preprocessor) 312, which may process multiple
inputs to form one output. The second is post-processing (class
Postprocessor) 314, which may take the result of the
pre-processing, feed it to a fuzzy function and get a fuzzy score.
Some of the rules may be conditional, such as the variable blood
pressure systolic, where the thresholds vary depending on the age
of the applicant. Class Condition 316 may represent such a
condition, if there is any. Classes FixedScore 318, Minimal and
Maximal may define some special preprocessing functions, and class
Linear 320 may define the general linear fuzzy function as
illustrated in FIG. 1.
[0057] According to an embodiment of the invention, there may be
two phases at runtime for rule engine 302. The first phase may be
initialization. In the process, the rule definition file in XML
format configures the rule engine. All the rule engine parameters
are defined in the process, for example, number of rules, the fuzzy
thresholds, pre and post processing and aggregation operation
(including class Intersection 322 and Sum Missing 324) and class
ThresholdLevel 326. The second phase may be scoring. After correct
initialization, the fireEngine method in rule engine 302 may take
an input parameter--an instance of class Case 328 containing all
the required variable values, and output an instance of class
Result 330, which encapsulates all the decision results, including
rateclass placement, the fuzzy scores for each variable and each
rateclass, and the aggregation scores. Class ResultLogger 332 may
log the output. Other object models for a java implementation may
also be used.
[0058] FIG. 4 is a flowchart illustrating the steps performed in a
process for underwriting an insurance application using fuzzy logic
rules according to an embodiment of the invention. At step 400, a
request to underwrite an insurance application may be received. The
request to underwrite may come directly from a consumer (e.g., the
person being insured), an insurance agent or another person. The
request to underwrite comprises information about one or more
components of the insurance application. According to an embodiment
of the invention, the components may include the various
characteristics associated with the individual to be insured, such
as a cholesterol level, a blood pressure level, a pulse, and other
characteristics.
[0059] At step 410, at least one decision component is evaluated.
As described above, evaluating a decision component may comprise
evaluating a decision component using a fuzzy logic rule. To
perform the evaluation, a rule may be defined and assigned to the
decision component. While each rule is generally only assigned one
decision component, it is understood that more than one decision
component may be assigned to each rule. Further, parameters for
each rule may be defined, as also described above.
[0060] At step 420, at least one measurement is assigned to the at
least one decision component. As described above with regard to the
application of a fuzzy logic rule, a measurement may be assigned to
the decision component from a sliding scale, such as between zero
(0) and one (1). Other types of measurements may also be
assigned.
[0061] At step 430, each decision component is assigned a specific
component category based on the assigned measurement. As described
above, a number of specific component categories arc defined. Based
on the assigned measurements, each decision component is assigned
to one or more specific component categories.
[0062] By way of the examples above, the specific component
categories may be defined as cat1, cat2, cat3, and cat4. Cat1 may
only be assigned decision components at a certain level or higher.
Similarly, cat2 may on I y be assigned decision components at a
second level or higher and so on. Other methods for assigning a
specific component category may also be used.
[0063] At step 440, the insurance application is assigned to a
category. According to an embodiment of the invention, the
categories to which the insurance application is assigned are the
same as the categories to which the insurance decision components
are assigned. As described above, the insurance application may be
assigned to a category based upon how the decision components were
assigned. Thus, by way of example, an insurance application may be
assigned to cat I only if two or fewer decision components are
assigned to cat2 and all other decision components are assigned to
cat I. Other methods for assigning an insurance application to a
category may also be used.
[0064] At step 450, an insurance policy is issued. Based on the
category to which it is assigned, certain amounts are paid to
maintain the insurance policy in a manner that is well known in the
industry. It is understood that based on a category, an insurance
policy may not be issued. The customers may decide the premiums are
too high. Alternatively, the insurance company may determine that
the risk is too great, and decide not to issue the insurance
policy.
Case Based Reasoning
[0065] A rule-based reasoning (RBR) system may provide for an
underwriting process by following a generative approach, typically
a rule-chaining approach, in which a deductive path is created from
the evidence (facts) to the decisions (goals). A case-based
reasoning (CBR) system, on the other hand, may follow an analogical
approach rather than a deductive approach. In such a system, a
reasoner may determine the correct rate class suitable for
underwriting by noticing a similarity of an application for
insurance with one or more previously underwritten insurance
applications and by adapting known solutions of such previously
underwritten insurance applications instead of developing a
solution from scratch. A plurality of underwriting descriptions and
their solutions are stored in a CBR Case Base and are the basis for
measurement of the CBR performance. According to an embodiment of
the invention, a CBR system may be only as good as the cases within
its Case Base (also referred to as "CB") and its ability to
retrieve the most relevant cases in response to a new
situation.
[0066] A case-based reasoning system can provide an alternative to
a rules-based expert system, and may be especially appropriate when
a number of rules needed to capture an expert's knowledge is
unmanageable, when a domain theory is too weak or incomplete, or
when such domain theory is too dynamic. The CBR system has been
successful in areas where individual cases or precedents govern the
decision-making processes.
[0067] In many aspects, a case-based reasoning system and process
is a problem solving method different from other artificial
intelligence approaches. In particular, instead of using only
general domain dependent heuristic knowledge, such as in the case
of an expert system, specific knowledge of concrete, previously
experienced, problem situations may be used with CBR. Another
important characteristic may be that CBR implies incremental
learning, as a new experience is memorized and available for future
problem solving each time a problem is solved. CBR may involve
solving new problems by identifying and adapting solutions to
similar problems stored in a library of past experiences.
[0068] According to an embodiment of the invention, an inference
cycle of the CBR process may comprise a plurality of steps, as
illustrated in the flow chart of FIG. 5. At step 502, probing and
retrieving one or more relevant cases from a case library is
performed. Ranking the retrieved relevant cases, based on a
similarity measure occurs at step 504. At step 506, one or more
best cases are selected. At step 508, one or more retrieved
relevant cases are adapted to a current case. The retrieved,
relevant cases are evaluated versus the current case, based on a
confidence factor at step 510. The newly solved case is stored in
the case memory at step 512.
[0069] These steps will be illustrated below within the context of
insurance underwriting. However, one of ordinary skill in the art
will recognize that these steps may be used in other contexts as
well. For purposes of this example only, assume that an applicant
provides his/her vital sign information (e.g., an age, a weight, a
height, a systolic blood pressure level and a diastolic blood
pressure level, a cholesterol level and a ratio, etc.) as a vector
equal to:
X=[x.sub.1, x.sub.2 . . . , x.sub.n].
[0070] Furthermore, in this example, assume that two of the values
corresponding to the cholesterol level, and a weight-to-height
ratio, are above normal levels, while the others fall within normal
ranges. The first two components of vector X correspond to the
cholesterol level (x.sub.1) and the weight-to-height ratio
(x.sub.2). For purposes of this example, the applicant has an
abnormally high cholesterol ratio (8.5%) and is over-weight
(weight-to-height ratio=3.8 lb/inch). Furthermore, the applicant
has one medical condition/history, for instance a history of
hypertension. This condition may require the applicant to provide
additional detailed information related to the history of
hypertension, e.g., a cardiomegaly, a chest pain, a blood pressure
mean and a trend over the past three months (where mean is the
average of the blood pressure readings over a particular time
period and trend corresponds to the slope of the reading such as
upward, or downward, etc.) The detailed information may be
contained in a vector Y=[y.sub.1, y.sub.2, . . . , y.sub.p], where
the value of p will vary according to the applicant's medical
condition.
[0071] The first step in the CBR methodology may be to represent a
new case (probe) as a query in a structured query language (SQL),
which may be formulated against a database of previously placed
applicants (cases). According to an embodiment of the invention,
the SQL query may be of the form: [0072] Q: [f.sub.1(x),
f.sub.2(x), . . . , f.sub.n(x)] AND [Condition=label] where
[f.sub.1(x), f.sub.2(x), . . . , f.sub.n(x)], will will be a vector
of n fuzzy preference functions, one of each of the elements of
vector X, and a label will be an index representing the applicant's
current medical condition. For this example, the CBR system may
retrieve all previous applicants with a history of hypertension,
whose vital signs were normal, except for a cholesterol ratio and a
weight-to-height ratio. In other words, the SQL query may be for
all cases matching the same condition and similar vital information
as the applicant. An example of such a SQL query may be: [0073]
Q1=[Support(Around (8.5%; x)), Support (Around (3.8; x)), Support
(Normal(i)), . . . , Support (Normal(n))] AND
[Condition=Hypertension]
[0074] The meaning of Normal(i) may be determined by a fuzzy logic
set representing a soft threshold for a variable, x(i), as it is
used in the stricter class rate, (e.g., Preferred Best in the case
of Life Insurance.) FIG. 6 illustrates the case of Normal G), where
x(j) corresponds to the cholesterol ratio. For example, it may be
determined from the most experienced underwriters of the insurance
company that under no circumstances can the applicant get the best
class rate if his/her cholesterol ratio is above X I by more than
five points. In that example, one may use X Ib-X Ia=5. The specific
values for X Ia and X Ib may be parameters of the model, and will
be explained below in greater detail. To obtain the partial degree
of satisfaction when the cholesterol ratio value falls within the
range [X1a, X1b], a continuous switching function may be used which
interpolates between the values I and 0. The simplest such function
is a straight line, but other functions may also be used.
[0075] In a linear membership function as shown in FIG. 6, the
values X Ia and X Ib are the low and high cutoffs, respectively. A
strict yes/no rule may be recovered in the limit that X Ia=X Ib.
Thus, many methods that mix fuzzy and strict rules in any
proportion m:iy be covered as a subset of this method.
[0076] Around (a; x) may be determined by a fuzzy relationship,
whose membership function can be interpreted as the degree to which
the value x meets the property of "being around a." If Around (a;
x)=1, then the value of x may be close to a well within a desired
tolerance. The support of the fuzzy relationship Around (a; x) may
be defined as the interval of values of x for which Around (a;
x)>0, as illustrated in FIG. 7. If Around (a; x)=0 then the
value of x is too far from a, beyond any acceptable tolerance.
[0077] The core of the fuzzy relationship Around (a; x) may be
defined as the interval of values of x for which Around (a; x)=1,
as illustrated in FIG. 8. Any value belonging to the core fully
satisfies the property and, in terms of a preference, it is
indistinguishable from any other value in the core.
[0078] A trapezoidal membership distribution representing the
relationship may have a natural preference interpretation. The
support of the distribution may represent a range of tolerable
values and correspond to an interval-value used in an initial SQL
retrieval query. The core may represent the most desirable range of
values and may establish a top preference. By definition, a feature
value falling inside the core will receive a preference value of 1.
