U.S. patent application number 17/245440 was filed with the patent office on 2021-08-12 for methods and systems for evaluation of risk of substance use disorders.
This patent application is currently assigned to Prescient Medicine Holdings, Inc.. The applicant listed for this patent is Prescient Medicine Holdings, Inc.. Invention is credited to Keri Donaldson.
Application Number | 20210249141 17/245440 |
Document ID | / |
Family ID | 1000005553529 |
Filed Date | 2021-08-12 |
United States Patent
Application |
20210249141 |
Kind Code |
A1 |
Donaldson; Keri |
August 12, 2021 |
METHODS AND SYSTEMS FOR EVALUATION OF RISK OF SUBSTANCE USE
DISORDERS
Abstract
Provided here are systems and methods for predicting risk of a
substance use disorder and for providing decision support to
healthcare professionals to implement a treatment regimen
recommendation and mitigate any potential risk of a substance use
disorder.
Inventors: |
Donaldson; Keri;
(Hummelstown, PA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Prescient Medicine Holdings, Inc. |
Hummelstown |
PA |
US |
|
|
Assignee: |
Prescient Medicine Holdings,
Inc.
Hummelstown
PA
|
Family ID: |
1000005553529 |
Appl. No.: |
17/245440 |
Filed: |
April 30, 2021 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
17004838 |
Aug 27, 2020 |
10998105 |
|
|
17245440 |
|
|
|
|
62892212 |
Aug 27, 2019 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G16H 50/20 20180101;
G16H 10/60 20180101; G16H 20/10 20180101; G16H 50/30 20180101 |
International
Class: |
G16H 50/30 20060101
G16H050/30; G16H 20/10 20060101 G16H020/10; G16H 10/60 20060101
G16H010/60; G16H 50/20 20060101 G16H050/20 |
Claims
1. A non-transitory machine-readable storage medium storing
processor-executable instructions that, when executed by at least
one processor, cause the at least one processor to: in response to
receipt of a sample from a patient, determine if the sample
contains a specified set of allelic variants; in response to a
determination that the sample does not include the specified set of
allelic variants, transmit a response to a user indicating that a
score indicative of risk to a substance use disorder is not
available; in response to a determination that the sample includes
the specified set of allelic variants, determine a first plurality
of SNP (single-nucleotide polymorphism) profiles for each of the
specified set of allelic variants in the sample; and determine,
using a machine learning ensemble model, the score indicative of
risk to a substance use disorder for the patient based on the first
plurality of SNP profiles for each of the specified set of allelic
variants in the sample, wherein the machine learning ensemble model
is trained with inputs comprising a second plurality of SNP
profiles associated with the specified set of allelic variants from
a first plurality of test subjects and a third plurality of SNP
profiles associated with the specified set of allelic variants from
a second plurality of test subjects and outputs comprising a first
plurality of substance use indicators specifying that the first
plurality of test subjects have been diagnosed with the substance
use disorder and a second plurality of substance use indicators
specifying that the second plurality of test subjects have not been
diagnosed with the substance use disorder.
2. The non-transitory machine-readable storage medium of claim 1,
wherein clinical data of the patient is provided along with the
sample.
3. The non-transitory machine-readable storage medium of claim 1,
wherein the score indicative of risk to the substance use disorder
for the patient from the machine learning ensemble model is a value
between 0 and 1.
4. The non-transitory machine-readable storage medium of claim 3,
further comprising instructions that cause the at least one
processor to: in response to the score based on the first plurality
of SNP profiles for each of the specified set of allelic variants
in the sample being greater than 0.33, return an output indicating
a predisposition to develop the substance use disorder.
5. The non-transitory machine-readable storage medium of claim 3,
further comprising instructions that cause the at least one
processor to: in response to the score based on the first plurality
of SNP profiles for each of the specified set of allelic variants
in the sample is less than or equal to 0.33, return an output
indicating a lower likelihood to develop the substance use
disorder.
6. The non-transitory machine-readable storage medium of claim 1,
wherein the substance use disorder includes one or more of an
opioid use disorder, an alcohol use disorder, a cannabinoid use
disorder, and a cocaine use disorder.
