U.S. patent application number 17/339583 was filed with the patent office on 2022-03-24 for discovering genomes to use in machine learning techniques.
The applicant listed for this patent is Analytics For Life Inc.. Invention is credited to Timothy Burton, Abhinav Doomra, Paul Grouchy, Sunny Gupta, Ali Khosousi, Ian Shadforth.
Application Number | 20220093215 17/339583 |
Document ID | / |
Family ID | 1000006013204 |
Filed Date | 2022-03-24 |
United States Patent
Application |
20220093215 |
Kind Code |
A1 |
Grouchy; Paul ; et
al. |
March 24, 2022 |
DISCOVERING GENOMES TO USE IN MACHINE LEARNING TECHNIQUES
Abstract
A facility for identifying combinations of feature and machine
learning algorithm parameters, where each combination can be
combined with one or more machine learning algorithms to train a
model, is disclosed. The facility evaluates each genome based on
the ability of a model trained using that genome and a machine
learning algorithm to produce accurate results when applied to a
validation data set by, for example, generating a fitness or
validation score for the trained model and the corresponding genome
used to train the model. Genomes that produce fitness scores that
exceed a fitness threshold are selected for mutation, mutated, and
the process is repeated. These trained models can then be applied
to new data to generate predictions for the underlying subject
matter.
Inventors: |
Grouchy; Paul; (Toronto,
CA) ; Burton; Timothy; (Ottowa, CA) ;
Khosousi; Ali; (Toronto, CA) ; Doomra; Abhinav;
(North York, CA) ; Gupta; Sunny; (Toronto, CA)
; Shadforth; Ian; (Morrisville, NC) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Analytics For Life Inc. |
Toronto |
|
CA |
|
|
Family ID: |
1000006013204 |
Appl. No.: |
17/339583 |
Filed: |
June 4, 2021 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
15653441 |
Jul 18, 2017 |
11062792 |
|
|
17339583 |
|
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06N 3/126 20130101;
G06F 16/00 20190101; G16Z 99/00 20190201; G06N 20/10 20190101; G06N
3/086 20130101; G06N 20/20 20190101; G16B 40/00 20190201; G06N
5/003 20130101; G16B 5/00 20190201; G01N 33/50 20130101 |
International
Class: |
G16B 40/00 20060101
G16B040/00; G16B 5/00 20060101 G16B005/00; G06N 20/20 20060101
G06N020/20; G06N 3/12 20060101 G06N003/12; G06F 16/00 20060101
G06F016/00; G01N 33/50 20060101 G01N033/50 |
Claims
1. A system, having a memory and a processor, for discovering
machine learning genomes, the system comprising: a first component
configured to generate a plurality of genomes, wherein each genome
identifies at least one feature and at least one parameter for at
least one machine learning algorithm, wherein generating a first
genome of the plurality of genomes comprises: randomly selecting,
from among a set of features, one or more of the features, randomly
selecting, from among a set of parameters for at least one machine
learning algorithm, one or more of the parameters, and assigning at
least one random value to each of the selected parameters; a second
component configured to, for each generated genome, train a one or
more models using the generated genome, and for each model trained
using the generated genome, calculate a fitness score for the
trained model at least in part by applying the trained model to a
validation data set, and produce a fitness score for the generated
genome based at least in part on the fitness scores generated for
the models trained using the generated genome; a third component
configured to identify, from among the generated genomes, a
plurality of genomes having a fitness score that exceeds a fitness
threshold; and a fourth component configured to, for each of the
identified genomes, mutate the identified genome, wherein at least
one of the components comprises computer-executable instructions
stored in the memory for execution by the system.
2. The system of claim 1, further comprising: a fifth component
configured to, for the first genome comprising a first set of
features, identify correlated features from among the first set of
features at least in part by: for each feature of the first set of
features, applying a feature generator associated with the feature
to a training set of data to generate a feature vector for the
feature, for at least one pair of feature vectors, calculating a
distance between each feature vector of the pair of feature
vectors, determining that the calculated distance is less than a
distance threshold, in response to determining that the calculated
distance is less than a distance threshold, removing, from the
first genome, a feature corresponding to at least one feature
vector of the pair of feature vectors, wherein each feature vector
includes, for each of a plurality of patients, a single value
generated by applying a first feature generator to at least one
representation of physiological data representative of the
patient.
3. The system of claim 2, wherein the removing, from the first
genome, of at least one feature corresponding to at least one
feature vector of a first pair of feature vectors comprises:
randomly selecting one of feature vector of the first pair of
feature vectors, identifying, from among the features of the first
genome, a feature corresponding to the randomly selected feature
vector; and removing, from the first genome, the identified
feature.
4. The system of claim 1, further comprising: a fifth component
configured to, for the first genome comprising a first set of
features, generate a graph comprising a vertex for each feature of
the first set of features; a sixth component configure to generate
an edge between vertices whose corresponding features have a
correlation value that exceeds a correlation threshold or a
distance value that is less than a distance threshold; and a
seventh component configured to remove vertices from the graph
until no connected vertices remain in the graph.
5. The system of claim 1, further comprising: a machine configured
to receive physiological signal data from at least one patient; a
fifth component configured to, for each patient, apply at least one
of the trained models to at least a portion of the physiological
signal data received for the patient by the machine, and generate a
prediction for the patient based at least in part on the
application of the at least one of the trained models to at least a
portion of the received physiological signal.
6. A method, performed by a computing system having a memory and a
processor, for discovering machine learning genomes, the method
comprising: generating, with the processor, a plurality of genomes,
wherein each genome identifies at least one feature and at least
one parameter for at least one machine learning algorithm; for each
generated genome, training at least one model using the generated
genome, and producing a fitness score for the genome based at least
in part on the trained at least one model; identifying, from among
the generated genomes, at least one genome having a fitness score
that exceeds a fitness threshold; and mutating each identified
genome.
7. The method of claim 6, wherein generating a first genome of the
plurality of genomes comprises: randomly selecting, from among a
set of features, one or more of the features; randomly selecting,
from among a set of parameters for at least one machine learning
algorithm, one or more of the parameters; and assigning at least
one value to each of the selected parameters.
8. The method of claim 7, wherein generating the first genome
further comprises: for each feature of the randomly selected
features, retrieving a feature vector for the feature based at
least in part on a feature generator associated with the feature
and a training set of data; identifying pairs of correlated feature
vectors from among the generated feature vectors; and for each
identified pair of correlated feature vectors, identifying one
feature vector of the pair of correlated feature vectors, removing,
from the first genome, the feature associated with the feature
generator used to generate the identified feature vector; randomly
selecting, from among the set of features, a feature to add to the
first genome, and adding the randomly selected feature to the first
genome.
9. The method of claim 8, wherein identifying pairs of correlated
feature vectors comprises: for each pair of feature vectors,
calculating a distance metric for the pair of feature vectors, and
determining whether the distance metric calculated for the pair of
feature vectors is less than a distance threshold, wherein the
distance threshold is determined based at least in part on the
calculated distance metrics determined for each pair of feature
vectors.
10. The method of claim 6, wherein producing a fitness score for a
first genome comprises: identifying a number of false positives
generated by applying, to two or more validation data sets, a model
trained using the first genome; and identifying a number of false
negatives generated by applying, to two or more validation data
sets, a model trained using the first genome.
11. The method of claim 6, wherein producing a fitness score for a
first genome comprises: generating, for at least one model trained
using the first genome, a receiver operating characteristic curve;
and calculating an area under the generated receiver operating
characteristic curve.
12. The method of claim 6, wherein producing a fitness score for a
first genome comprises calculating, for at least one model trained
using the first genome, one or more of the errors selected from the
group comprising: mean squared prediction error, mean absolute
error, interquartile error, and log loss error, receiver-operator
characteristic curve error, and f-score error.
13. The method of claim 6, wherein mutating a first identified
genome comprises: selecting at least one feature of the first
identified genome; and removing, from the first identified genome,
each of the selected features of the first identified genome.
14. The method of claim 6, wherein mutating the first identified
genome further comprises: randomly selecting, from among the set of
features, a plurality of the features; and adding, to the first
identified genome, each of the randomly selected plurality of
features.
15. The method of claim 6, wherein mutating a first identified
genome comprises: modifying at least one feature of the first
identified genome.
16. The method of claim 5, wherein mutating a first identified
genome comprises: modifying at least one machine learning algorithm
parameter of the first identified genome.
17. A computer-readable medium storing instructions that, if
executed by a computing system having a memory and a processor,
cause the computing system to perform a method for discovering
machine learning genomes, the method comprising: generating a
plurality of genomes, wherein each genome identifies at least one
feature; for each generated genome, training at least one model
using the generated genome, and producing a fitness score for the
genome based at least in part on the trained at least one model;
and identifying, from among the generated genomes, one or more
genomes having a fitness score that exceeds a fitness
threshold.