As the feature value moves away from a most desirable range, its
associated preference value will decrease from I to 0. By
retrieving the cases having cholesterol ratios falling in the
support of Around (8.5%; x) and having weight-to-height ratios
falling in the support of Around (3.8; x) all possible relevant
cases may be retrieved.
[0079] In executing an SQL query Q1 of the above example against
the CBR database, N cases may be retrieved. By construction, all N
cases must have all of their vital values inside the support of the
corresponding element x(i) defined by QI. Furthermore, all cases
must be related to the same medical condition, (e.g.,
hypertension).
[0080] At this point, considering the outputs of each of the N
retrieved cases may provide a first preliminary decision. According
to an embodiment of the invention, a decision may be made only on
the retrieved cases, i.e., only using the first n variables and the
label used in the SQL query Q1. Each retrieved case may be referred
to as a case C.sub.k (k between 1 and N), and an output
classification of case Ck as Ok, where Ok is a variable having an
attribute value indicating the rate class assigned to the applicant
corresponding to case Ck. By way of example, Ok may assume one out
of T possible values, i.e., Ok=L, where L E {R.sub.1, R.sub.2, . .
. , R.sub.T}. For instance, in the case of Life insurance products,
L={Preferred-Best, Preferred, Preferred-Nicotine, . . . , Standard,
. . . , Table-32}. Other values may also be used.
[0081] In this example, the SQL query Q1 retrieves 40 cases (N=40).
FIG. 9 illustrates the histogram (distribution of the retrieved
cases over the rate classes) of the results of the SQL query QI. As
seen in FIG. 9, a first preliminary decision indicates Table-II as
being the most likely rate class for the new applicant represented
by the SQL query Q1.
[0082] All N cases may have all their vital values inside the
support of the corresponding element x(i) defined by the SQL query
Q1 and they are all related to the same medical condition, (e.g.,
hypertension). Therefore, each case may also contain p additional
elements corresponding to the variables specific to the medical
condition. A case Ck (k between 1 and N) may he represented as an
r-dimensional vector, where r=n+p. The first n elements correspond
to the n vital sign described by the vector X, namely [x.sub.1,k,
x.sub.2,k, . . . , x.sub.n,k]. The remaining p elements may
correspond to the specific features related to the condition
hypertension, namely [x.sub.(n+1), x.sub.(n+2), k, . . . ,
X.sub.r,k]. The value of p may vary according to the value of the
label, i.e., the medical condition.
[0083] A degree of matching between case Ck and the SQL query Q1
may be determined. To this extent, the n-dimensional vector
M(C.sub.k, Q1) may be defined as an evaluation of each of the
functions [f.sub.ix), f.sub.2(x), . . . , f.sub.n(x)] from the SQL
Query Q1 with the first n elements of Ck. namely [x.sub.1,k,
x.sub.2,k, . . . , x.sub.n,k]:
M(C.sub.k, Q1)=[f.sub.1(x.sub.1,k), f.sub.2(x.sub.2,k), . . . ,
f.sub.n(x.sub.n,k)]
[0084] At the end of this evaluation, each case will have a
preference vector whose elements take values in the (0,1] interval
(where the notation (0,1] indicates that this is an open interval
at 0 (i.e., it does not include the value 0), and a closed interval
at 1 (i.e., it includes the value 1)). These values may represent a
partial degree of membership of the feature value in each case and
the fuzzy relationships representing preference criteria in the SQL
query Q1. Since this preference vector represents a partial order,
the CBR system aggregates its elements to generate a ranking of the
case, according to their overall preference.
[0085] A determination is made of an n-dimensional weight vector
W=[w.sub.1, w.sub.2, . . . , w.sub.n] in which the element w.sub.i
takes a value in the interval [0,1] and determines the relative
importance of feature i in M(C.sub.k, Q1), i.e., relevance of
f.sub.j (x.sub.i,k). According to an embodiment of the invention,
this can be done via direct elicitation from an underwriter or
using pair-wise comparisons, following Saaty's method. By way of
example, if all features are equally important, all their
corresponding weights may be equal to 1. Other methods may also be
used. Once the weight vector has been determined, several
aggregating functions are used to rank the cases, where the
aggregating function will map an n-dimensional unitary hypercube
into a one-dimensional unit interval, i.e.
[0,1].sup.n->[0,1].
[0086] To consider compensation among the elements, a definition is
made of the aggregating function A[W,M(C.sub.k, Q1)] as a weighted
sum of its elements, i.e.
A [ W , M ( C k , Q 1 ) ] = i = 1 n w i f i ( x i , k )
##EQU00001##
[0087] Alternatively, a strict intersection aggregation without
compensation may be obtained using a weighted minimum, i.e.
A[W,M(C.sub.k, Q1)]=Minimum.sub.1,n[max(1-w.sub.j),
f(x.sub.i,k)]
Regardless of the aggregating function selected, it may be
considered as a measure of similarity between the each retrieved
case C.sub.k and the query Q1, and may be referred to as S(k, 1).
Using this measure, cases may be sorted according to an overall
degree of preference, which may be interpreted as a measure of
similarity between each retrieved case C.sub.k and the query
Q1.
[0088] In the first preliminary decision, the output of case
C.sub.k may be referred to as O.sub.k, where O.sub.k is a variable
whose attribute value indicates a rate class assigned to the
applicant corresponding to a case C.sub.k. Assume, for example,
that O.sub.k can take one out of T possible values, i.e.,
O.sub.k=L, where L .di-elect cons. (R.sub.1, R.sub.2, . . . ,
R.sub.T}. For instance, in the case of Life insurance products,
L={Preferred-Best, Preferred, Preferred-Nicotine, . . . , Standard,
. . . , Table-32}. However, not all cases are equally similar to
our probe. FIG. 10 illustrates a distribution of the similarity
measure S(k, 1) over the T for the retrieved N cases (e.g., N=40 in
the present example).
[0089] According to an embodiment of the invention, a minimum
similarity value may be considered for a case. For instance, to
only consider similar cases, a threshold may be established on the
similarity value. By way of example, only cases with a similarity
greater or equal to 0.5 may be considered. According to an
embodiment of the invention, a determination may be made of a fuzzy
cardinality of each of the rate classes, by adding up the
similarity values in each class. Other distributions may also be
evaluated.
[0090] A histogram may be drawn that aggregates the original
retrieval frequency with the similarity of the retrieved cases, and
may be referred to as a pseudo-histogram. This process may be
similar to a N-Nearest Neighbor approach, where the N retrieved
cases represent the N points in the neighborhood, and the value of
S(k, 1) represents the complement of the distance between the point
K and the probe, i.e., the similarity between each case and a
query. The rate class Ri, with the largest cumulative measure may
be proposed as a solution. By way of example, Table-II is the
solution indicated by either option.
[0091] A decision may be made on how many cases will be used to
refine a solution. Having sorted the cases along the first n
dimensions, the remaining p dimensions may be analyzed
corresponding to the features related to the specific medical
condition. Some of these medical conditions may have variables with
binary or attribute values (e.g., chest pain (YIN), malignant
hypertension (N), Mild, Treated, etc.), while others ones may have
continuous values (e.g., cardiomegaly (% of enlargement), systolic
and diastolic blood pressure averaged and trend in past 3 months,
24 months, etc.).
[0092] An attribute-value and a binary-value may be used to select,
among the N retrieved cases, the cases that have the same values.
This may be the same as performing a second SQL query, thereby
refining the first SQL query Q1. From the originally retrieved N
cases, the cases with the correct binary or attribute values may be
selected. This may be done for all of the attribute-values and the
binary-valued variables, or for a subset of the most important
variables. After this selection, the original set of cases will
likely have been reduced. However, when a Case Base is not
sufficiently large, a reduction in the number of variables used to
perform this selection may be needed. Assuming that there are now L
cases (where L<N), these cases may still he sorted according to
a value of a similarity metric S(k, 1).
[0093] A third preliminary decision may be obtained by re-computing
the distribution of the similarity measure S(k, 1) over the T
values for the output O.sub.k, and then proposing as a solution the
class Ri with the largest cumulative measure using the same
pseudo-histogram method described above.
[0094] A similarity measure over the numerical features related to
the medical condition may be obtained by establishing a fuzzy
relationship Around(a; x) similar to the one described above. This
fuzzy relationship would establish a neighborhood of cases with
similar condition intensities. By performing an evaluation and an
aggregation similar to one described above, a similarity measure
may be obtained by medical condition, and may be referred to as
I(k, 1).
[0095] A final decision may involve creating a linear combination
of both similarity measures:
F(k,1)=.alpha.S(k,1)+(1-.alpha.)I(k,1),
thereby providing the distribution of the final similarity measure
F(k, 1) over the T values of O.sub.k. According to an embodiment of
the invention, the final decision or solution may be the class
R.sub.i with the largest cumulative measure using the same
pseudo-histogram method.
[0096] A reliability of the solution may be measured in several
ways, and as a function of many internal parameters computed during
this process. According to an embodiment of the invention, the
number of retrieved (N) and refined (L) cases (e.g., area of the
histogram) may be measured. Larger values of N+L may imply a higher
reliability of the solution. According to another embodiment of the
invention, the fuzzy cardinality of the retrieved and refined cases
(i.e., area of the pseudo-histogram) may be measured. Larger values
may imply a higher reliability of the solution. According to a
further embodiment of the invention, the shape of the
pseudo-histogram of the values of O.sub.k, (i.e., spread of the
histogram) may be measured, where a tighter distribution (smaller
sigmas) would be more reliable than scattered ones. According to
another embodiment of the invention, the mode of the
pseudo-histogram of the values of O.sub.k, (e.g., maximum value of
the histogram) may be measured. Higher values of the mode may be
more reliable than lower ones. A contribution of one or more of
these measurements may be used to determine reliability. Other
measurements may also be used.
[0097] Using a training set, a conditional probability of
misclassification as a function of each of the above parameters may
be determined, as well. Then, the (fuzzy) ranges of those
parameters may be determined and a confidence factor may be
computed.
[0098] If the solution does not pass a confidence threshold (e.g.,
because it does not have enough retrieved cases, has a scattered
pseudo-histogram, etc.), then the CBR system may suggest a solution
to the individual underwriter and delegate to him/her the final
decision. Alternatively, if the confidence factor is above the
confidence threshold, then the CBR system may validate the
underwriter's decision. Regardless of the decision maker, once the
decision is made, the new case and its corresponding solution are
stored in the Case Base, becoming available for new queries.
[0099] According to an embodiment of the invention, clean cases
(previously placed by rule base) may be used to tune the CBR
parameters (e.g., membership functions, weights, and similarity
metrics), thereby abating risk. Other methods for abating risk may
also be used.
[0100] By defining and using three stages of preliminary decisions,
the CBR system may display tests, thereby generating useful
information for the underwriter while the Case Base is still under
development. As more information (cases and variables describing
each case) is stored in the Case Base, the CBR system may be able
to use a more specific decision stage.