7. A method of determining a subject's risk for developing a
substance use disorder, the method comprising: analyzing a sample
of the subject to obtain one or more SNP profiles; in response to
the one or more SNP profiles each including a set of specified
allelic variants: determining, via a processor of a computing
device and a machine learning ensemble model stored in a
machine-readable storage medium of the computing device, a score
indicating a risk of the subject for developing a substance use
disorder based on the one or more SNP profiles of each of the set
of specified allelic variants; and in response to the score
indicating a high risk for the subject to develop the substance use
disorder: determining, via the processor, a treatment regimen
recommendation for the subject based on the score, and
transmitting, via the processor, the score and the treatment
regimen recommendation to a user device.
8. The method of claim 7, wherein the score indicating the
subject's risk for developing the substance use disorder is a value
from the machine learning ensemble model between 0 and 1, wherein
the machine learning ensemble model includes a pre-determined
threshold, wherein the score being greater than the pre-determined
threshold indicates the subject has a higher risk for developing
the substance use disorder, and wherein the score being less than
or equal to than the pre-determined threshold indicates the subject
has a lower risk for developing the substance use disorder.
9. The method of claim 7, wherein the computing device includes a
SNP analyzer to analyze the sample of the subject to obtain the one
or more SNP profiles of each of a set of specified allelic
variants.
10. The method of claim 7, wherein the method further comprises: in
response to a determination of the score, generating, via the
processor, a report, the report including the treatment regimen
recommendation and the score.
11. The method of claim 10, wherein the method further comprises:
in response to generation of the report, transmitting, via the
processor, the report to a state prescription drug monitoring
program (PDMP) database.
12. The method of claim 7, wherein the method further comprises:
prior to receipt of the sample, determining, via the processor,
whether a prior determined score indicating the subject's risk for
developing a substance use disorder is available for the subject;
and in response to a determination that the prior determined score
is not available for the subject, transmitting, via the processor
of the computing device, a request for another sample of the
subject; in response to a determination that the prior determined
score is available for the subject, determining, via the processor,
a treatment regimen recommendation based on the prior determined
score.
13. The method of claim 12, wherein the prior determined score is
stored in a patient record database containing electronic health
records of the subject.
14. The method of claim 7, wherein the set of specified allelic
variants includes one or more of allelic variants of genes:
serotonin 2A receptor, galanin, ATP binding cassette transporter 1,
catechol-O-methyltransferase, dopamine transporter, dopamine D2
receptor, dopamine D1 receptor, methylene tetrahydrofolate
reductase, dopamine beta hydroxylase, delta opioid receptor, a
first mu opioid receptor (OPRM1), dopamine D4 receptor,
gamma-aminobutyric acid, kappa opioid receptor, and a second mu
opioid receptor (MUOR).
15. The method of claim 7, wherein the set of specified allelic
variants includes allelic variants of genes: serotonin 2A receptor,
galanin, ATP binding cassette transporter 1,
catechol-O-methyltransferase, dopamine transporter, dopamine D2
receptor, dopamine D1 receptor, methylene tetrahydrofolate
reductase, dopamine beta hydroxylase, delta opioid receptor, a
first mu opioid receptor (OPRM1), dopamine D4 receptor,
gamma-aminobutyric acid, kappa opioid receptor, and a second mu
opioid receptor (MUOR).
16. The method of claim 7, wherein a determination of the score is
further based on a subject's clinical data.
17. A non-transitory machine-readable storage medium encoded with
instructions executable by a processing resource, the
non-transitory machine-readable storage medium comprising
instructions to: store one or more sets of data, each of the one or
more sets of data including one or more SNP profiles for each of a
specified set of allelic variants, each of the one or more SNP
profiles associated with a subject and value indicating whether the
subject is diagnosed with substance use disorder, each of the set
of specified allelic variants includes one or more of allelic
variants of genes: serotonin 2A receptor, galanin, ATP binding
cassette transporter 1, catechol-O-methyltransferase, dopamine
transporter, dopamine D2 receptor, dopamine D1 receptor, methylene
tetrahydrofolate reductase, dopamine beta hydroxylase, delta opioid
receptor, a first mu opioid receptor (OPRM1), dopamine D4 receptor,
gamma-aminobutyric acid, kappa opioid receptor, and a second mu
opioid receptor (MUOR); generate one or more subsets of data based
on the one or more sets of data, each subset of data including SNP
profiles associated with a same value indicating whether subjects
are diagnosed with substance use disorder; and train a machine
learning model with training data defined by the one or more
subsets of data, such that the training of the machine learning
model provides a substance use disorder predictor.
18. The non-transitory machine-readable storage medium of claim 17,
wherein the one or more sets of data include clinical data, the
clinical data including age, sex, ethnicity, race, and, if a
subject has been exposed to a specific substance, a time since
first exposure.