18. The computer-readable medium of claim 17, wherein each genome
further identifies at least one parameter for at least one machine
learning algorithm.
19. The computer-readable medium of claim 17, the method further
comprising: mutating each identified genome having a fitness score
that exceeds the fitness threshold.
20. The computer-readable medium of claim 17, the method further
comprising: computing the fitness threshold at least in part by,
determining an overall fitness score based on the fitness scores
produced for each of the generated genomes.
21-26. (canceled)
Description
CROSS-REFERENCE TO RELATED APPLICATION(S)
[0001] This application is a continuation of U.S. patent
application Ser. No. 15/653,441, filed on July 18, 2017, entitled
"DISCOVERING GENOMES TO USE IN MACHINE LEARNING TECHNIQUES," which
is incorporated by reference herein in its entirety. This
application is related to U.S. patent application Ser. No.
13/970,580, filed on Aug. 19, 2013, entitled "NON-INVASIVE METHOD
AND SYSTEM FOR CHARACTERIZING CARDIOVASCULAR SYSTEMS," now U.S.
Pat. No. 9,289,150; U.S. patent application Ser. No. 15/061,090,
filed on Mar. 4, 2016, entitled "NON-INVASIVE METHOD AND SYSTEM FOR
CHARACTERIZING CARDIOVASCULAR SYSTEMS," now U.S. Pat. No. 9,655,536
U.S. patent application Ser. No. 15/588,148, filed on May 5, 2017,
entitled "NON-INVASIVE METHOD AND SYSTEM FOR CHARACTERIZING
CARDIOVASCULAR SYSTEMS," now U.S. Pat. No. 9,968,275; U.S. patent
application Ser. No. 13/605,364, filed on Sep. 6, 2012, entitled
"SYSTEM AND METHOD FOR EVALUATING AN ELECTROPHYSIOLOGICAL SIGNAL,"
now U.S. Pat. No. 8,923,958; U.S. patent application Ser. No.
13/970,582, filed on Aug. 19, 2013, entitled "NON-INVASIVE METHOD
AND SYSTEM FOR CHARACTERIZING CARDIOVASCULAR SYSTEMS FOR ALL-CAUSE
MORTALITY AND SUDDEN CARDIAC DEATH RISK," now U.S. Pat. No.
9,408,543; U.S. patent application Ser. No. 15/207,214, filed on
Jul. 11, 2016, entitled "NON-INVASIVE METHOD AND SYSTEM FOR
CHARACTERIZING CARDIOVASCULAR SYSTEMS FOR ALL-CAUSE MORTALITY AND
SUDDEN CARDIAC DEATH RISK," now U.S. Pat. No. 9,955,883; U.S.
patent application Ser. No. 14/295,615, filed on Jun. 4, 2014,
entitled "NONINVASIVE ELECTROCARDIOGRAPHIC METHOD FOR ESTIMATING
MAMMALIAN CARDIAC CHAMBER SIZE AND MECHANICAL FUNCTION," now U.S.
Pat. No. 9,737,229; U.S. patent application Ser. No. 14/077,993,
filed on Nov. 12, 2013, entitled "NONINVASIVE ELECTROCARDIOGRAPHIC
METHOD FOR ESTIMATING MAMMALIAN CARDIAC CHAMBER SIZE AND MECHANICAL
FUNCTION," now U.S. Pat. No. 10,039,468; U.S. patent application
Ser. No. 14/596,541, filed on Jan. 14, 2015, entitled "NONINVASIVE
METHOD FOR ESTIMATING GLUCOSE, GLYCOSYLATED HEMOGLOBIN AND OTHER
BLOOD CONSTITUENTS," now U.S. Pat. No. 9,597,021; U.S. patent
application Ser. No. 15/460,341, filed on Mar. 16, 2017, entitled
"NONINVASIVE METHOD FOR ESTIMATING GLUCOSE, GLYCOSYLATED HEMOGLOBIN
AND OTHER BLOOD CONSTITUENTS," now U.S. Pat. No. 10,765,350; U.S.
patent application Ser. No. 14/620,388, filed on Feb. 12, 2015,
entitled "METHOD AND SYSTEM FOR CHARACTERIZING CARDIOVASCULAR
SYSTEMS FROM SINGLE CHANNEL DATA," now U.S. patent application Ser.
No. 14/620,388; U.S. patent application Ser. No. 15/192,639, filed
on Jun. 24, 2016, entitled "METHODS AND SYSTEMS USING MATHEMATICAL
ANALYSIS AND MACHINE LEARNING TO DIAGNOSE DISEASE," now U.S. Pat.
No. 9,910,964; U.S. patent application Ser. No. 15/248,838, filed
on Aug. 26, 2016, entitled "BIOSIGNAL ACQUISITION DEVICE," now U.S.
Pat. No. 10,542,897; U.S. Provisional Patent Application No.
62/397,895, filed on Sep. 21, 2016, entitled "GRAPHICAL USER
INTERFACE FOR CARDIAC PHASE-SPACE TOMOGRAPHY," U.S. patent
application Ser. No. 15/633,330, filed Jun. 26, 2017, entitled
"NON-INVASIVE METHOD AND SYSTEM FOR MEASURING MYOCARDIAL ISCHEMIA,
STENOSIS IDENTIFICATION, LOCALIZATION AND FRACTIONAL FLOW RESERVE
ESTIMATION," now U.S. Pat. No. 10,362,950; and U.S. patent
application Ser. No. 15/653,433 (Attorney Docket No.
124077-8001.US00), filed on Jul. 18, 2017, entitled "DISCOVERING
NOVEL FEATURES TO USE IN MACHINE LEARNING TECHNIQUES, SUCH AS
MACHINE LEARNING TECHNIQUES FOR DIAGNOSING MEDICAL CONDITIONS."
Each of the above-identified applications and issued patents is
hereby incorporated by reference in its entirety.
BACKGROUND
[0002] Machine learning techniques predict outcomes based on sets
of input data. For example, machine learning techniques are being
used to predict weather patterns, geological activity, provide
medical diagnoses, and so on. Machine learning techniques rely on a
set of features generated using a training set of data (i.e., a
data set of observations, in each of which an outcome to be
predicted is known), each of which represents some measurable
aspect of observed data, to generate and tune one or more
predictive models. For example, observed signals (e.g., heartbeat
signals from a number of subjects) may be analyzed to collect
frequency, average values, and other statistical information about
these signals. A machine learning technique may use these features
to generate and tune a model that relates these features to one or
more conditions, such as some form of cardiovascular disease (CVD),
including coronary artery disease (CAD), and then apply that model
to data sources with unknown outcomes, such as an undiagnosed
patient or future weather patterns, and so on. Conventionally,
these features are manually selected and combined by data
scientists working with domain experts.
BRIEF DESCRIPTION OF THE DRAWINGS
[0003] FIG. 1 is a block diagram illustrating an environment in
which the facility operates in some embodiments.
[0004] FIG. 2 is a flow diagram illustrating the processing of a
discover features component in some embodiments.
[0005] FIG. 3 is a flow diagram illustrating the processing of a
process component in some embodiments.
[0006] FIG. 4 is a flow diagram illustrating the processing of an
apply feature generators component in some embodiments.
[0007] FIG. 5 is a flow diagram illustrating the processing of an
identify novel feature vectors component in some embodiments.
[0008] FIG. 6 is a flow diagram illustrating the processing of a
discover genomes component in accordance with some embodiments.
[0009] FIG. 7 is a flow diagram illustrating the processing of a
spawn genome component in accordance with some embodiments.
[0010] FIG. 8 is a flow diagram illustrating the processing of an
identify high-performing genomes component in accordance with some
embodiments.
DETAILED DESCRIPTION
[0011] Because machine learning techniques rely on features and/or
combinations of features, the process of feature selection and
combination typically is an important part of a machine learning
process. Moreover, because a large number of diverse machine
learning algorithms exist (e.g., decision trees, artificial neural
networks (ANNs), deep ANNs, genetic (and meta-genetic) algorithms,
and so on), the choice of algorithm and any associated parameters
can also be important. For example, different machine learning
algorithms (or family of machine learn algorithms) may be best
suited for different types of data and/or the types of predictions
to be made. Furthermore, different machine learning algorithms may
present various tradeoffs with respect to resources (e.g., memory,
processor utilization), speed, accuracy, and so on. Typically,
models are trained using machine learning algorithms, features, and
parameters selected by individuals based on the preferences of
those individuals and/or criteria specified by those individuals.