[0101] According to an embodiment of the invention, the first two
preliminary decision stages may only require the same vital
information used for clean applications and the symbolic (i.e.,
label) information of the medical condition. A third decision'
stage may make use of a subset of the variables describing the
medical condition thereby refining the most similar cases. The
subset of variables may be chosen by an expert underwriter as a
function of their relevance to the insured risk (mortality,
morbidity, etc.). This step will allow the CBR system to refine the
set of N retrieved cases, and select the most similar L cases, on
the basis of the most important binary and attribute variables
describing the medical condition. The final two preliminary
decision stages may only require the same vital information used
for clean applications and the symbolic (i.e., label) information
of the medical condition.
[0102] According to an embodiment of the invention, it may be
important that at all times the value of N (for the first two
decision stages) and the value of L (for the third decision stage)
be large enough to ensure significance. The number of cases used
may be one of the parameters used to compute the confidence factor
described above.
[0103] In the first step of the example, the new case (probe) was
represented as a SQL query, and it was assumed that only one
medical condition was present. The complete SQL query Q may have
been formulated as:
Q: [f.sub.1,(x), f.sub.2(x), . . . , f.sub.n(x)] AND
[Condition=label] AND [Condition number=1]
If the applicant has more than one medical condition, the applicant
may be compared with other applicants having the same medical
conditions. By way of another example extending the original
example used, the applicant is assumed to have an abnormally high
cholesterol ratio (8.5%) and be over-weight (weight-to-height
ratio=lb/inch). Furthermore, the applicant discloses that he/she
has two medical conditions, (e.g., hypertension and diabetes).
[0104] In a densely populated Case Base, the applicant may be
represented by the query: [0105] Q: [f.sub.1,(x), f.sub.2(x), . . .
, f.sub.n(x)] AND [Condition 1=label] AND [Condition 2=label 2] AND
[Condition number=2] This query may be instantiated as: [0106] Q1:
[Support(Around (8.5%, x)), Support (Around (3.8; x)), Support
(Normal(i)), . . . , Support (Normal(n))] AND
[Condition=Hypertension] AND [Condition=Diabetes] AND [Condition
number=2] With a well-populated Case Base, this may be a process
for handling multiple medical conditions in complex cases.
[0107] As more conditions are added to a query, fewer cases will
likely be retrieved. If the retrieved number of cases N is not
significant, a useful decision may not be produced. An alternative
(surrogate) solution may be to decompose a query into two separate
queries, treating each medical condition separately. For instance,
assuming that the modified query Q1 requesting two simultaneous
conditions does not yield any meaningful result, the CBR system may
decompose the query Q1 into a plurality of queries, Q1-A and Q1-B:
[0108] where Q1-A: [Support(Around (8.5%, x)), Support (Around
(3.8; x)), Support (Normal(i)), . . . , Support (Normal(n))] AND
[Condition=Hypertension] AND [Condition number=I]; and [0109] where
Q1-B: [Support(Around (8.5%, x)), Support (Around (3.8; x)),
Support (Normal(i)), . . . , Support (Normal(n))] AND
[Condition=Diabetes] AND [Condition number=1] Each query may be
treated separately and may obtain a decision on the rate class for
each of the queries. In other words, it may be assumed that there
are two applicants, both overweight and with a high cholesterol
ratio, one with hypertension and one with diabetes.
[0110] After obtaining suggested placements in the appropriate rate
class, (e.g., RC-A and RC-B, respectively) the answers may be
combined according to a set of aggregation rules representing the
union of multiple rate classes induced by the presence of multiple
medical conditions. According to an embodiment of the invention,
these rules may be elicited from experienced underwriters. A
look-up table, as illustrated in FIG. 11, may represent this rule
set. FIG. 11 is just an example that shows a linear aggregation of
the rate classes. Assume that the rate class assigned to query Q1-A
is RC-A=Table 6 and the rate class assigned to query Q1-B is
RC-B=Table 8. The combined rate class generated from the
aggregation rule is RC=Table 14. Other tables may be designed to
over-penalize the occurrence of multiple conditions as their
presence might affect risk and, therefore, claims, in a non-linear
fashion. For example RC-A=Table 6 and RC-B=Table 8 could be
aggregated into RC=Table 18 by a stricter table. Other aggregation
process may also be used.
[0111] Additionally, these tables may be used in an associative
fashion. In other words, when an applicant has three or more
medical conditions, the CBR system may aggregate the rate classes
derived from the first two medical conditions, obtain the result
and aggregate the result with the rate class obtained from the
third medical condition, and so on, as illustrated in FIG. 11. This
method is a surrogate alternative that may be used when enough
cases with multiple conditions are included in the Case Base.
[0112] According to an embodiment of the invention, a CBR engine
may be encoded into a Java based computer code, which can query a
database to obtain the case parameters, and write its decision in
the database as well. This embodiment may be suitable for batch
processing, and for use in a fully automated underwriting
environment.
Calculation of Confidence Factor
[0113] A described above, CBR may be used to automate decisions in
a variety of circumstances, such as, but not limited to, business,
commercial, and manufacturing processes. Specifically, it may
provide a method and system to determine at run-time a degree of
confidence associated with the output of a Case Based Decision
Engine, also referred to as CBE. Such a confidence measure may
enable a determination to be made on when a CBE decision is
trustworthy enough to automate its execution and when the CBE
decision is not as reliable and may need further consideration. If
a CBE decision is not determined to be as reliable, a CBE analysis
may still be beneficial by providing an indicator, forwarding it to
a human decision maker, and improving the human decision maker's
productivity with an initial screening that may limit the
complexity of the final decision. The run-time assessment of the
confidence measure may enable the routing mechanism and increases
the usefulness of a CBE.
[0114] An embodiment of the invention may comprise two parts: a)
the run-time computation of a confidence factor for a query; and b)
the determination of the threshold Lu be used with the computed
confidence factor. FIG. 12 is a flowchart illustrating a process
for determining a run-time computation of a confidence factor
according to an embodiment of the invention. At Step 1200, a
confidence factor process is initiated. At Step 1210, CBE internal
parameters that may affect the probability of misclassification are
identified. At Step 1220, the conditional probability of
misclassification for each of the identified parameters is
estimated. At Step 1230, the conditional probability of
misclassification is translated into a soft constraint for each
parameter. At Step 1240, a run-time function to evaluate the
confidence factor for each new query is defined. The determination
of the threshold for the confidence factor may be obtained by using
a gradient-based search. It is understood that other steps may he
performed within this process, and/or the order of steps may be
changed. The process of FIG. 12 will now be described in greater
detail below.
[0115] According to an embodiment of the invention, CBE may be used
to automate the underwriting process of insurance policies. By way
of example, CBE may be used for underwriting life insurance
applications, as illustrated below. It is understood, however, that
the applicability of this invention is much broader, as it may
apply to any Case-Based Decision Engine(s).
[0116] According to an embodiment of the invention, an advantage of
the present invention may include improving deployment of a method
and system of automated insurance underwriting, based on the
analysis of previous similar cases, as it may allow for an
incremental deployment of the CBE, instead of postponing deployment
until an entire case base has been completely populated. Further, a
determination may be made for which applications (e.g.,
characterized by specific medical conditions) the CBE can provide
sufficiently high confidence in the output to shift its use from a
human underwriter productivity tool to an automated placement tool.
As a case base (also referred to as a "CB") is augmented and/or
updated by new resolved applications, the quality of the retrieved
cases may improve. Another advantage of the present invention may
be that the quality of the case base may be monitored, thereby
indicating the portion of the case base that requires growth or
scrubbing. For instance, monitoring may enable identification of
regions in the CB with insufficient coverage (small area
histograms, low similarity levels), regions containing inconsistent
decisions (bimodal histograms), and ambiguous regions (very broad
histograms).
[0117] In addition, by establishing a confidence threshold, a
determination may be made whether the output can be used directly
to place the application or if it will be a suggestion to be
revised by the human underwriter, where such a determination may be
made for each application processed by the CBE. Further, according
to an embodiment of the invention, a process may be used after the
deployment of the CBE, as part of maintenance of the case base. As
the case base is enriched by the influx of new cases, the
distribution of its cases may also vary. Regions of the case base
that were sparsely populated might now contain a larger number of
cases. Therefore, as part of the tuning of the CBE, one may
periodically recompute certain steps within the process to update
the soft constraints on each of the parameters. As part of the same
maintenance, one may also periodically update the value of the best
threshold to be used in the process.
[0118] While the present invention is described in relation to
applicability to the improvement of the performance of a Case Based
Engine for Digital Underwriting, it is understood that the method
and system described herein may be applied to any Case Based
Reasoning system, to annotate the quality of its output and decide
whether or not to act upon the generated output. By way of example,
CBR systems may have applications in manufacturing, scheduling,
design, diagnosis, planning, and other areas.
[0119] As described above, the CBE relies on having a densely
populated Case Base ("CB") from which to retrieve the precedents
for the new application (i.e., the similar cases). According to an
embodiment of the invention, until the CB contains a sufficiently
large number of cases for most possible applications, the CBE
output may not be reliable. Such an output may, by way of example,
be used as a productivity aid for a human underwriter, rather than
an automation tool.
[0120] For each processed application, a measure of confidence in
the CBE output is computed so that a final decision maker (CBE or
human underwriter) may be identified. As the decision engine
generates its output from the retrieval, selection, and adaptation
of the most similar cases, such a confidence measure may reflect
the quality of the match between the input (the application under
consideration) and the current knowledge, e.g., the cases used by
the CBE for its decision.
[0121] The confidence measure proposed by this invention needs to
reflect the quality of the match between the current application
under consideration and the cases used for the CBE decision. This
measure needs to be evaluated within the context of the statistics
for misclassification gathered from the training set. More
specifically, according to an embodiment of the invention, the
steps described below may be performed. These steps may include,
but are not limited to, the following: 1) Formulate a query against
the CB, reflecting the characteristics of the new application as
query constraints; 2) Retrieve the most relevant cases from the
case library. For purposes of illustration, assume that N cases
have been retrieved, where N is greater than 0 (i.e., not a null
query or an empty retrieved set of cases). A histogram of the N
cases is generated over the universe of their responses, i.e., a
frequency of the rate class; 3) Rank the retrieved cases using a
similarity measure; 4) Select the best cases thereby reducing the
total number of useful retrieved cases from N to L; and 5) Adapt
the L refined solutions to the current case in order to derive a
solution for the case. By way of example, selecting the mode of the
histogram may be used to derive a solution.