19. The non-transitory machine-readable storage medium of claim 17,
wherein each subset of data is pre-processed prior to training the
machine learning model, the pre-processing including normalizing
SNP profiles for each of the specified set of allelic variants
based on importance of an alleles variant on substance use disorder
predictability.
20. The non-transitory machine-readable storage medium of claim 17
wherein the substance use disorder includes opioid use disorder,
alcohol abuse disorder, cannabinoid use disorder, cocaine use
disorder, and nicotine use disorder.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This is a continuation of U.S. Non-Provisional application
Ser. No. 17/004,838, filed Aug. 27, 2020, titled "Methods and
Systems for Evaluation of Risk of Substance Use Disorders," now
U.S. Pat. No. 10,998,105, issued May 4, 2021, which claims priority
to and the benefit of U.S. Provisional Application No. 62/892,212,
filed Aug. 27, 2019, titled "Methods and Systems for Predicting
Risk of Substance Use Disorders," the entire contents of which are
incorporated herein by reference.
TECHNICAL FIELD
[0002] The present disclosure relates generally to the field of
treating patients in response to evaluating a risk of substance use
disorders. More specifically, the present disclosure relates to
systems and associated methods for predicting risk of a substance
use disorder and providing decision support for appropriate
treatment regimens.
BACKGROUND
[0003] Substance abuse and dependence are defined collectively in
the Diagnostic and Statistical Manual of Mental Disorders, 5th
Edition (DSM-5) as substance use disorder (SUD). SUD is a medical
condition in which an individual has a diminished ability to
control the use of one or more legal or illegal substances. These
substances include alcohol, cannabis, opioids, and nicotine. In a
2017 National Survey on Drug Use and Health, a substance use
disorder involving the use of alcohol or illicit drugs in the past
year affected approximately 19.7 million people (aged 12 or older).
When these results were further parsed, about 14.5 million people
had an alcohol use disorder and 7.5 million people who had an
illicit drug use disorder. While drug use disorder involving
marijuana affected 4.1 million people, opioid use disorder (OUD)
involving prescription pain relievers or heroin affected about 2.1
million people. According to the Centers for Disease Control and
Prevention (CDC), the number of deaths in the United States due to
opioid overdose rose more than 10% from 42,249 in 2016 to 47,600 in
2017, with prescription opioid medications accounting for 35.8% of
all opioid-related deaths in 2017. General trends from 2006
indicate that rates of opioid prescription have decreased, but an
estimated 17.4% of the U.S. population filled at least 1
prescription in 2017 with the average person receiving 3.4
prescriptions. Additionally, misuse of these medications continues
to be problematic, as an estimated 11.1 million individuals (12
years or older) misused prescription pain relievers in 2017,
including hydrocodone, oxycodone, and fentanyl products.
[0004] Different strategies have been implemented to help alleviate
this public health crisis, including publication of opioid
prescribing guidelines by the CDC in 2016. There is some evidence
that the guidelines have contributed to decreased prescription
rates for high-dosage opioid medications and lower percentages of
patients with overlapping opioid and benzodiazepine prescriptions.
However, the CDC guidelines have had little effect on opioid
overdose deaths, and have resulted in some unintended consequences,
such as the abrupt discontinuation of opioid medications in
patients who still need pain relief and the dismissal of patients
from medical care. Moreover, several medical associations have
raised concerns that efforts to limit opioid misuse and abuse have
had a detrimental effect of unduly restricting pain management for
patients with legitimate medical need for opioids, including those
with cancer, cancer survivors, and others with debilitating chronic
pain. These unintended consequences underscore the challenge of
balancing the benefits of opioids for pain management and the risk
of developing an opioid use disorder.
SUMMARY
[0005] Applicant recognized the problems noted above and has
conceived and developed systems and associated methods for
predicting risk of a substance use disorder and providing decision
support for appropriate treatment regimens. Certain embodiments
include methods for predicting risk of a substance use disorder
using machine learning models and associated systems for providing
decision support to healthcare professionals to implement a
treatment regimen and mitigate the risk of the substance use
disorder. Certain embodiments include methods and systems for
determining health outcomes and/or healthcare resource utilization
based on evaluating the risk of a substance use disorder using
machine learning models. In certain embodiments, the systems and
methods described herein are used by an insurance entity to provide
insurance coverage for a treatment regimen. In certain embodiments,
the systems and methods described herein are used by a prescription
drug monitoring program (PDMP). These embodiments include storing
an individual's risk of substance use disorders as part of the
individual's medical records. A health care provider, such as a
physician or a nurse or a pharmacist, can access the individual's
risk of substance use disorder in the PDMP to inform their opioid
prescribing or opioid dispensing decisions, and therefore can
either protect an individual at risk by providing a non-opioid
option or provide an opioid-regimen to an individual who would
benefit from the opioid.