The inventors have recognized that it can be expensive and
time-consuming manually to identify features, machine learning
algorithms, and corresponding parameters and even more difficult to
produce features, machine learning algorithms, and corresponding
parameters that produce more accurate models and, therefore, more
accurate predictions. Accordingly, the inventors have conceived and
reduced to practice a facility that performs automatic discovery of
combinations of features, machine learning algorithms, and/or
machine learning parameters.
[0012] In some embodiments, the facility operates as part of a
machine learning pipeline that constructs and evaluates predictive
models, such as those for disease diagnosis, based on time-series
and/or other signals, such as physiological signals. The machine
learning process uses features to identify patterns within a
training set of data and, based on these patterns, generates
predictive models. These predictive models can be validated using
validation data sets (i.e., data sets for which an outcome is known
but that were not used to train the model) and applied to new input
data in order to predict outcomes from the input data, such as
providing a diagnosis for a medical condition, etc. As new data and
new features are produced or acquired, the machine learning process
improves upon the predictive capabilities of these models by
incorporating new features and, in some cases, discarding others,
such as those that are determined to be too similar to other
features.
[0013] In particular, the facility seeks to identify combinations
of features and machine learning algorithm parameters where each
combination can be used to train one or more models. A combination
of features and/or machine learning parameters is sometimes
referred herein to as a "genome." The facility evaluates each
genome based on the ability of a model trained using a machine
learning algorithm and that genome to produce accurate results when
applied to a validation data set by, for example, generating a
fitness or validation score for the trained model and the
corresponding genome used to train the model. In some cases, the
facility uses the validation score as a fitness score while in
other cases the validation score is an element of a fitness score
(e.g., fitness score=training score+validation score). In some
cases, multiple models may be trained using a genome and the
resulting fitness scores can be aggregated to generate an
aggregated fitness score for the genome.
[0014] By way of example, the facility for identifying combinations
of features and machine learning algorithm parameters can be used
for a medical diagnosis predictive modeling task. In this example,
the facility receives, for a number of patients or subjects, one or
more sets of physiological data that relate to some type of
physiological output or condition of the patient over a period of
time (e.g., less than a second, on the order of a few seconds,
about ten seconds, about 30 seconds and up to about five minutes,
about an hour or more, etc.), such as electroencephalograms, and so
on. These data may be received in real-time or near real-time,
concurrent or nearly concurrent with the operation of the facility,
or they may be received at an earlier time. In some cases, the
facility discards certain portions of the signal to ensure that the
signals from each patient commence at a stable and consistent
initial condition. Furthermore, the data may be normalized to
remove potentially misleading information. For example, the
facility can normalize the amplitude of signal data (e.g.,
transforming to a z-score), to account for variations in signal
strength caused by sensor contact or other non-physiological data.
As another example, in the case of a cardiac signal, the facility
can perform a peak search and discard any data before a first
heartbeat identified in the signal and after a last heartbeat
identified in the signal.
[0015] In some embodiments, the facility applies a set of feature
generators to a set of signals to generate, for each combination of
a signal and a feature generator, a feature value for the signal.
Thus, each feature value is representative of some property of the
underlying signal data. In one example, the facility receives
patient data for each of 1000 patients and applies one or more
feature generators to the data to generate, for each application of
a feature generator to the data of a single patient, a feature
value (or set of feature values). The facility collects the feature
values generated by a single feature generator in a "feature
vector," such that the feature vector stores one feature value per
patient. Once the feature vectors are generated, they can be
compared to determine how different each is relative to each of the
other feature vectors. The facility computes a distance metric for
each feature vector to assess the novelty of the corresponding
feature generator. Based on the assessed novelty, the facility (1)
provides the feature generators that produced the novel feature
vectors to the machine learning process for the purpose of basing
new predictive models on the provided feature generators and (2)
modifies these feature generators to create a new generation of
feature generators. The facility repeats this evolutionary process
to identify even more novel features for use by the machine
learning process.
[0016] In some embodiments, for each received set of data, the
facility computes or identifies separate sets of one or more values
from the data. For example, in the case of data generated as part
of an electrocardiogram, the facility identifies global and local
maxima and minima within the data, computes frequency/period
information from the data, calculates average values of the data
over a certain period of time (e.g., the average duration and
values generated during a QRS complex), and so on. In some cases,
the facility transforms the received data and extracts sets of one
or more values from the transformed data. The facility can
transform received signal data in any number of ways, such as
taking one or more (successive) derivatives of the data, taking one
or more partial derivatives of the data, integrating the data,
calculating the gradient of the data, applying a function to the
data, applying a Fourier transform, applying linear or matrix
transformations, generating topology metrics/features, generating
computational geometry metrics/features, generating differential
manifold metrics/features, and so on. In this manner, the facility
generates multiple perspectives of the data in order to yield a
diverse set of features. While these transformations are provided
by way of example, one of ordinary skill will recognize that the
data can be transformed in any number of ways.
[0017] In one example, the facility receives multiple input signals
(e.g., input signals collected by different electrodes or leads
connected to a patient, multimodal signals, such as signals from
leads of wide-band biopotential measuring equipment and a channel
of S.sub.pO.sub.2 (blood oxygen saturation), and so on) and/or
transformed signals and extracts values from the signal data by
computing, for each signal, an average value of the signal over the
sampling period. In this example, four signals per patient are
represented, although one of ordinary skill in the art will
recognize that any number of signals may be monitored and/or
received for processing and further analysis by the facility. Thus,
in this example, the extracted data of each patient can be
represented as a set of these average values over time, such
as:
TABLE-US-00001 TABLE 1 Patient A B C D 1 0.24 0 0 30 2 0.2 0.6 4.2
5 ... n .32 2 4 .02
Table 1 represents a set of average signal values (A, B, C, and D)
for each of n patients. Although average values have been used
here, one of ordinary skill in the art will recognize that any type
of data can be extracted or computed from the underlying data
signals, such as the amount of time that a signal exceeded a
threshold value, the values for one signal while the value of
another signal exceeded a threshold value, and so on.
[0018] In some embodiments, after data have been extracted from the
received signal, the facility applies one or more feature
generators to the received or generated data, such as the extracted
data, the raw or preprocessed signal data, the transformed data,
and so on. A feature generator receives as input at least a portion
or representation of the signal data and produces a corresponding
output value (or set of values) (i.e., a "feature"). One set of
feature generators includes the following equations:
F .times. 1 = A + C - D , ( Eq .times. .times. 1 ) F .times.
.times. 2 = A * S .function. ( 4 ) * B D + C + D , and ( Eq .times.
.times. 2 ) F .times. .times. 3 = S .function. ( 1 ) * D , ( Eq
.times. .times. 3 ) ##EQU00001##
where each of A, B, C, and D represents a value extracted from a
specific patient's data and S(t) represents, for each signal, the
value of the signal at time t. In Eq 1, for example, F1 represents
the name of the feature while the equation A+C-D represents the
corresponding feature generator. In some cases, the facility
employs composite feature generators in which one feature generator
serves as an input to another feature generator, such as:
F .times. 4 = F .times. 1 * F .times. 2 F .times. .times. 3 3 + . 0
.times. 57 ( Eq .times. .times. 4 ) ##EQU00002##
In this example, the facility applies feature generators to the
extracted data of each patient represented in Table 1 to generate,
for each feature generator, a feature vector of three values (one
for each patient), such as those represented in Table 2 below:
TABLE-US-00002 TABLE 2 Patient F1 F2 F3 1 -29.76 5.48 905.83 2 -0.6
6.67 9.57 ... n 4.3 185.74 0.04
[0019] In this example, the facility has applied each feature
generator F1, F2, and F3 to the extracted data shown in Table 1 to
generate, for each feature generator, a corresponding feature
vector that includes a value for each patient. For example, the
feature vector generated by applying feature generator Fl to the
extracted data includes a value of -29.76 for Patient 1, a value of
-0.6 for patient 2, and so on. Thus, each feature vector
represents, for a specific feature generator, a signature (not
necessarily unique) for the corresponding feature generator based
on at least a portion of each patient's physiological data (i.e.,
the patients represented in the physiological data to which the
feature generators were applied). In some examples, feature
generators are expressed using different structures or models, such
as expression trees, neural networks, etc. One of ordinary skill in
the art will recognize that the facility may employ any number of
feature generators and any number of sets of physiological data (or
portions thereof) in the generation of feature vectors. In some
embodiments, the facility randomly selects a number of
previously-generated feature generators for use in generating
feature vectors rather than employing each and every available
feature generator. In some embodiments, the facility creates and/or
modifies feature generators by, for example, randomly generating
expression trees, randomly assigning weights to connections within
a neural network, and so on.