[0122] To determine the confidence in the decision, it may be
desirable to understand what the probability of generating a
correct or incorrect classification is. Specifically, it may be
desirable to identify which factors affect misclassifications, and,
for a given case, use these factors to assess if it is more or less
likely to generate a wrong decision. According to an embodiment of
the invention, unless a decision is binary, the decision will
consist of placing the case under considerations in one of several
bins. Hence, there may be different degrees of misclassification,
depending on the distance of the CBE decision from the correct
value. Given the different costs associated with different degrees
of misclassification, the factors impacting the decision may be
used with the likely degree of misclassification.
[0123] One aspect of the present invention deals with the process
and method used to accomplish this result. At Step 1210 the CBE
internal parameters that might affect the probability of
misclassification may be determined. Each of these parameters may
be referred to as an x. Furthermore, assume that there are M
parameters (i.e., i=1, . . . M, forming a parameter vector
X=[x.sub.1, x.sub.2, . . . x.sub.M].
[0124] Parameters that may affect the probability of
misclassification include, but are not limited to, the following
potential list of candidates: [0125] x.sub.1: N=Number of retrieved
cases (i.e., cardinality of retrieved set and area of histogram in
FIG. 9), e.g., N=40 cases. [0126] x.sub.2: variability of retrieved
cases (measure of dispersion of histogram in FIG. 9). [0127]
x.sub.3: number of retrieved cases thresholded by similarity value
(area of histogram in FIG. 10) e.g., 25 cases. [0128] x.sub.4:
variability of retrieved cases thresholded by similarity value.
(measure of dispersion of histogram in FIG. 10). [0129] x.sub.5:
L=number of refined cases. (i.e., cardinality of refined set) e.g.,
21 cases. [0130] x.sub.6: variability of refined cases. [0131]
x.sub.7: number of refined cases, thresholded by similarity value
e.g., 16 cases. [0132] x.sub.8: variability of refined cases
thresholded by similarity value. [0133] x.sub.9: measure of
strength of mode (percentage of cases in mode of histogram) e.g.,
50%.
[0134] According to an embodiment of the invention, other
parameters may include: [0135] x.sub.10: number of retrieved cases
weighted by similarities. (i.e. fuzzy cardinality of retrieved set
(area of histogram in FIG. 9)). [0136] x.sub.11: variability of
retrieved cases weighted by similarities (measure of dispersion of
histogram in FIG. 9). [0137] x.sub.12: number of refined cases
weighted by similarities (i.e. fuzzy cardinality of refined set).
[0138] x.sub.13: variability of refined cases weighted for
similarities.
[0139] These parameters may be query-dependent, (e.g., they may
vary for each new application). This may be in contrast to static
design parameters, such as, but not limited to, similarity weights,
retrieval parameters, and confidence threshold. Static parameters
may be tuned at development time (e.g., when a system is initially
developed) and periodically revised at maintenance time(s) (e.g.,
during maintenance periods for a system). According to an
embodiment of the invention, static parameters may be considered
fixed while evaluating parameters [x1-x9:].
[0140] According to an embodiment of the invention, the above
parameters may likely be positively correlated. By way of example,
the number or refined cases L may depend on the total number of
cases N. The relative impact of these parameters may be evaluated
via a statistical correlation analysis, CART, C4.5 or other
algorithms to identify and eliminate those parameters that
contribute the least amount of additional information. By way of
another example, methods may be used to handle partially redundant
information in a way that avoids double counting of the evidence.
The use of a minimum operator in the computation of the Confidence
Factor, as is described below, is such an example.
[0141] According to an embodiment of the invention, at step 1220,
the conditional probability of misclassification for each parameter
Xi (for i=1 . . . 9) may be estimated. By way of example, this step
may be achieved by running a set of experiments with a training
set. Given a certified Case Base (e.g., a CB containing a number K
of cases whose associated decisions were certified correct), the
following steps may then be followed: [0142] (1) For each of the K
cases in the CB, one case is selected (from the CB) and may be
considered as the probe, i.e., the case whose decision we want to
determine (1310). [0143] (2) The Case Based Engine (CBE) and the
(K-1) cases remaining in the CB may then be used to determine the
rate class (i.e., the placement decision for the probe) (1320).
[0144] (3) The decision derived from the CBE may then be compared
with the original certified decision of the probe (1330). [0145]
(4) The comparison and its associated parameters [x.sub.1-x.sub.9]
may then be recorded. [0146] (5) The selected case may be placed in
the CB and another case selected. (i.e., back to step (1) (1340)).
[0147] (6) Perform steps (2) through (5) until all the K cases in
the CB have been used as probes (1350).
[0148] This process is illustrated in FIG. 13. Once the process is
completed, the results may be collected and analyzed. The
comparison matrix of FIG. 14 illustrates a comparison between a
probe's decision derived from the CBE and the probe's certified
reference decision. The cells located on the comparison matrix's
main diagonal may contain the percentage of correct
classifications. The cells off the main diagonal may contain the
percentage of incorrect classifications. As was previously
mentioned, there may be different degrees of misclassification,
depending on the distance of a CBE decision from the corresponding
reference decision.
[0149] At this point, it may be desirable to estimate the
conditional probability of misclassification given each of
parameters [x1-x9]. Since each case in the comparison matrix has
its associated parameters [x1-x9] recorded, a histogram of the
distance from the correct decision for each of these parameters may
be generated.
[0150] This process may be illustrated by a simple example. As was
previously described, the value of the first parameter x1: [0151]
x.sub.1: N=Number of retrieved cases. (i.e., cardinality of
retrieved set (area of histogram in FIG. 9))
[0152] FIG. 15 shows an example of cross-tabulation of
classification distances and number of retrieved cases for each
probe. By way of this example, the processing of 573 probes is
shown, achieving a correct classification for 242 of them.
Additionally, 214 were classified as one rate class off (where 114
at (-1) and 100 at (+1) equal 214). Further, 99 were two rate
classes off (where 64 at (-2) and 35 at (+2) equal 99), and 18 were
3 or more classes off. These 573 cases may also be subdivided in
ten bins, representing ranges of the number of retrieved cases used
for each probe. By way of example, 41 cases had between 1 and 4
retrieved cases (first column), while 58 cases used more than 40
retrieved cases (last column). FIG. 16 illustrates the same
cross-tabulation using percentages instead of the number of cases.
According to an embodiment of the invention, this table may be
referred to as matrix D(i, j), where i=I . . . 7 (the seven
distances considered), and j=I . . . 10 (the ten bins
considered).
[0153] Note that this table contains the same percentages
illustrated in FIG. 15, once we normalize the values by the total
number of cases, tabulated for different values of x 1. For
instance, the total percentage of Correct Classifications (CC) in
FIG. 14 may be defined as the sum of the elements on the main
diagonal, i.e.
% CC = i = 1 T M ( i , i ) ##EQU00002##
[0154] The same percentage may be obtained by adding the
percentages distributed along the fourth row (corresponding to
Distance 0), i.e.
% CC = j = 1 10 D ( 4 , j ) ##EQU00003##
[0155] The percentage of correct classification may increase with
the number of cases retrieved for each probe (fourth row,
distance=0). By analyzing a given column on this table, an estimate
may be derived of the probability of correct/incorrect
classification, given that the number of cases is in the range of
values corresponding to the column.
[0156] According to an embodiment of the invention, step 1230 may
comprise translating the conditional probability of
misclassification into a soft constraint for each parameter
x.sub.i(for i= . . . 9). By way of example, all misclassifications
are determined to be equally undesirable, the only concern may be
with the row corresponding to distance equal 0 (i.e., correct
classification), as illustrated in FIG. 17. By way of another
example, it may be desirable to penalize more those
misclassifications that are two or three rate classes away from the
correct decision. Therefore, an overall performance function may be
formulated that aggregates the rewards of correct classifications
with increasing penalties for misclassifications. Although various
types of aggregating function may be used to achieve these ends,
one possible aggregating function may use a weighted sum of rewards
and penalties. Specifically, for each bin (range of values) of the
parameter x 1 under consideration, a reward/penalty wi may he
considered. For instance:
f ( Bin k ) = i = 1 7 w 1 D ( i , k ) ##EQU00004##
[0157] Where, for example, the weight vector W[w.sub.i, i=1 . . . 7
is W=-11, -6, -1,4, -1, -6. -11]
[0158] This weight vector indicates that misclassifying a decision
by three or more rate classes is eleven times worse than a
misclassification that is one rate class away. Except for the
fourth element, which indicates the reward for correct
classifications, all other elements in vector W indicate the
penalty value for the corresponding degree of misclassification.
FIG. 18 illustrates the result of applying the performance function
j(Bin.sub.k) to the values of FIG. 16, i.e., Matrix D.
[0159] By interpreting the values of FIG. 18 as degree of
preference, a fuzzy membership function Ci(x.sub.i), is derived,
indicating the tolerable and desirable ranges for each parameter
x.sub.1. According to an embodiment of the invention, a possible
way to convert the values of FIG. 18 to a fuzzy membership function
is to replace any negative value with a zero and then normalize the
elements by the largest value. In this example, the result of this
process is illustrated in FIGS. 19 and 20.
[0160] As previously described, the membership function of a fuzzy
set is a mapping from the universe of discourse (the range of
values of the performance function) into the interval [0,1]. The
membership function has a natural preference interpretation. The
support of the membership function Ci(x,) represents the range of
tolerable (i.e., acceptable) values of x.sub.1. The support of the
fuzzy set Ci(x.sub.i) is defined as the interval of values of x for
which Ci(x.sub.i)>0. Similarly, the core may represent the most
desirable range of values and establish a top preference. The core
of the membership function Ci(x.sub.i) may be defined as the
interval of values x, for which Ci(x.sub.i)=1. In the example of
FIG. 20, the support is [22, infinity] and the core is [40,
infinity]. By definition, a feature value falling inside the core
will receive a preference value of 1. As the feature value moves
away from the most desirable range, its associated preference value
will decrease from 1 to 0. At this point, the information may be
translated into a soft constraint representing our preference for
the values of parameter x. The soft constraint may be referred to
as Ci(x.sub.i), as illustrated in FIG. 20.
[0161] According to an embodiment of the invention, a fourth step
of this invention may be to define a nm-time function to evaluate
the confidence measure for each new query. By way of example, after
executing the third step for each of the nine parameters, nine soft
constraints may be obtained Ci(x.sub.i) i=1, . . . , 9. A soft
constraint evaluation (SCE) vector is generated that contains the
degree to which each parameter satisfies its corresponding soft
constraint; SCE [C.sub.1(x.sub.1), . . . , C.sub.9(x.sub.9)]. The
Confidence Factor (CF.sub.j) to be associated to each new case j
may be computed at run-time as the intersection of all the soft
constraints evaluations contained in the SCE vector.