[0020] In some embodiments, after the facility generates a number
of feature vectors, the facility employs some form of novelty
search to identify the most "novel" feature vectors among the
generated feature vectors. Novelty corresponds to how different a
particular feature vector is from each of a comparison set of other
feature vectors (made up of any feature vectors generated by the
facility during a current iteration and feature vectors produced by
feature generators selected in any earlier iteration); the greater
the difference from the feature vectors of the comparison set, the
greater the novelty. The facility uses a form of distance as a
measure of novelty (i.e., how "far" each feature vector is from the
other feature vectors). In this case, for each generated feature
vector, the facility calculates the distance between that feature
vector and each of the other generated feature vectors and performs
an aggregation of the generated distance values, such as
calculating an average or mean (e.g., arithmetic, geometric,
harmonic, etc.) distance value for the feature vector, or a total
(sum) distance between the feature vector and each of the other
generated feature vectors, identifying a mode distance value, a
median distance value, a maximum distance value for the feature
vector, and so on. For example, using the feature vectors of Table
2 (for patients 1, 2, and n), the distances for each set of feature
vectors could be calculated as such:
F .times. .times. 1 .times. - .times. F .times. .times. 2 .times.
.times. distance .times. : .times. ( - 2 .times. 9.76 - 5.48 ) 2 +
( - 0.6 - 6.67 ) 2 + ( 4.3 - 1 .times. 8 .times. 5 . 7 .times. 4 )
2 = 1 .times. 8 .times. 4 . 9 .times. 7. ##EQU00003## F .times.
.times. 1 .times. - .times. F .times. .times. 3 .times. .times.
distance .times. : .times. ( - 2 .times. 9.76 - 905.83 ) 2 + ( -
0.6 - 9.57 ) 2 + ( 4.3 - 0.04 ) 2 = 93 .times. 6 . 2 .times. 3
##EQU00003.2## F .times. .times. 2 .times. - .times. F .times.
.times. 3 .times. .times. distance .times. : .times. ( 5.48 -
905.83 ) 2 + ( 6.67 - 9.57 ) 2 + ( 185.74 - 0.04 ) 2 = 91 .times. 9
. 7 .times. 0 . ##EQU00003.3##
[0021] In this example, the total Euclidean distance between each
of the feature vectors has been calculated as a means for
calculating a difference between each of two vectors. In addition
to the feature vectors generated by a current set (i.e., a current
generation) of feature generators, the facility includes feature
vectors produced by feature generators selected in an earlier
generation. In some examples, the facility applies a weight, such
as a randomly generated weight, to each of the feature vectors
and/or normalizes each set of feature vectors prior to comparison.
Thus, the distance measurements for each of the feature vectors in
this example are as follows:
TABLE-US-00003 TABLE 3 Feature Distance Distance Distance Average
MAX Generator to F1 to F2 to F3 Distance Distance F1 -- 184.97
936.23 560.60 936.23 F2 184.97 -- 919.70 552.34 919.70 F3 936.23
919.70 -- 927.97 936.23
[0022] In this example, the facility identifies the most "novel"
feature vectors based on the calculated distances, which act as a
"novelty score" or "fitness score" for each of the feature vectors.
The facility identifies the feature vectors with the greatest
average distance to other vectors (e.g., the feature vector
generated by F3), the feature vectors with the greatest MAX
distance (e.g., the feature vectors generated by F1 and F3), and so
on. In some examples, the number of novel feature vectors
identified is fixed (or capped) at a predetermined number, such as
five, ten, 100, 500, etc. In other examples, the number of novel
feature vectors to be identified is determined dynamically, such as
the top 10% of analyzed feature vectors based on novelty scores,
any feature vectors having a novelty scores that is more than a
predetermined number of standard deviations beyond a mean novelty
score for the analyzed feature vectors, and so on. The feature
generators that produced each of these identified novel feature
vectors can then be added to the set of features available for use
as inputs to models constructed and evaluated by the machine
learning pipeline. Those models can be applied to patient data for,
e.g., diagnostic, predictive, therapeutic, or other analytic,
scientific, health-related or other purposes.
[0023] In some embodiments, in addition to providing the feature
generators used to generate the identified novel feature vectors
for use by the machine learning process, the facility randomly
mutates or modifies the feature generators used to generate the
identified novel feature vectors. Each mutation effects some change
in the corresponding feature generator and creates a new version of
the feature generator that can be used to contribute to a new
generation of feature generators. The facility uses this new
feature generator to generate new feature vectors, and then
assesses the novelty of the new feature vectors. Moreover, the
corresponding feature generator can be further mutated to continue
this process of feature vector and feature generation creation. For
example, a feature generator expressed in the form of an equation,
such as F1.sub.0=A+C-D, can be mutated by randomly selecting one or
more element(s) of the equation and replacing the selected
element(s) with other elements (e.g., randomly selected elements).
In this example, the equation can be changed by replacing A with B
to create F1.sub.1=B+C-D or replacing C-D with
C - B 2 3 ##EQU00004##
to create
F .times. 1 1 = B + C - B 2 3 . ##EQU00005##
In this case, the subscripted 0 and 1 have been included to
represent a generational marker or count for each of the feature
generators. In other words, F1.sub.0 represents F1 above (Eq 1) at
generation 0 (i.e., the first generation), F1.sub.1 represents a
mutated version of F1 at generation 1 (i.e., the second
generation), and so on. In some cases, an earlier generation (or a
transformation thereof) is included as an element in subsequent
generations, such as F2.sub.1= {square root over
(F2.sub.0)}+C.sup.2 or {square root over (F2.sub.n-1)}+C.sup.2
(n.noteq.0).
[0024] In some embodiments, the facility obtains features in
different ways. For example, the facility may receive from a user,
such as a domain expert, a set of features (and corresponding
feature generators) that the user has identified as being optimal
and/or that the user desires to be tested. As another example, the
features may be editorially selected from one or more feature
stores. In some cases, features automatically generated by the
facility can be combined with other features to create various
hybrid features. Even features of unknown provenance may be
used.
[0025] In some embodiments, the facility identifies genomes to
train models, identifies, from among these genomes, the "best"
(highest rated) genomes, and mutates the identified genomes to
produce even more genomes that can be used to train models. After
using a genome to train one or more models, the facility applies
each trained model to a validation data set so that the trained
model can be scored (e.g., how well does the trained model
correctly identify and/or classify subjects in the underlying
validation data set). The facility mutates the genomes that produce
the best results (e.g., have the highest validation or fitness
scores), trains new models using these mutated genomes, and repeats
this process until one or more termination criteria are met (e.g.,
a predetermined number of generations, no additional high scoring
(higher than a predetermined or dynamically generated threshold)
genomes are generated during a predetermined or dynamically number
(e.g., 1, 5, 8, 17, etc.) of previous generations, a combination
thereof, etc.).
[0026] In some embodiments, the facility uses previously identified
or generated genomes as a first set of genomes (i.e., a first
generation) from which to discover genomes for machine learning
algorithms. In other examples, the facility automatically generates
a first generation of genomes by, for each genome, randomly (with
or without replacement) selecting one or more feature vectors from
one or more previously generated sets of feature vectors (e.g., a
feature vector produced by applying a feature generator to a set of
training data). A genome may also include one or more machine
learning algorithm parameters to the machine learning algorithm,
such as the number of predictors (e.g., regressors, classifiers,
the number and/or the maximum number of decision trees to use for a
machine learning algorithm, etc.) to use for an underlying ensemble
method associated with the algorithm, a maximum depth for a machine
learning algorithm (e.g., maximum depth for decision trees), and so
on. In the event that the genome is configured to be used with one
specific machine learning algorithm, the genome can be configured
to define a value for each machine learning parameter associated
with that machine learning algorithm. In other cases, one of the
elements of the genome selects among different machine learning
algorithms and may be mutated so that the genome and its
corresponding parameter values are used with different machine
learning algorithms to train models over the evolutionary process.
For example, during a first generation, a genome may identify a
machine learning algorithm that relies on decision trees while a
mutated version of that same genome identifies a machine learning
algorithm that uses one or more support vector machines, linear
models, etc. In these cases, the genome may specify a modeling
parameter for each and every machine learning algorithm that may be
combined with the genome to train a model. Thus, a single genome
may include machine learning parameters for multiple machine
learning algorithms. However, a genome need not include each and
every modeling parameter for a corresponding machine learning
algorithm. In the event that a model is to be trained using a
particular machine learning algorithm and a genome that does not
include a value for a machine learning parameter of that machine
learning algorithm, the facility can retrieve a default value for
these parameters from, for example, a machine learning parameter
store.
[0027] For example, a set of genomes may be represented as:
TABLE-US-00004 TABLE 4 G1.sub.1 MLA = 4 F23 F78798 F32 F55 F453
F234 G2.sub.1 MLA = 9 F9701 F223 F1 F63 F349 P9:1 = 7 G3.sub.1 MLA
= 2 F823 F525 F732 F525 F125 G4.sub.1 MLA = 6 F597 F135 F404 F31
P6:1 = 5 P6:2 = 150 . . . G20.sub.1 MLA = 1 F43 F65 P1:1 = 8 P1:2 =
218 P1:3 = 0.3
where each row corresponds to a different genome (named in the
first column from the left) from among a first generation of
selected or generated genomes and identifies a machine learning
algorithm ("MLA"; second column from the left) to use to train a
model using the genome, such as an index into a machine learning
algorithm store. For example, genome G3.sub.1 specifies a machine
learning algorithm corresponding to index 2 in a machine learning
algorithm store (MLA=2). In this example, each non-bolded region
(to the right of the second column) identifies a different feature.
A genome can also include a corresponding feature generator or a
reference to the corresponding feature generator, such as a link to
feature generator store. As discussed above, these features may be
generated automatically by the facility and/or retrieved from
another source.
[0028] Furthermore, each bolded region in Table 4 represents a
value for a particular machine learning parameter. In this example
set of genomes, machine learning parameters are represented by an
indicator or reference (e.g., P6:1) followed by an equals sign and
a corresponding value. For example, machine learning algorithm
parameter P6:1 has a corresponding value of 8 in genome G20.sub.1.
In this example set of genomes, each machine learning parameter is
presented as an index into a two-dimensional array, such that
"P6:1" represents the "first" machine learning parameter of the
"sixth" machine learning algorithm (i.e., the machine learning
parameter with an index of 1 for the machine learning algorithm
with an index of 6). As discussed above, a genome may specify
values for any or all machine learning parameters that may be used
to training a model using the genome (or a mutated version of that
genome). Moreover, as is clear from Table 4, genomes may be of
varying length. For example, genome G1.sub.1 includes values for
six features and zero machine learning parameters while gnome
G2.sub.1 includes values for two features and three machine
learning parameters. Accordingly, the facility may employ
variable-length genomes in the machine learning processes.
[0029] In some embodiments, the facility may filter features from
within genomes and/or filters genomes themselves to avoid
redundancy among each. In order to filter features and/or genomes,
the facility generates correlation values for each pair and
discards one item of the pair. To identify and filter correlated
features from a genome, the facility generates, for each of the
features, a feature vector by applying a feature generator
associated with the feature to a training set of data to produce a
set of values. The facility compares each of the generated feature
vectors to the other generated feature vectors to determine whether
any of the feature vectors are "highly" correlated (i.e., not
"novel" within the selected set of feature vectors). For example,
the component may calculate a distance value for each of the
generated feature vectors relative to the other feature vectors (as
discussed above with respect to identifying novel feature
generators) and, if the distance between any pair (set of two) is
less than or equal to a distance threshold (i.e., "highly"
correlated or not "novel"), discard a feature corresponding to one
of the pair of feature vectors. Moreover, the facility may replace
the discarded feature with a new feature, such as a
randomly-selected feature. Similarly, the facility may identify and
discard redundant genomes by generating, for each feature of the
genome, a feature vector, calculating distance metrics for each
pair (set of two) of genomes based on the generated feature
vectors, and identifying pairs of genomes whose calculated
distances do not exceed a genome distance threshold. For each
identified pair of genomes, the facility may discard or mutate one
or both of the genomes to reduce correlation and redundancy among a
group of genomes. Although distance is used in this example as a
metric for determining a correlation between two vectors or sets of
vectors, one of ordinary skill in the art will recognize that
correlations between two or sets of vectors can be calculated in
other ways, such as normalized cross-correlation, and so on. In
some embodiments, the facility may employ additional or other
techniques to filter genomes, such as generating a graph where
features represent vertices in the graph which are connected via
edges in the graph. An edge between two features is generated if,
for example, a correlation value between the two features exceeds a
predetermined correlation threshold and/or the distance between the
two features is less than a predetermined distance threshold. Once
the graph is generated, the facility removes connected vertices
(features) from the graph until no edges remain in the graph (an
edge being removed when a connected vertex is removed) and selects
the remaining non-connected vertices (features) for inclusion in
the "filtered" genome. In some cases, the facility may randomly
select connected vertices for removal. Moreover, the facility may
perform this process multiple times for a set of vertices
(features) and then select a preferred "filtered" genome, such as
the genome with the most or least vertices (features) removed.
[0030] In order to test the fitness or validity of each genome, the
facility trains at least one model using the features, machine
learning parameters, and/or machine learning algorithm(s) of that
genome. For example, the facility can use AdaBoost ("Adaptive
Boosting") techniques to train a model using the corresponding
features, machine learning parameters, machine learning algorithm,
and a training set of data. However, one of ordinary skill in the
art will recognize that many different techniques can be used to
train one or more models given a genome or a set of genomes. After
the model is trained, the facility applies the trained model to one
or more sets of validation data to assess how well the trained
model identifies and/or classifies previously-identified or
classified subjects within the validation data set. For example, a
genome may be generated to train models to identify patients
represented in a data set who are likely to have diabetes. Once a
model is trained using one of these genomes, the trained model can
be applied to a validation set of data to determine a validation
score that reflects how well the trained model identifies patients
from the validation set that are known to have or now have
diabetes; scoring (adding) one "point" for every correct
determination (e.g. true positives and true negatives) and losing
(subtracting) one "point" for every incorrect determination (e.g.,
false positives and false negatives). Thus, an overall score for
the trained model can be determined based on how many "points" the
trained model scores when applied to one or more sets of validation
data. One of ordinary skill in the art will recognize that several
techniques may be used to generate a fitness score for a trained
model, such as calculating the area under a corresponding receiver
operating characteristic (ROC) curve, calculating a mean squared
prediction error, f scores, sensitivity, specificity, negative and
positive predictive values, diagnostic odds ratios, and so on. In
this example, where a single machine learning algorithm is trained
using the genome, the generated fitness score may be similarly
attributed to the genome. In other case, the genome may be used to
train multiple machine learning algorithms and each of those
trained machine learning algorithms may be applied to multiple
validation sets to produce, for each genome used to train machine
algorithms, multiple fitness scores. In these cases, the facility
generates a fitness score for the corresponding genome by
aggregating each of the fitness scores generated for the machine
learning algorithms trained using the genome. In some cases, the
generated fitness scores may be aggregated and/or filtered prior to
aggregation.
[0031] In some embodiments, after the facility has produced fitness
scores for each of the genomes, the facility identifies the "best"
genomes based on these fitness scores. For example, the facility
can establish a fitness threshold based on the produced fitness
scores and identify the "best" genomes as those genomes whose
resulting fitness scores exceed the fitness threshold. The fitness
threshold may be generated or determined in any number of ways,
such as receiving a fitness threshold from a user, calculating a
fitness threshold based on the set of fitness scores (e.g.,
average, average plus 15%, top fifteen, top n-th percentile (where
n is provided by a user or generated automatically by the
facility)), and so on. The facility then stores each of the genomes
in association with their corresponding fitness scores and selects
the genomes identified as "best" for mutation (i.e., the genomes
having a fitness score that exceeds a fitness threshold).
[0032] In some embodiments, the facility mutates a genome by
adding, removing, or changing any one or more of the feature
vectors or machine learning parameters of the genome. For example,
Table 5 below represents a number of mutations to the genomes
represented above in Table 4.
TABLE-US-00005 TABLE 5 G1.sub.2 MLA = 5 F23 F78798 F32 G2.sub.2 MLA
= 9 F223 F1 F63 F349 F584 P9:1 = 12 G4.sub.2 MLA = 6 F597 F135 F404
F31 F24 F982 P6:1 = 5 P6:2 = 150 . . . G20.sub.2 MLA = 1 F43
F65*F14 P1:1 = 8 P1:2 = 218 P1:3 = 0.3
In this example, each row corresponds to a different genome (named
in the first column from the left) from among a second generation
of genomes selected for mutation. In this example, based on its low
fitness score, the facility did not select genome G3.sub.1 for
mutation and, therefore, Table 5 does not include a corresponding
entry for a mutated version of this genome. Moreover, genome
G1.sub.1 has been mutated (represented as G1.sub.2) by removing
three feature vectors (represented with a strikethrough) and
changing the references machine learning algorithm index from 4 to
5. Furthermore, the facility has mutated genome G2.sub.1 by 1)
removing feature vector F9701, 2) adding feature vector F584, and
3) adjusting machine learning parameter P9.sub.1 from 7 to 12;
genome G4.sub.1 by adding features F24 and F982; and genome
Gn.sub.1 by multiplying values generated by F65 by values generated
by F14. These mutated genomes can then be used to train one or more
machine learning algorithms, scored by applying the trained machine
learning algorithm to one or more validation data sets, selected
for mutation, mutated, and so on. The facility performs this
process until a termination point is reached, such as when a
predetermined number of generations has been produced (e.g., six,
30, 100,000, etc.), and so on.