CF j = i = 1 9 C i ( x i ) = Min i = 1 9 C i ( x i )
##EQU00005##
[0162] According to an embodiment of the invention, all elements in
the Soft Constraint Evaluation (SCE) vector may be real numbers in
the interval [0,1]. Therefore the Confidence Factor CFj will also
be a real number in the interval [0,1]. Nine potential soft
constraints represent the most desirable fuzzy ranges for the nine
parameters described above. Given a new probe, its computed
parameter vector X=[x.sub.1-x.sub.9] may used be to determine the
degree to which all soft constraints are satisfied (SCE), leading
to the computation of its Confidence Factor CF.
[0163] As previously described above, a four-step process was
described to compute at run-time the confidence factor. The minimum
threshold for the confidence value may be determined by a series of
experiments with the data, to avoid being too restrictive or too
inclusive. A higher-than-needed threshold may decrease the coverage
provided by the CBE by rejecting too many correct solutions (False
Negatives). As the threshold is lowered, the number of accepted
solutions is increased and therefore, an increase in coverage is
obtained. However, a lower-than needed threshold may decrease the
accuracy provided by the CBE by accepting too many incorrect
solutions (False Positives). Therefore, it may be desirable to
obtain a threshold using a method that balances these two
concepts.
[0164] According to an embodiment of the invention, coverage for
any given threshold level r may include accepting n(r) cases out of
K. Given a Case Base with K cases, the function g.sub.1(t) may be
defined as a measure or coverage:
g.sub.1(.tau.)=n(.tau.)/K
[0165] For accuracy, the performance function}; as previously
defined, may be used (e.g., aggregate the rewards of correct
classifications with the increasing penalties for
misclassifications) and may be adapted to the entire Case Base to
evaluate its accuracy for any given threshold r. As the value of r
is modified, more decisions may be accepted or rejected, modifying
the entries of the comparison matrix M=[M(i, j)].
2 ( .tau. ) = i = 1 T K * R * M ( i , i ) + i = 1 T j = 1 , j
.noteq. 1 T p ( i , j ) * R * M ( i , j ) ##EQU00006##
[0166] Specifically, the function g.sub.2(.tau.) may be defined as
a measure of relative accuracy, where M(i, j) is the (i, j) element
of the comparison matrix illustrated in FIG. 14. It may represent
the percentage of cases classified in cell i while the correct
classification was cell j. Therefore (i=j) implies a correct
classification. The percentage may be computed over the total cases
for which the decision has been accepted (i.e., its corresponding
confidence was above the threshold). Further, K*R may be a reward
for correct classification (where K indicates a static multiple of
basic reward R), and p(i, j)*R may be the penalty for incorrect
classification (p(i, j) determine a dynamic multiple of basic
reward R).
[0167] For simplicity, R=1 may he used. The penalty function p(i,
j) may indicate the increasing penalty for misclassifications
farther away from the correct one. Many possible versions of
function p(i, j) can be used. By way of example, the vector W={-11,
-6, -1, 4, -1, -6, -11] corresponds to the values:
K=4 and
p(i,j)=5|i-j|+4
[0168] A linear penalty function p(i, j) is illustrated m FIG. 30.
It will be recognized by those of ordinary skill in the art that
other linear functions may also be used. If over-penalization for
larger misclassifications is desired, a non-linear penalty function
may be used, such as p(i, j)=-3(i-j).sup.2+4, such as that
illustrated in FIG. 31.
[0169] The selection of a penalty function may be left as a choice
to a user to represent the cost of different misclassifications.
According to an embodiment of the invention, if there were no
differences among such costs, then a simplified version of
g.sub.2(r) could be used to measure the CBE accuracy, e.g.
2 ( .tau. ) = i = 1 T K * R * M ( i , i ) ##EQU00007##
[0170] Functions g.sub.1(t) and g.sub.2(t) may be defined to
measure coverage and relative accuracy, respectively. Function
g.sub.1(t) may be a monotonically non-increasing with the value t
(larger values of t will not increase coverage) , while g.sub.2(t)
may be a monotonically non-decreasing with the value t (larger
values of t will not decrease relative accuracy, unless the set is
empty.). The two functions may be aggregated into a global accuracy
function A(t) to evaluate the overall system performance under
different thresholds t:
A(.tau.)=g.sub.1(.tau.).times.g.sub.2(.tau.)
where x indicates scalar multiplication
[0171] The function A(t) provides a measure of accuracy combined
with the coverage of cases. FIG. 21 illustrates an example of the
computation of Coverage, Relative Accuracy, and Global Accuracy as
a function of threshold t. In this example, t=0.1 has the largest
coverage, t=0.7 has the largest relative accuracy, and t=0.5 has
the largest global accuracy.
[0172] There are many approaches that may be used to maximize the
aggregate function A(t) to obtain the best value for threshold t.
Any reasonable optimization algorithm (such as a gradient-based
search, or a combined gradient and binary search) may be used to
this effect. For example, in FIG. 21, the value of A(t) may be
computed for nine values of t. According to an embodiment of the
invention, values may be explored to determine a best threshold. By
way of example only, the neighborhood of t=0.5 may be explored,
such as by a gradient method, to determine that the value t=0.55 is
the best threshold.
[0173] As described above, the present invention provides many
advantages. According to an embodiment of the present invention,
incremental deployment of the CBE may be achieved, instead of
postponing its deployment until an entire Case Base has been
completely populated. Further, a determination may be made for
which applications (e.g., characterized by specific medical
conditions) the CBE can provide sufficiently high confidence in the
output to shift its use from a human underwriter productivity tool
to an automated placement tool.
[0174] According to an embodiment of the invention, as the Case
Base is augmented and or updated by new resolved applications, the
quality of the retrieved cases may change. The present invention
may enable monitoring of the quality of the Case Base, indicating
the part of the CB requiring growth or scrubbing. By way of
example, regions within the Case Base with insufficient coverage
(small area histograms, low similarity levels) may be identified,
as well as regions containing inconsistent decisions (bimodal
histograms), and ambiguous regions (very broad histograms).
[0175] According to an embodiment of the invention, by establishing
a confidence threshold, a determination can be made, for each
application processed by the CBE, if the output can be used
directly to place the application or if it will be a suggestion to
be revised by a human underwriter.
[0176] According to an embodiment of the invention, a process as
described above may be used after the deployment of the CBE, as
part of the Case Base maintenance. As the Case Based is enriched by
the influx of new cases, the distribution of its cases may also
vary. Regions of the CB that were sparsely populated might now
contain a larger number of cases. Therefore, as part of the tuning
of the CBE, one should periodically recompute various steps within
the process to update the soft constraints on each of the
parameters. As part of the same maintenance, the value of the best
threshold may also be updated and used m the process.
Network-Based Underwriting System
[0177] FIG. 22 illustrates a system 2200 according to an embodiment
of the present invention. The system 2200 comprises a plurality of
computer devices 2205 (or "computers") used by a plurality of users
to connect to a network 2202 through a plurality of connection
providers (CPs) 2210. The network 2202 may be any network that
permits multiple computers to connect and interact. According to an
embodiment of the invention, the network 2202 may be comprised of a
dedicated line to connect the plurality of the users, such as the
Internet, an intranet, a local area network (LAN), a wide area
network (WAN), a wireless network, or other type of network. Each
of the CPs 2210 may be a provider that connects the users to the
network 2202. For example, the CP 2210 may be an Internet service
provider (ISP), a dial-up access means, such as a modem, or other
manner of connecting to the network 2202. In actual practice, there
may be significantly more users connected to the system 2200 than
shown in FIG. 22. This would mean that there would be additional
users who are connected through the same CPs 2210 shown or through
another CP 2210. Nevertheless, for purposes of illustration, the
discussion will presume three computer devices 2205 are connected
to the network 2202 through two CPs 2210.
[0178] According to an embodiment of the invention, the computer
devices 2205a-2205c may each make use of any device (e.g., a
computer, a wireless telephone, a personal digital assistant, etc.)
capable of accessing the network 2202 through the CP 2210.
Alternatively, some or all of the computer devices 2205a-2205c may
access the network 2202 through a direct connection, such as a TI
line, or similar connection. FIG. 22 shows the three computer
devices 2205a-2205c, each having a connection to the network 2202
through the CP 2210a and the CP 2210b. The computer devices
2205a-2205c may each make use of a personal computer such as a
computer located in a user's home, or may use other devices which
allow the user to access and interact with others on the network
2202. A central controller module 2212 may also have a connection
to the network 2202 as described above. The central controller
module 2212 may communicate with one or more modules, such as one
or more data storage modules 2236, one or more evaluation modules
2224, one or more case database modules 2240 or other modules
discussed in greater detail below.
[0179] Each of the computer devices 2205a-2205c used may contain a
processor module 2204, a display module 2208, and a user interface
module 2206. Each of the computer devices 2205a-2205c may have at
least one user interface module 2206 for interacting and
controlling the computer. The user interface module 2206 may be
comprised of one or more of a keyboard, a joystick, a touchpad, a
mouse, a scanner or any similar device or combination of devices.
Each of the computers 2205a-2205c may also include a display module
2208, such as a CRT display or other device. According to an
embodiment of the invention, a developer, a user of a production
system, and/or a change management module may use a computer device
2205.
[0180] The central controller module 2212 may maintain a connection
to the network 2202 such as through a transmitter module 2214 and a
receiver module 2216. The transmitter module 2214 and the receiver
module 2216 may be comprised of conventional devices that enable
the central controller module 2212 to interact with the network
2202. According to an embodiment of the invention, the transmitter
module 2214 and the receiver module 2216 may be integral with the
central controller module 2212. According to another embodiment of
the invention, the transmitter module 2214 and the receiver module
2216 may be portions of one connection device. The connection to
the network 2202 by the central controller module 2212 and the
computer devices 2205 may be a high speed, large bandwidth
connection, such as through a T1 or a T3 line, a cable connection,
a telephone line connection, a DSL connection, or another similar
type of connection. The central controller module 2212 functions to
permit the computer devices 2205a-2205c to interact with each other
in connection with various applications, messaging services and
other services which may be provided through the system 2200.
[0181] The central controller module 2212 preferably comprises
either a single server computer or a plurality of server computers
configured to appear to the computer devices 2205a-2205c as a
single resource. The central controller module 2212 communicates
with a number of modules. Each module will now be described in
greater detail.
[0182] A processor module 2218 may be responsible for carrying out
processing within the system 2200. According to an embodiment of
the invention, the processor module 2218 may handle high-level
processing, and may comprise a math co-processor or other
processing devices.
[0183] A decision component category module 2220 and an application
category module 2222 may handle categories for various insurance
policies and decision components. As described above, each decision
component and each application may be assigned a category. The
decision component category module 2220 may include information
related to the category assigned for each decision component,
including a cross-reference Lo the application associated with each
decision component, the assigned category or categories, and/or
other information. The application category module 2222 may include
information related to the category assigned for each application,
including a cross-reference to the decision components associated
with each application, the assigned category or categories, and/or
other information.
[0184] An evaluation module 2224 may include an evaluation of a
decision component using one or more rules, where the rules may be
fuzzy logic rules. The evaluation module 2224 may direct the
application of one or more fuzzy logic rules to one or more
decision components. Further, the evaluation module 2224 may direct
the application of one or more fuzzy logic rules to one or more
policies within a case database 2240, to be described in greater
detail below. Evaluation module policies within a case database
2240, are to be described in greater detail below.
[0185] A measurement module 2226 may include measurements assigned
to one or more decision components. As described above, a
measurement may be assigned to each decision component based on an
evaluation, such as an evaluation with a fuzzy logic rule. The
measurement module 2226 may associate a measurement with each
decision component, direct the generation of the measurement,
and/or include information related to a measurement.
[0186] An issue module 2228 may handle issuing an insurance policy
based on the evaluation and measurements of one or more decision
components and the application itself. According to an embodiment
of the invention, decisions whether to ultimately issue an
insurance policy or not to issue an insurance policy may be
communicated to an applicant through the issue module 2228. The
issue module 2228 may associate issuance of an insurance policy
with an applicant, with various measurement(s) and evaluation(s) of
one or more policies and/or decision components and other
information.
[0187] A retrieval module 2230 may be responsible for retrieving
cases from a case database module 2240. According to an embodiment
of the invention, queries submitted by a user for case-based
reasoning may be coordinated through the retrieval module 2230 for
retrieving cases. Other information and functions related for case
retrieval may also be available.
[0188] A ranking module 2232 may be responsible for ranking cases
retrieved based on one or more queries received from a user.
According to an embodiment of the invention, the ranking module
2232 may maintain information related to cases and associated with
one or more queries. The ranking module 2232 may associate each
case with the ranking(s) associated with one or more queries. Other
information may also be associated with the ranking module
2232.
[0189] A rate class module 2234 may handle various designations of
rate classes for one or more insurance policies. According to an
embodiment of the invention, each application may be assigned a
rate class, where the premiums paid by the applicant are based on
the rate class. The rate class module 2234 may associate a rate
class with each insurance application, and may assign a rate class
based on evaluation and measurements of various applications and
decision components, as well as based on a decision by one or more
underwriters. Other information may also be associated with the
rate class module 2234.
[0190] Data may be stored in a data storage module 2236. The data
storage module 2236 stores a plurality of digital files. According
to an embodiment of the invention, a plurality of data storage
modules 2236 may be used and located on one or more data storage
devices, where the data storage devices are combined or separate
from the controller module 2212. One or more data storage modules
2236 may also be used to archive information.
[0191] An adaptation module 2238 may be responsible for adapting
the results of one or more queries to determine which previous
cases are most similar to the application for the present
application for insurance. Other information may also be associated
with the adaptation module 2238.
[0192] All cases used in a case based reasoning may be stored in a
case database module 2240. According to an embodiment of the
invention, a plurality of case database modules 2240 may be used
and located on one or more data storage devices, where the data
storage devices are combined or separate from the controller module
2212.
[0193] While the system 2200 of Pig. 22 discloses the requester
device 2205 connected to the network 2202, it should be understood
that a personal digital assistant ("FDA"), a mobile telephone, a
television, or another device that permits access to the network
2202 may be used to arrive at the system of the present
invention.
[0194] According to another embodiment of the invention, a
computer-usable and writeable medium having a plurality of computer
readable program code stored therein may be provided for practicing
the process of the present invention. The process and system of the
present invention may be implemented within a variety of operating
systems, such as a Windows.RTM. operating system, various versions
of a Unix-based operating system (e.g., a Hewlett Packard, a Red
Hat, or a Linux version of a Unix-based operating system), or
various versions of an AS/400-based operating system. For example,
the computer-usable and writeable medium may be comprised of a CD
ROM, a floppy disk, a hard disk, or any other computer-usable
medium. One or more of the components of the system 2200 may
comprise computer readable program code in the form of functional
instructions stored in the computer-usable medium such that when
the computer-usable medium is installed on the system 2200, those
components cause the system 2200 to perform the functions
described. The computer readable program code for the present
invention may also be bundled with other computer readable program
software.
[0195] According to one embodiment, the central controller module
2212, the transmitter module 2214, the receiver module 2216, the
processor module 2218, the decision component category module 2220,
application category module 2222, evaluation module 2224,
measurement module 2226, issue module 2228, retrieval module 2230,
ranking module 2232, rate class module 2234, data storage module
2236, adaptation module 2238, and case database module 2240 may
each comprise computer-readable code that, when installed on a
computer, performs the functions described above. Also, only some
of the components may be provided in computer-readable code.
[0196] Additionally, various entities and combinations of entities
may employ a computer to implement the components performing the
above-described functions. According to an embodiment of the
invention, the computer may be a standard computer comprising an
input device, an output device, a processor device, and a data
storage device. According to other embodiments of the invention,
various components may be computers in different departments within
the same corporation or entity. Other computer configurations may
also be used. According to another embodiment of the invention,
various components may be separate entities such as corporations or
limited liability companies. Other embodiments, in compliance with
applicable laws and regulations, may also be used.
[0197] According to one specific embodiment of the present
invention, the system may comprise components of a software system.
The system may operate on a network and may be connected to other
systems sharing a common database. Other hardware arrangements may
also be provided.
[0198] Other embodiments, uses and advantages of the present
invention will be apparent to those skilled in the art from
consideration of the specification and practice of the invention
disclosed herein. The specification and examples should be
considered exemplary only. The intended scope of the invention is
only limited by the claims appended hereto.
Information Summarization
[0199] The fuzzy rule-based decision engine and the case-based
decision engine may need to capture the medical/actuarial knowledge
required to evaluate and underwrite an application. They may do so
by using a rule set or a case base, respectively. However, both
decision engines may also need access to all the relevant
information that characterizes the new application. While the
structured component of this information can be captured as data
and stored into a database ("DB"), the free-form nature of an
attending physician statement (APS) may not be suitable to
automated parsing and interpretation. Therefore, for each
application requiring an APS, a summarization tool may be used that
will convert all the essential input variables from that statement
into a structured form, suitable for storage in a DB and for
supporting automated decision systems. Furthermore, if the decision
engines were not capable of handling this new application, then the
use of the APS summarization tool may be a productivity aid for a
human underwriter, rather than an automation tool.
[0200] The present invention may be used in connection with an
engine to automate decisions in business, commercial, or
manufacturing processes. Such an engine may be based on (but not
limited to) rules and/or cases. A process and system may be
provided to structure and summarize key information required by a
reasoning system. According to an embodiment of the invention,
summarized information required by a reasoning system may be used
to underwrite insurance applications, and establish a rate class
corresponding to the perceived risk of the applicant. Such risk may
be characterized by several information sources, such as, but not
limited to, the application form, the APS, laboratory data, medical
insurance consortium data bases, motor vehicle registration data
bases, etc. Once this information has been gathered and compiled,
the application risk may be evaluated by a human underwriter or by
an automated decision system. This evaluation is carried out
leveraging the medical and actuarial knowledge of the human
underwriter, which is captured in its essence by the automated
reasoning system. According to an embodiment of the invention, an
APS summarization tool may capture the relevant variables that
characterize a given medical impairment, allowing an automated
reasoning system to determine the degree of severity of such
impairment and to estimate the underlying insurance risk.
[0201] According to an embodiment of the invention, a focus of this
invention on the individual medical impairments of a patient may
provide 1) incremental deployment of the Automated Underwriting
system as summaries for new impairments can be developed and added;
2) efficient coverage, by addressing the most frequent impairments
first, according to a Pareto analysis of their frequencies; 3)
efficient description of the impairment, by including in the
summary only the variables that could have an impact on the
decision.
[0202] By way of example, an aspect of the present invention will
be described in terms of underwriting of an application for a fixed
life insurance policy. Although the description focuses on the use
of a reasoning system to automate the underwriting process of
insurance policies, it will be understood by one of ordinary skill
in the art that the applicability of this invention may be much
broader, as it may apply to other reasoning system
applications.
[0203] According to an embodiment of the invention, a method for
executing and manipulating an APS summarization tool may occur as
illustrated in FIG. 23. At step 2300, a summarizer with the
appropriate medical knowledge would log into a web-based system to
begin the summarization process. According to an embodiment of the
invention, the APS summarization system may include a general form
plus various condition specific forms, which are then filled out by
the summarizer. The summarizer may first fill out the general form,
which contains data fields relevant to all applicants. Condition
specific forms arc then filled out as needed, as the summarizer
discovers various features present m the APS being summarized.
[0204] At step 2302, a summarizer may verify that the APS
corresponds to the correct applicant. This may be done by matching
information on the APS itself with information about the applicant
provided by the system. By way of example, an applicant's name,
date of birth, and social security number could be matched. If a
match is not made, the summarizer may note this by checking the
appropriate checkbox. According to an embodiment of the invention,
at step 2304, failure to match an APS to an applicant would end the
summarizer's session for that applicant, and the summarizer may
recommend corrective action.
[0205] At step 2306, the general form is filled out. FIG. 24
illustrates a general form within a graphical user interface 2400
according to an embodiment of the invention. Graphical user
interface 2400 may comprise access to any network browser, such as
Netscape Navigator, Microsoft Explorer, or others. Other means of
accessing a network may also be used. Graphical user interface 2400
may include a control area 2402, whereby a summarizer may control
various aspects of graphical user interface 2400. Control may
include moving to various portions of the network via the graphical
user interface 2400, printing information from the network,
searching for information within the network, and other functions
used within a browser.
[0206] According to an embodiment of the invention, a general form
2400 may provide a fixed structure 2406 to capture the data within
the system. According to an embodiment of the invention, different
sections of the form may be organized into fields that are
structured to provide only a fixed set of choices for the
summarizer. This may be done to standardize the different pieces of
information contained in the APS. By way of example, a fixed set of
choices may be provided to a summarizer via a pull-down menu 2408.
For fields that cannot be treated as pull-down menus (e.g., dates,
numeric values of lab tests), such as entry field 2410 labeled as
"Initial date," validation may be performed to ensure that data
entry errors are minimized, and to check that values are within
allowable pre-determined limits. According to an embodiment of the
invention, validation may include a "client-side" validation,
designed to give the summarizer an immediate response if any of the
data is incorrectly entered. A "client-side" validation may be
achieved through JavaScript code embedded in the web pages.
According to an embodiment of the invention, validation may include
a "server-side" validation, which may be performed after data
submission. "Server-side" validation may be designed primarily as a
fail-safe check to prevent erroneous data from entering the
business-critical database.