[0033] FIG. 1 is a block diagram illustrating an environment 100 in
which the facility operates in accordance with some embodiments of
the disclosed technology. In this example, environment 100 includes
service provider 110, signal recorder 140 (e.g., a wide-band
biopotential measuring equipment), data providers 150, patient 160,
and network 160. In this example, service provider includes
facility 120, which includes discover features component 121,
process component 122, apply feature generators component 123,
identify novel feature vectors component 124, discover genomes
component 125, spawn genome component 126, identify high-performing
genomes component 127, patient data store 130, model store 131,
feature vector store 132, and feature generator store 133. Discover
features component 121 is invoked by the facility to identify and
mutate feature generators based on received data. Process component
122 is invoked by the discover features component 121 to process
and transform patient signal data, such as raw signal data from
signal recorder 140 (e.g., one or more measurement devices and/or
systems used to collect the underlying data such as wide-band
biopotential measuring equipment, etc.), 3-D image data, etc. Apply
feature generators component 123 is invoked by the discover
features component to apply a set of one or more feature generators
to the processed and transformed patient signal data. Identify
novel feature vectors component 124 is invoked by the discover
features component to identify the most novel feature vectors from
among a group of feature vectors generated by, for example, one or
more feature generators. Discover genomes component 125 is invoked
by the facility 120 to generate, analyze, and mutate genomes for
use by machine learning algorithms. Spawn genome component 126 is
invoked by the discover genomes component to generate a genome
comprising any number of feature vectors and/or machine learning
parameters. Identify high-performing genomes component 127 is
invoked by the discover genomes component to identify, from among a
group of genomes, the genomes that have a corresponding fitness
score that exceeds a fitness threshold. Patient data store 130
includes physiological patient data, such as raw physiological data
(including but not limited to data obtained via, e.g., signal
recorder 140), transformed physiological data, biographical
information, demographic information, etc. These data may be stored
anonymously to protect the privacy of each of the corresponding
patients and may be processed and encrypted to ensure that its
transmission and storage complies with any governing laws and their
implementing regulations, such as the U.S. Health Insurance
Portability and Accountability Act of 1996 (as amended), the
European Data Protection Directive, the Canadian Personal
Information Protection and Electronics Documents Act, the
Australian Privacy Act of 1988, Japan's Personal Information
Protection Act of 2015 (as amended), state and provincial laws and
regulations, and so on. Model store 131 stores information about
models generated by applying machine learning techniques to
training data, such as the machine learning techniques described in
Christopher M. Bishop, Pattern Recognition and Machine Learning
(2006) (Library of Congress Control Number: 2006922522; ISBN-10:
0-387-31073-8), which is herein incorporated by reference in its
entirety. Feature vector store 132 stores sets of feature vectors
generated by applying one or more feature generators to a set of
physiological data. Feature generator store 133 stores sets of
feature generators that can be applied to patient physiological
data and can include multiple generations of feature generators.
Genome store 134 stores generated and/or mutated genomes created by
the facility and/or other sources. Machine learning parameters
store 135 stores, for each of a number of machine learning
algorithms, a set of parameters that can serve as inputs to that
machine learning algorithm and additional information related to
the parameter, such as a maximum value for a corresponding
parameter, a minimum value for a corresponding parameter, a default
value for a corresponding parameter, etc. Machine learning
algorithm store 133 stores the logic for each of a number of
machine learning algorithms, each of which can be selectively
trained and validated by the facility. In this example, a signal
recorder 140 is connected to patient 160 via electrodes 145 and
includes facility 120, one or more output devices 142, such as a
monitor, printer, speaker, etc., and one or more input devices 144,
such as settings controls, keyboard, biometric data reader, etc.
Thus, as in this example, the facility can be configured to operate
remotely from a patient and other diagnostics equipment and/or in
conjunction with or part of the diagnostics equipment such as a
wide-band biopotential measuring equipment (i.e., any device
configured to capture unfiltered electrophysiological signals,
including those with spectral components that are not altered).
Accordingly, the facility can be configured to operate in real-time
with the reading of physiological data and/or can be applied to
previously recorded physiological data. Data providers 150, each of
which includes a data store 152, can provide information for
analysis or use by the facility such as, physiological patient data
recorded off-site (e.g., at a hospital or clinic without access to
a facility on premises, third-party data providers, etc.), feature
vectors and/or feature generators produced or generated elsewhere,
and so on. Network 170 represents communications links over which
the various elements of environment 100 may communicate, such as
the internet, a local area network, and so on.
[0034] In various examples, these computer systems and other
devices can include server computer systems, desktop computer
systems, laptop computer systems, netbooks, tablets, mobile phones,
personal digital assistants, televisions, cameras, automobile
computers, electronic media players, appliances, wearable devices,
other hardware, and/or the like. In some embodiments, the facility
120 may operate on specific-purpose computing systems, such as
wide-band biopotential measuring equipment (or any device
configured to capture unfiltered electrophysiological signals,
including electrophysiological signals with unaltered spectral
components), electroencephalogram equipment, radiology equipment,
sound recording equipment, and so on. In various examples, the
computer systems and devices include one or more of each of the
following: a central processing unit ("CPU") configured to execute
computer programs; a computer memory configured to store programs
and data while they are being used, including a multithreaded
program being tested, a debugger, the facility, an operating system
including a kernel, and device drivers; a persistent storage
device, such as a hard drive or flash drive configured to
persistently store programs and data (e.g., firmware and the like);
a computer-readable storage media drive, such as a floppy, flash,
CD-ROM, or DVD drive, configured to read programs and data stored
on a computer-readable storage medium, such as a floppy disk, flash
memory device, CD-ROM, or DVD; and a network connection configured
to connect the computer system to other computer systems to send
and/or receive data, such as via the internet, a Local Area Network
(LAN), a Wide Area Network (WAN), a point-to-point dial-up
connection, a cell phone network, or another network and its
networking hardware in various examples including routers,
switches, and various types of transmitters, receivers, or
computer-readable transmission media. While computer systems
configured as described above may be used to support the operation
of the facility, those skilled in the art will readily appreciate
that the facility may be implemented using devices of various types
and configurations, and having various components. Elements of the
facility may be described in the general context of
computer-executable instructions, such as program modules, executed
by one or more computers or other devices. Generally, program
modules include routines, programs, objects, components, data
structures, and/or the like configured to perform particular tasks
or implement particular abstract data types and may be encrypted.
Furthermore, the functionality of the program modules may be
combined or distributed as desired in various examples. Moreover,
display pages may be implemented in any of various ways, such as in
C++ or as web pages in XML (Extensible Markup Language), HTML
(HyperText Markup Language), JavaScript, AJAX (Asynchronous
JavaScript and XML) techniques, or any other scripts or methods of
creating displayable data, such as the Wireless Access Protocol
(WAP). Typically, the functionality of the program modules may be
combined or distributed as desired in various embodiments,
including cloud-based implementations, web applications, mobile
applications for mobile devices, and so on.
[0035] The following discussion provides a brief, general
description of a suitable computing environment in which the
disclosed technology can be implemented. Although not required,
aspects of the disclosed technology are described in the general
context of computer-executable instructions, such as routines
executed by a general-purpose data processing device, e.g., a
server computer, wireless device, or personal computer. Those
skilled in the relevant art will appreciate that aspects of the
disclosed technology can be practiced with other communications,
data processing, or computer system configurations, including:
internet or otherwise network-capable appliances, hand-held devices
(including personal digital assistants (PDAs)), wearable computers
(e.g., fitness-oriented wearable computing devices), all manner of
cellular or mobile phones (including Voice over IP (VoIP) phones),
dumb terminals, media players, gaming devices, multi-processor
systems, microprocessor-based or programmable consumer electronics,
set-top boxes, network PCs, mini-computers, mainframe computers,
and the like. Indeed, the terms "computer," "server," "host," "host
system," and the like are generally used interchangeably herein,
and refer to any of the above devices and systems, as well as any
data processor.
[0036] Aspects of the disclosed technology can be embodied in a
special purpose computer or data processor, such as
application-specific integrated circuits (ASIC), field-programmable
gate arrays (FPGA), graphics processing units (GPU), manycore
processors, and so on, that is specifically programmed, configured,
or constructed to perform one or more of the computer-executable
instructions explained in detail herein. While aspects of the
disclosed technology, such as certain functions, are described as
being performed exclusively on a single device, the disclosed
technology can also be practiced in distributed computing
environments where functions or modules are shared among disparate
processing devices, which are linked through a communications
network such as a Local Area Network (LAN), Wide Area Network
(WAN), or the internet. In a distributed computing environment,
program modules may be located in both local and remote memory
storage devices.
[0037] Aspects of the disclosed technology may be stored or
distributed on tangible computer-readable media, including
magnetically or optically readable computer discs, hard-wired or
preprogrammed chips (e.g., EEPROM semiconductor chips),
nanotechnology memory, biological memory, or other
computer-readable storage media. Alternatively,
computer-implemented instructions, data structures, screen
displays, and other data under aspects of the disclosed technology
may be distributed over the internet or over other networks
(including wireless networks), on a propagated signal on a
propagation medium (e.g., electromagnetic wave(s), sound wave,
etc.) over a period of time, or they may be provided on any analog
or digital network (packet switched, circuit switched, or other
scheme). Furthermore, the term computer-readable storage medium
does not encompass signals (e.g., propagating signals) or
transitory media.
[0038] FIG. 2 is a flow diagram illustrating the processing of a
discover features component 121 in accordance with some embodiments
of the disclosed technology. The discover features component is
invoked by the facility to identify novel feature vectors based on
selected patient data. In block 205, the component receives
physiological signal data, such as raw signal data received
directly from signal recorder, previously-generated physiological
signal from another device or site, etc. Several techniques exist
for collecting and analyzing physiological signals (e.g.,
electrophysiological signals, biosignals) from patients for
diagnostic and other purposes including, for example, activity
trackers, echocardiograms, wide-band biopotential measuring
equipment, electroencephalograms, electromyograms,
electrooculography, galvanic skin response, heart rate monitors,
magnetic resonance imaging, magnetoencephalograms, mechanomyograms,
wearable technology devices (e.g., FITBITs), and so on. While data
provided by these systems can be helpful in identifying medical
concerns and diagnosing medical conditions, they are often only a
starting point for the diagnosis process. Moreover, given the
specific nature of most of these systems, the data they analyze is
often over-filtered to reduce complexity for the system itself or
for, e.g., a technician, physician, or other health care provider
(in such cases, to reduce visual complexity, etc.) thereby
eliminating data that potentially have untapped diagnostic value.
In block 210, the component invokes a process signal data component
to process and transform the received signal data, which can
produce multiple sets of data and transformed data. In block 215,
the component sets a generation value equal to 0. In block 220, the
component generates one or more feature generators by, for example,
randomly generating an expression tree, randomly generating a set
of weights for a neural network, randomly mutating one or more of a
set of previously-generated feature generators, and so on. In block
225, the component invokes an apply feature generators component to
apply the generated feature generators to one or more sets of the
processed signal data to produce a set of feature vectors. In block
230, the component invokes an identify novel feature vectors
component to identify the most novel feature vectors from among the
group of feature vectors generated by the feature generators. In
block 235, the component stores the feature generators that
produced the identified feature vectors in, for example, a feature
generator store. In block 240, the component increments the
generation variable. In decision block 245, if the generation
variable is greater than or equal to a generation threshold, then
the component completes, else the component continues at block 250.
The component may also use other stopping conditions, such as a
number of generations of feature generators that do not produce at
least a threshold number of novel feature vectors. In block 250,
the component copies and mutates the identified feature generators
and then loops back to block 225 to apply the mutated feature
generators to one or more sets of the processed signal data. As
discussed above, the component may apply any type or types of
mutations to a feature generator, such as applying multiple point
mutations and/or a random recombination to one or more expression
trees, randomly generating a set of connection weights for a neural
network, and so on.
[0039] FIG. 3 is a flow diagram illustrating the processing of a
process component 122 in accordance with some embodiments of the
disclosed technology. The process component is invoked by the
discover features component to process and transform patient signal
data. In blocks 305 through 365, the component loops through each
signal (or data set) of a set of received signals (or set of data
sets), each signal representing physiological data received from a
patient. In block 310, the component pre-processes the received
signal, such as applying one or more signal filters to the signal,
performing a peak search on the data and discarding extraneous
information, down-sampling the received signal, up-sampling the
received signal, sub-sampling the received signal, converting an
analog signal to a digital signal, converting image data to signal
data, and so on. In block 315, the component stores the
pre-processed signal in, for example, a patient data store. The
signal data may be stored anonymously (i.e., without explicitly or
implicitly identifying the corresponding patient, etc.). However,
different instances of signal data associated with the same patient
may be associated with an anonymized unique identifier so that
multiple signals from a single patient can be used in conjunction
for training and diagnostic purposes. In block 320, the component
extracts one or more values from the stored signal data. In block
325, the component stores the one or more extracted values. In
block 330, the component identifies any transformations to be
applied to the signal. For example, the facility may store an
indication of a set of transformations or transformation functions
(e.g., Fourier transforms, functions to apply to the signal,
derivatives, partial derivatives, and so on) to apply to a
particular signal. As another example, the facility may randomly
select, from among a catalog of transformations, one or more
transformations to apply to the signal data. In blocks 335 through
360, the component loops through each of the transformations and
applies the transformation to the signal. In block 340, the
component applies the transformation to the signal (e.g.,
calculating the third derivative with respect to a particular
variable, calculating the result of a composite function generated
by applying one function to the signal data (i.e., a function
representative of the signal data), etc.). In block 345, the
component stores the transformed signal data in, for example, a
patient data store. In block 350, the component extracts one or
more values from the transformed signal data. In block 355, the
component stores the one or more extracted values. In block 360, if
there are any identified transformations yet to be applied, then
the component selects the next transformation and loops back to
block 335 to apply the transformation to the signal data, else the
component continues at block 365. In block 365, if there are any
signals yet to be analyzed, then the component selects the next
signal and loops back to block 305 to process the next signal, else
the component completes.
[0040] FIG. 4 is a flow diagram illustrating the processing of an
apply feature generators component 123 in accordance with some
embodiments of the disclosed technology. The apply feature
generators component is invoked by the discover features component
121 to apply a set of one or more feature generators to signal
data, such as pre-processed and transformed signal data, modeled
signal data, etc. In blocks 410 through 470, the component loops
through each of a received set of feature generators and applies
the feature generator to each signal in a received set of signal
data. For example, the received signal data can include multiple
sets of signal data for each of multiple patients, multiple
transformations of that data, and so on. In blocks 420 through 450,
the component loops through each of the signals to apply the
feature generators to the signal data. In block 430, the component
applies the currently-selected feature generator to the
currently-selected signal data. For example, the component may
apply the feature generator to each of a pre-processed version of
the currently-selected data signal and any transformed version of
that data. As another example, the component "plugs in" or
substitutes coefficients generated by modeled signal data into a
feature generator with a set of variables to produce an output
feature value. As another example, the component can apply one or
more elements of modeled signal data to a neural network to produce
an output feature value. In block 440, the component stores the
output value. In block 450, if there are any signals yet to be
analyzed, then the component selects the next signal and loops back
to block 420 to process the next signal, else the component
continues at block 460. In block 460, the component generates a
feature vector that includes each of the generated feature values
and stores the feature vector in association with the feature
generator in, for example, a feature vector store. For example, the
feature vector may comprise an array of features and a link to, or
identifier of, the corresponding feature generator. The component
may also associate the feature vector with the signal data used to
generate the feature vector. In block 470, if there are any feature
generators yet to be processed, then the component selects the next
feature generator and loops back to block 410 to process the
feature generator, else the component completes.
[0041] FIG. 5 is a flow diagram illustrating the processing of an
identify novel feature vectors component 124 in accordance with
some embodiments of the disclosed technology. In this example, the
facility receives a set of feature vectors and, for each feature
vector, information related to the corresponding feature generator,
such as an identifier for the feature generator. In block 505, the
component collects a comparison set of feature vectors that
includes, for example, feature vectors generated by feature
generators of earlier generations that were found to be novel and
the feature vectors generated by a current generation of feature
vectors. For example, the component can randomly select a set of
novel feature vectors from a feature store. In some cases, a
request to retrieve feature vectors includes upper and lower limits
on the number of features values for each feature vector to be
retrieved, such as no less than 50 (lower threshold) and no more
than 5000 (upper threshold). In blocks 510 through 540, the
component loops through each feature vector of a current generation
of feature generators to determine how different each of their
corresponding feature vectors is to each of the feature vectors of
the comparison set of feature vectors. In blocks 515 through 530,
the component loops through each feature vector of the comparison
set of feature vectors to compare each feature vector to the
feature vector of the currently-selected feature generator. In
block 520, the component calculates a difference value between the
currently-selected feature vector of the comparison set and the
feature vector of the currently-selected feature generator. For
example, the component can calculate a distance value between each
of the feature vectors. In block 525, the component stores the
calculated difference value. In block 530, if there are any feature
vectors yet to be compared, then the component selects the next
feature vector and loops back to block 515 to process the feature
vector, else the component continues at block 535. In block 535,
the component calculates a novelty score for the currently-selected
feature generator based on the stored difference values, such as an
average or maximum distance, and stores the novelty score in
association with the feature generator (e.g., in a feature
generator store). In block 540, if there are any feature generators
yet to be assessed, then the component selects the next feature
generator and loops back to block 515 to process the feature
generator, else the component continues at block 545. In blocks 545
through 560, the component tests whether each of the feature
vectors is novel, based on the calculated novelty scores, and
identifies any corresponding feature generators. In decision block
550, if the novelty score for the currently-selected feature
generator is greater than a novelty threshold, then the component
continues at block 555, else the component continues at block 560.