[0207] According to an embodiment of the invention, link section
2404 may provide access to other portions of general form 2400. As
illustrated in FIG. 24, link section 2404 may include links (such
as hypertext links) to portions of general form 2400 that relate to
blood pressure, family history, nicotine use, build, lipids,
alcohol use, cardiovascular fitness and tests, final check,
comments, abnormal physical symptoms, abnormal blood results,
abnormal urine results, abnormal pap test, mammogram, abnormal
colonoscopy, chest x-ray, pulmonary function, substance abuse, and
non-medical history. Other information within a general form 2400
may also be provided, and as such, may be linked through link
section 2404.
[0208] According to an embodiment of the invention, an APS summary
may distinguish between a blank data field and answers such as
"don't know" or "not applicable," thereby ensuring the completeness
of the summary. For a general form submission, a final validation
pass may be performed at step 2308 to alert the summarizer if
certain required fields are blank. If required fields are blank,
the system may require a summarizer to return to step 2306 and
complete the general form. If the summarizer wishes to indicate
that the particular piece of information is not known, they may be
required to specifically indicate so, thereby maintaining
information about what information is specifically not known.
However, it will be recognized that not all fields will necessarily
require information. For example, certain fields may be
"conditionally mandatory," meaning that they require an answer only
if other fields have been filled out in a particular way. Use of
conditionally mandatory fields may ensure that all necessary
information is gathered. Further, ensuring that all required fields
have been filled may also ensure that the necessary information is
gathered.
[0209] When the general form has been filled out and validated at
step 2308, with all of the required fields entered, it may be
necessary to complete one or more condition-specific forms. At step
2310, it is determined if any condition-specific forms arc
required. If no condition specific forms are required, the results
may be submitted to a database or other storage device for use at a
later time at step 2320.
[0210] If a condition-specific form is required, a summarizer may
select a condition-specific form to fill-in at step 2312. According
to an embodiment of the invention, a summarizer may move from the
general form to any of the condition-specific forms by following a
hypertext link embedded within the general form. By way of example,
a link to a condition-specific form may be similar to, and/or same
as links located within link portion 2404. Further, links to
condition-specific forms may be located within link portion 2404. A
portion of the knowledge of which condition-specific forms are
necessary may be obtained while filling out the general form. In
the current example of life insurance underwriting, these
condition-specific forms may include hypertension, diabetes,
etc.
[0211] FIG. 25 illustrates an example of a condition-specific form
for hypertension within a graphical user interface 2500 according
to an embodiment of the invention. Graphical user interface 2500
may comprise access to any network browser, such as Netscape
Navigator, Microsoft Explorer, or other browser. Other manners of
accessing a network may also be used. Graphical user interface 2500
may include a control area 2502, whereby a summarizer may control
various aspects of graphic user interface 2500. Control may include
moving to various portions of the network via the graphic user
interface 2500, printing information from the network, searching
for information within the network, and other functions used within
a browser.
[0212] Graphical user interface 2500 displays the
hypertension-specific form, which may include various sections for
inputting information related to hypertension. In the hypertension
specific form illustrated in FIG. 25, initial identification
section 2504 may enable a summarizer to provide initial
identification information, including whether an applicant has
hypertension, the type of hypertension, whether it was secondary
hypertension, and if so, how the cause was removed or cured.
According to an embodiment of the invention, pull down menus may be
used to ensure that information entered is standardized for each
patient. Other information may also be gathered in initial
identification section 2504.
[0213] EKG section 2506 may enable a summarizer to provide EKG
information, including EKG readings within a specified time period
(e.g., 6 months), chest X-rays within a specified time period
(e.g., 6 months), and other information related to EKG readings.
According to an embodiment of the invention, pull down menus may be
used to ensure that information entered is standardized for each
patient. Patient cooperation section 2508 may enable a summarizer
to provide information related to a patient's cooperation,
including whether the patient has cooperated, whether a patient's
blood pressure is under control, and if so, for how many months,
and other information related to a patient's cooperation in dealing
with hypertension. According to an embodiment of the invention,
pull down menus may be used to ensure that information entered is
standardized for each patient.
[0214] Blood pressure section 2510 may enable a summarizer to enter
blood pressure readings corresponding to various dates. According
to an embodiment of the invention, separate entry fields may be
provided for the date the blood pressure reading was taken, (e.g.,
systolic reading (SBP) and the diastolic reading (DBP)). Other
information may also be entered in blood pressure section 2510.
Further, it will he understood by those skilled in the art that
other information related to hypertension may also be entered in a
hypertension form displayed on graphical user interface 2500.
[0215] At step 2314, a summarizer fills out a condition-specific
form. For a condition-specific form, a final validation pass may be
performed at step 2316 to alert the summarizer if certain required
fields are blank. If required fields are blank, the system may
require a summarizer to return to step 2314 and complete the
condition-specific form. As with a general form, if the summarizer
wishes to indicate that the particular piece of information is not
known, they may be required to specifically indicate so, thereby
facilitating the tracking of what information is specifically not
known. However, it will be recognized that not all fields will
necessarily require information. For example, certain fields may be
"conditionally mandatory," meaning that they require an answer only
if other fields have been filled out in a particular way. Use of
conditionally mandatory fields may ensure that all necessary
information is gathered. Further, ensuring that all required fields
have been filled may also ensure that the necessary information is
gathered.
[0216] If the condition-specific form has been filled out and
validated at step 2316, with all of the required fields entered, a
summarizer may determine if additional condition-specific forms are
necessary at step 2318. If additional condition-specific forms are
necessary, a summarizer may return to step 2312 and select the
appropriate condition-specific form in which to enter information.
If no additional condition-specific forms are required, the results
may be submitted to a database or other storage device for use at a
later time at step 2320.
[0217] Once the summarization is complete for a general form and
any selected condition-specific forms, the summarizer may submit
the results, such as described in step 2320. The data may then be
transferred over a network, such as the Internet, and stored in a
database for later use. According to an embodiment of the
invention, different categorical data fields may be presented to
the summarizer as text, but for space efficiency are encoded as
integer values in the database. A "translation table" to the
corresponding field meanings may then be provided as part of the
design of the APS summary. The APS summarizer may provide a
structured list of topics, thereby enabling a trained person to
summarize the most significant information currently contained in a
handwritten or typewritten APS. Further, the APS summarizer may
provide an efficient description of the data content of the APS. As
stated above, the APS itself can be several tens of pages of
doctor's notes. The APS summary is designed to capture only the
data fields that are relevant to the problem at hand. In addition,
a structured and organized description of the APS data may be
provided. An APS itself can adhere to any arbitrary order because
of different doctor's styles. The APS summary may provide a single
consistent format for the data as required for an automated system,
and/or which facilitates the human underwriter's job greatly.
[0218] Since the APS summary may be captured in a database, the
information contained in it may be easily available to any
computer-based application. Again, this is a requirement for an
automated underwriting system, but it may provide many other
advantages as well. For example, the APS data may otherwise be very
difficult to analyze statistically, to categorize, or to classify.
Since the APS summary forms can be web-based, the physical location
of the summarizers may be immaterial. The original APS sheets can
be received in location X, scanned, sent over the Internet to
location Y, where the APS summary is filled out, and the digital
data from the summary can be submitted and stored on a database
server in location Z. Further, the automated decision engine can be
in any fourth location, as could an individual running queries
against the APS summary database for statistical analysis or
reporting purposes.
[0219] According to an embodiment of the invention, general and
condition specific forms may be written in HTML and JavaScript,
which provide the validation functionality. A system for storing
filled out summary data into a remote database has also been
created. This system was created using JavaBeans and JSP. Testing
by experienced underwriters has been performed. The HTML summary
forms are displayed to the underwriters via a web browser, and the
data from an actual APS is entered onto the form. The underwriter
comments and feedback are captured on the form as well, and used to
aid the continual improvement of the forms. In choosing which
condition-specific forms to create, a statistical analysis was done
of the frequencies of the various medical conditions. The
conditions that are most frequent were chosen to be worked on
first. The APS summary does not have to cover all conditions before
it is put into production. Deployment of the APS summary may be
progressive, covering new conditions one by one as new forms become
available. Applicants with APS requirements that are not covered in
the current APS summary may be underwritten using the usual
procedures. Condition-specific forms may therefore be added to the
APS summary in order to increase coverage of applicants by the
digital underwriting system.
Optimization of Fuzzy Rule-Based and Case-Based Decision
Engines
[0220] According to an embodiment of the present invention, fuzzy
rule-based and case-based reasoning may be used to automate
decisions in business, commercial, or manufacturing process.
Specifically, a process and system to automate the determination of
optimal design parameters that impact the quality of the output of
the decision engines is described.
[0221] According to an embodiment of the invention, the
optimization aspect may provide a structured and robust search and
optimization methodology for identifying and tuning the decision
thresholds (cutoffs) of the fuzzy rules and internal parameters of
the fuzzy rule-based decision engine ("RBE"), and the internal
parameters of the case-based decision engine ("CBE"). These
benefits may include a minimization of the degree of rate class
assignment mismatch between that of an expert human underwriter and
automated rate class decisions. Further, the maintenance of the
accuracy of rule-based and case-based decision-making as decision
guidelines evolve with time may be achieved. In addition,
identification of ideal parameter combinations that govern the
automated decision-making process may occur.
[0222] The system and process of the present invention may apply to
a class of stochastic global search algorithms known as
evolutionary algorithms to perform parameter identification and
tuning. Such algorithms may be executed utilizing principles of
natural evolution and may be robust adaptive search schemes
suitable for searching non-linear, discontinuous, and
high-dimensional spaces. Moreover, this tuning approach may not
require an explicit mathematical description of the
multi-dimensional search space. Instead, this tuning approach may
rely solely on an objective function that is capable of producing a
relative measure of alternative solutions. According to an
embodiment of the invention, an evolutionary algorithm may be used
for optimization within an RBE and CBE. By way of example, an
evolutionary algorithm ("EA") may include genetic algorithms,
evolutionary programming, evolution strategies, and genetic
programming. The principles of these related techniques may define
a general paradigm that is based on a simulation of natural
evolution. EAs may perform their search by maintaining at any time
t a population P(t)={P.sub.1(t), P.sub.2(t), . . . , P.sub.p(t)} of
individuals. In this example, "genetic" operators that model
simplified rules of biological evolution are applied to create the
new and desirably more superior population P(t+1). Such a process
may continue until a sufficiently good population is achieved, or
some other termination condition is satisfied. Each Pi(t) .di-elect
cons. P(t), represents via an internal data structure, a potential
solution to the original problem. The choice of an appropriate data
structure for representing solutions may be more an "art" than a
"science" due to the plurality of data structures suitable for a
given problem. However, the choice of an appropriate representation
may be a critical step in a successful application of EAs. Effort
may be required to select a data structure that is compact,
minimally superfluous, and can avoid creation of infeasible
individuals. For instance, if the problem domain requires finding
an optimal real vector from the space defined by dissimilarly
bounded real coordinates, it may be more appropriate to choose as a
representation a real-set-array (e.g., bounded sets of real
numbers) instead of a representation capable of generating bit
strings. A representation that generates bit strings may create
many infeasible individuals, and can be certainly longer than a
more compact sequence of real numbers. Closely linked to a choice
of representation of solutions may be a choice of a fitness
function .psi.: P(t).fwdarw.R, that assigns credit to candidate
solutions. Individuals in a population are assigned fitness values
according to some evaluation criterion. Fitness values may measure
how well individuals represent solutions to the problem. Highly fit
individuals are more likely to create offspring by recombination or
mutation operations. Weak individuals are less likely to be picked
for reproduction, so they eventually die out. A mutation operator
introduces genetic variations in the population by randomly
modifying some of the building blocks of individuals. Evolutionary
algorithms are essentially parallel by design, and at each
evolutionary step a breadth search of increasingly optimal
sub-regions of the options space is performed. Evolutionary search
is a powerful technique of solving problems, and is applicable to a
wide variety of practical problems that are nearly intractable with
other conventional optimization techniques. Practical evolutionary
search schemes do not guarantee convergence to the global optimum
in a predetermined finite time, but they are often capable of
finding very good and consistent approximate solutions. However,
they are shown to asymptotically converge under mild
conditions.
[0223] An evolutionary algorithm may be used within a process and
system for automating the tuning and maintenance of fuzzy
rule-based and case-based decision systems used for automated
decisions in insurance underwriting. While this approach is
demonstrated for insurance underwriting, it is broadly applicable
to diverse rule-based and case-based decision-making applications
in business, commercial, and manufacturing processes. Specifically,
we describe a structured and robust search and optimization
methodology based on a configurable multi-stage evolutionary
algorithm for identifying and tuning the decision thresholds of the
fuzzy rules and internal parameters of the fuzzy rule-based
decision engine and the internal parameters of the case-based
decision engine. The parameters of the decision systems impact the
quality of the decision-making, and are therefore critical.
Furthermore, this tuning methodology can be used periodically to
update and maintain the decision engines.
[0224] As stated above, these fuzzy logic systems may have many
parameters that can be freely chosen. These parameters may either
be fit to reproduce a given set of decisions, or set by management
in order to achieve certain results, or a combination of the two. A
large set of cases may be provided by the company as a "certified
case base." According to an embodiment of the invention, the
statistics of the certified case base may closely match the
statistics of insurance applications received in a reasonable time
window. According to an embodiment of the invention, there will be
man y more cases than free parameters, so that the system will be
over-determined. Then, an optimal solution may be found which
minimizes the classification error between a decision engine's
output and the supplied cases. When considering maintenance of a
system, it may be convenient and advantageous that the parameters
are chosen using optimization vs. a set of certified cases. New
fuzzy rules and certified cases may be added, or aggregation rules
may change. The fuzzy logic systems may be kept current, allowing
the insurance company to implement changes quickly and with zero
variability.
[0225] The parameter identification and tuning problem which may
presented in this invention can be mathematically described as a
minimization problem:
min .psi.(x) where .chi.=.chi..sub.1.times..chi..sub.2.times. . . .
.times..chi..sub.n.chi..sub.i.OR right. and
.psi.:.chi..fwdarw..sub.+X .di-elect cons. .chi.
where .chi. is an n-dimensional bounded hyper-volume (parametric
search space) i n the n-dimensional space of reals, x is a
parameter vector, and If/ is the objective function that maps the
parametric search space to the non-negative real line.
[0226] FIG. 26 illustrates such a minimization (optimization)
problem according to an embodiment of the invention in the context
of the application domain, where the search space x corresponds to
the space of decision engine designs induced by the parameters
imbedded in the decision engine, and the objective function If/
measures the corresponding degree of rate-class assignment mismatch
between that of the expert human underwriter and the
decision-engine for the certified case base. An evolutionary
algorithm iteratively generates trial solutions (trial parameter
vectors in the space x), and uses their corresponding consequent
degree of rate-class assignment mismatch as the search feedback.
Thus, at step 2602, a space of decision engine's designs is probed.
At step 2604, a mismatch matrix, which will be described in greater
detail below, is generated based on the rate-class decisions
generated for the cases by the decision engine. Penalties for
mismatching cases are assigned at step 2606. The evolutionary
algorithm uses the corresponding degree of rate-class assignment
mismatches, and the associated penalties to provide feedback to the
decision engine at step 2608. The system may then refine the
internal parameters and decision thresholds in the decision engine
at step 2602, and proceed through the process again. Thus, an
iterative process may be performed.
[0227] FIG. 27 illustrates an example of an encoded population
maintained by the evolutionary algorithm at a given generation.
According to an embodiment of the invention, each individual in the
population is a trial vector of design parameters representing
fuzzy rule thresholds and internal parameters of the decision
engine. Each percentage entry may represent a value of a trial
parameter that falls within a corresponding bounded real line. Each
trial solution vector may be used to initialize an instance of the
decision engine, following which each of the cases in the certified
case base is evaluated.
[0228] FIG. 28 illustrates a process schematic for an evaluation
system according to an embodiment of the invention. Trial design
parameters are provided at an input module 2802. The trial design
parameters are automatically input to decision engine 2804. Case
subset 2808 from certified case base 2806 is input into decision
engine 2804. Certified case base 2806 may comprises cases that have
been certified as being correct. Case subset 2808 may be a
predetermined number of cases from certified case base 2806.
According to an embodiment of the invention, case subset 2808 may
comprise two thousand (2000) certified cases. According to an
embodiment of the invention, case subset 2808 may comprise a number
of times the number of tunable parameters of decision engine 2804.
The cases within case subset 2808 are processed in decision engine
2804, and output to decision engine case decisions 2810.
[0229] Once all the cases i n the certified case base are
evaluated, a square confusion matrix 2814 is created. According to
an embodiment of the invention, confusion matrix 2814 may be
generated by comparing decision engine case decisions 2810 and
certified case decisions 2812. The rows of confusion matrix 2814
may correspond to certified case decisions 2812 as determined by an
expert human underwriter, and the columns of confusion matrix 2814
may correspond to the decision engine case decisions 2810 for the
cases in the certified case base. By way of example, assume a case
has been assigned a category S from certified case decision 2812
(from the matrix 2814) and a category PB from decision engine
decision 2810. Under these categorizations, the case would count
towards an entry in the cell at row 3 and column L In this example,
the certified case decision 2812 places the case in a higher risk
category, while the decision engine case decision 2810 places the
case in a lower risk category. Therefore, for this particular case,
the decision engine 2810 has been more liberal in decision-making.
By way of another example, if on the other hand both the certified
case decision 2812 and the decision engine case decisions 2810
agree as upon categorizing the case in class S, then the case would
count towards an entry in the cell at row 3 and column 3. By way of
another example, if the certified case decision 2812 is PB, but the
machine decision 2810 is S, then clearly the machine decision is
more strict.
[0230] According to an embodiment of the invention, it may be
desirable lo use a decision engine that is able to place the
maximum number of certified cases along the main diagonal of
confusion matrix 2814. It may also be desirable to determine those
parameters 2802 for decision engine 2804 that produce such results
(e.g., minimize the degree of rate class assignment confusion or
mismatch between certified case decisions 2812 and decision engine
case decisions 2810). Confusion matrix 2814 may be used as the
foundation to compute an aggregate mismatch penalty or score, using
penalty module 2816. According to an embodiment of the invention, a
penalty matrix may be derived from actuarial studies and is
element-by-element multiplied with the cells of the confusion
matrix 2814 to generate an aggregate penalty/score for a trial
vector of parameters in the evolutionary search. A summation over
the number of rows and columns of the matrix may occur, and that
should now be "T" (upper case T), as the confusion matrix M may be
of a dimension T.times.T. Other process systems may also be used to
achieve the present invention.
[0231] According to an embodiment of the invention, an evolutionary
algorithm may utilize only the selection and stochastic variation
(mutation) operations to evolve generations of trial solutions.
While the selection operation may seek to exploit known search
space regions, the mutation operation may seek to explore new
regions of the search space. Such an algorithm is known to those of
ordinary skill in the art. One example of the theoretical
foundation for such an algorithm class appears in Modeling and
Convergence Analysis of Distributed Coevolutionary Algorithms, Raj
Subbu and Arthur C. Sanderson, Proceedings of the IEEE
International Congress on Evolutionary Computation, 2000.
[0232] FIG. 29 illustrates an example of the mechanics of such an
evolutionary process. At step 2902, an initial population of trial
decision engine parameters is created. Proportional selection
occurs at step 2904 and an intermediate population is created at
step 2906. Stochastic variation occurs at step 2908, and a new
population is created at step 2910. The new population may then be
subject to proportional selection at step 2904, thereby creating an
iterative process.
[0233] According to an embodiment of the invention, the
evolutionary algorithm may use a specified fixed population size
and operate in one or more stages, each stage of which may be user
configurable. A stage is specified by a tuple consisting of a fixed
number of generations and normalized spread of a Gaussian
distribution governing randomized sampling. A given solution (also
called the parent) in generation i may be improved by cloning it to
create two identical child solutions from the parent solution.
[0234] The first child solution may be mutated according to a
uniform distribution within the allowable search bounds. The second
child solution may then be mutated according to the Gaussian
distribution for generation i. If the mutated solution falls
outside of the allowable search bounds, then the sampling is
repeated a few times until an acceptable sample is found. If no
acceptable sample is found within the allotted number of trials,
then the second child solution may be mutated according to a
uniform distribution. The best of the parent and two child
solutions is retained and is transferred to the population at
generation i+1. In addition, it is ensured via elitism that the
improvement in the best performing individual of each generation of
evolution i+n (where n is an increasing whole number) is a monotone
function. According to an embodiment of the invention, the process
may be repeated until i+n generation has been generated, where i+n
is a whole number.
[0235] While the invention has been particularly shown and
described within the framework of an insurance underwriting
application, it will be appreciated that variations and
modifications can be effected by a person of ordinary skill in the
art without departing from the scope of the invention. For example,
one of ordinary skill in the art will recognize that the fuzzy
rule-based or case-based engine of this invention can be applied to
any other transaction-oriented process in which underlying risk
estimation is required to determine the price structure (premium,
price, commission, etc.) of an offered product, such as insurance,
re-insurance, annuities, etc. Furthermore, the determination of the
confidence factor and the optimization of the decision engines
transcend the scope of insurance underwriting. A confidence factor
obtained in the manner described in this document could be
determined from any application of a case-based reasoner (whether
it is fuzzy or not). Similarly, the engine optimization process
described in this document can be applied to any engine in which
the structure of the engine has been defined and the parametric
values of the engine need to be specified to meet a predefined
performance metric. Furthermore, one of ordinary skill in the art
will recognize that such decision engines do not need to be
restricted to insurance underwriting applications.
* * * * *