The novelty threshold may be generated or determined in any number
of ways, such as receiving a novelty threshold from a user,
calculating a novelty threshold based on the set of novelty scores
(e.g., average, average plus 25%, top n (where n is provided by a
user or generated automatically by the facility), top tenth
percentile), and so on. In this manner, the novelty threshold may
change dynamically (e.g., from generation to generation) based on,
for example, the number of generations without a new feature
generator that exceeds the current novelty threshold to ensure that
the facility is producing and testing new feature generators and
corresponding features. In block 555, the component identifies the
currently-selected feature vector as novel. In block 560, if there
are any feature vectors yet to be processed, then the component
selects the next feature vector and loops back to block 545 to
process the feature vector, else the component completes.
[0042] FIG. 6 is a flow diagram illustrating the processing of a
discover genomes component 126 in accordance with some embodiments
of the disclosed technology. The facility invokes the discover
genomes component to generate and analyze genomes for use by
machine learning algorithms. In block 610, the facility initializes
a generation variable equal to 0. In block 620, the component
determines a number (n) of genomes to generate based on, for
example, user input, system parameters, or randomly. In block 630,
the component invokes a spawn genome component n time(s) to spawn
the appropriate number of genomes. In block 640, the component
invokes an identify high-performing genomes component to identify,
from among the spawned genomes, the genomes that have a fitness
score that exceeds a fitness threshold. In block 650, the component
increments the generation variable. In decision block 660, if the
generation variable is greater than or equal to a generation
threshold, then processing of the component completes, else the
component continues at block 670. In block 670, the component
mutates the high-performing genomes and then loops back to block
640 to identify high-performing genomes from among the mutated
genomes (e.g., the mutated genomes having fitness scores that
exceed a fitness threshold). The component may mutate the genome by
adding, changing, or removing (or any combination thereof) one or
more elements of the variable-length genome. For example, the
component may mutate one genome by replacing one feature with
another feature and adding a new feature to the mutated genome. In
another example, the component may select a new machine learning
algorithm to associate with the genome. In this case, the component
may also remove or mutate any irrelevant machine learning algorithm
parameters and/or replace them with machine learning parameter
values for the newly selected machine learning algorithm. As
another example, genomes may use sexual reproduction techniques as
a form of mutation by randomly selecting elements of multiple
genomes and combining those elements to form a new genome.
Moreover, one or more elements of a genome may be configured so
that they remain fixed (i.e., are not changed) during the
evolutionary process described herein.
[0043] FIG. 7 is a flow diagram illustrating the processing of a
spawn genome component 126 in accordance with some embodiments of
the disclosed technology. The discover genomes component 125
invokes the spawn genome component to generate a genome identifying
any number of features, machine learning parameters, and/or machine
learning algorithms. In block 710, the component identifies a set
of available features, such as features referenced in one or more
feature generator stores. In block 720, the component determines
the number of features to include in the genome to be generated.
For example, the component may determine the number of features to
include in the genome to be generated based on user input, system
parameters, or randomly. In block 730, the component randomly
selects, from among the identified features, the determined number
of features. In block 740, the component replaces correlated
features from among the selected features with randomly selected
features. In block 750, the component identifies a set of available
machine learning parameters. For example, the component may
identify, for each machine learning algorithm available to the
facility, a set of parameters associated with that machine learning
algorithm, which may be stored in a list or other data structure
available to the component (e.g., a machine learning parameter
store). In some cases, a genome may be generated for a single
machine learning algorithm (or fixed set of machine learning
algorithms). In this case, the component may identify only those
machine learning parameters (or a proper subset thereof) that are
associated with the single or fixed set of machine learning
algorithms. In other cases, a genome may include an element that
identifies a machine learning algorithm and that can be mutated. In
this case, the component may identify any or all machine learning
parameters of machine learning algorithms that are within the scope
of this mutation (i.e., the parameters of the machine learning
algorithms that the genome and its descendants may be associated
with to train models during the evolutionary process described
herein). In block 760, the component determines the number of
machine learning parameters to include in the genome to be
generated. For example, the component may determine the number of
machine learning parameters to include in the genome to be
generated based on user input, system parameters, or randomly. For
example, one genome may include each and every machine learning
parameter associated with a particular machine learning algorithm
or a set of machine learning algorithms while another genome
includes only a proper subset of machine learning parameters
associated with a particular machine learning algorithm. In block
770, the component randomly selects, from among the identified
machine learning parameters, the determined number of machine
learning parameters and assigns a value to the parameter based on
any associated constraints, such as randomly selecting a value
between a minimum value and a maximum value associated with the
parameter. In block 780, the component stores each of the selected
features and machine learning parameters in a genome data structure
and then returns the genome data structure.
[0044] FIG. 8 is a flow diagram illustrating the processing of an
identify high-performing genomes component 127 in accordance with
some embodiments of the disclosed technology. The discover genomes
component invokes the identify high-performing genomes component to
identify, from among a group of genomes, the genomes that have a
corresponding fitness score that exceeds a fitness threshold (i.e.,
the genomes that are "high-performing"). In blocks 810 through 850,
the component loops through a set of genomes provided to the
component, such a first generation set of genomes, a mutated set of
genomes, or some combination thereof. In block 820, the component
trains one or more models using the currently-selected genome,
including its features, machine learning parameters, and any
specified machine learning algorithm. In the event that a machine
learning parameter of the genome is not associated with the machine
learning algorithm used to train the model, that machine learning
parameter may be ignored. Similarly, if a particular machine
learning algorithm requires as input a particular machine learning
parameter that is not included in the currently-selected genome,
the facility (or the machine learning algorithm itself) may provide
a default value retrieved from, for example, a machine learning
parameter store. In block 830, the component generates a validation
or fitness score for the currently-selected genome by, for example,
applying the trained model to a set of validation data and
assessing the ability of the trained model to correctly identify or
classify subjects from among the validation data. In block 840, the
component stores the generated score(s) and/or an aggregation
thereof produced for the trained models in association with the
currently-selected genome. In block 850, if there are any genomes
yet to be scored, then the component selects the next genome and
loops back to block 810, else the component continues at block 860.
In blocks 860 through 890, the component assesses the score
generated for each genome and selects the "best" genomes for
mutation. In this example, the "best" genomes are those that
produce validation or fitness scores that exceed a fitness
threshold. In decision block 870, if the score generated for the
currently-selected genome exceeds a fitness threshold, then the
component continues at block 880, else the component continues at
block 890. In block 880, the component flags the currently-selected
genome for mutation. In some embodiments, the component may select
genomes for mutation based on criteria other than, or in addition
to, fitness scores. For example, the component may use novelty
scores or other scores to select genomes for mutation. In some
cases, the component may employ a tournament selection process in
which a number of genomes are randomly selected from the population
and the genome with the highest score among this "tournament" is
selected for reproduction. In this example, if only low scoring
genomes appear in a tournament, a low-scoring genome will be
selected for reproduction. In block 890, if there are any genomes
yet to be processed, then the component selects the next genome and
loops back to block 860, else the component returns the flagged
genomes and processing of the component completes.
[0045] From the foregoing, it will be appreciated that specific
embodiments of the disclosed technology have been described herein
for purposes of illustration, but that various modifications may be
made without deviating from the scope of the disclosed technology.
For example, the disclosed techniques can be applied to fields
outside of the medical field, such as predicting weather patterns,
geological activity, or any other field in which predictions are
made based on sampled input data. To reduce the number of claims,
certain aspects of the disclosed technology are presented below in
certain claim forms, but applicants contemplate the various aspects
of the disclosed technology in any number of claim forms.
Accordingly, the disclosed technology is not limited except as by
the appended claims.
* * * * *