[0006] Embodiments include systems to determine a subject's risk
for developing a substance use disorder. One such system includes
the following components: (a) a SNP analyzer to analyze a sample
from a patient and provide a first plurality of SNP profiles of
each of a set of specified allelic variants in the sample; (b) a
processor and a machine-readable storage medium storing (i) a
trained ensemble model, the trained ensemble model being trained
with inputs comprising a second plurality of SNP profiles
associated with the specified set of allelic variants from a first
plurality of test subjects and a third plurality of SNP profiles
associated with the specified set of allelic variants from a second
plurality of test subjects and outputs comprising a first plurality
of substance use indicators specifying that the first plurality of
test subjects have been diagnosed with the substance use disorder
and a second plurality of substance use indicators specifying that
the second plurality of test subjects have not been diagnosed with
the substance use disorder, and (ii) instructions, when executed by
the processor, are configured to provide the trained ensemble model
with the first plurality of SNP profiles to generate a score
indicative of risk to a substance use disorder for the patient
based on the first plurality of SNP profiles; in response to the
score based on the first plurality of SNP profiles being greater
than a pre-determined threshold, transmit to a user interface an
output indicating a predisposition of the patient to develop the
substance use disorder; and in response to the score based on the
first plurality of SNP profiles being less than or equal to the
pre-determined threshold, transmit to the user interface an output
indicating a lower likelihood of the patient to develop the
substance use disorder. The system can further include a database
storing a non-opioid based treatment regimen recommendation and an
opioid-based treatment regimen recommendation. In an embodiment,
the machine-readable storage medium further can include
instructions to: in response to the score based on the first
plurality of SNP profiles being greater than the pre-determined
threshold, retrieve the non-opioid based treatment regimen
recommendation from the database; and transmit to the user
interface an output indicating the predisposition of the patient to
develop the substance use disorder and the non-opioid based
treatment regimen recommendation. In an embodiment, the
machine-readable storage medium further can include instructions
to: in response to the score based on the first plurality of SNP
profiles being less than or equal to the pre-determined threshold,
retrieve the opioid based treatment regimen recommendation from the
database; and transmit to the user interface an output indicating
the predisposition of the patient to develop the substance use
disorder and the opioid based treatment regimen recommendation. In
an embodiment, a treatment regimen recommendation can include a
reduced dose of the opioid. In an embodiment, a treatment regimen
recommendation can include a reduced dose of the opioid in
combination with the non-opioid drug or other interventions. In an
embodiment, a treatment regimen recommendation can include a
limited time period for use of the opioid drug followed by extended
use of the non-opioid drug or other interventions. For example, a
treatment regimen recommendation can include use of an opioid drug
for no more than 3 days, then use a non-opioid drug.
[0007] In an embodiment, the trained ensemble model includes two or
more trained models. The trained ensemble model can be generated by
at least one of a random tree model and a support vector machine
model. In an embodiment, the sample is pre-processed prior to using
the trained ensemble model. In an embodiment, the pre-processing
includes ordinal encoding or one-hot encoding. A sample includes
any information from a biological material, such as one or more of
a SNP profile, other allelic variations, RNA expression data, or
other genetic information.
[0008] An embodiment of a method of determining a subject's risk
for developing a substance use disorder includes the step of
analyzing a sample of the subject to obtain one or more SNP
profiles of each of a set of specified allelic variants. In
response to the one or more SNP profiles of each of the set of
specified allelic variants, the method further includes
determining, via a processor of a computing device and a trained
ensemble model stored in a machine-readable storage medium of the
computing device, a score indicating the subject's risk for
developing a substance use disorder based on the one or more SNP
profiles of each of the set of specified allelic variants. The
method further includes determining, via the processor, a treatment
regimen recommendation for the subject based on the score, and
transmitting, via the processor, the score and the treatment
regimen recommendation to a user device. In an embodiment, the
score indicating the subject's risk for developing the substance
use disorder is a value from the trained ensemble model between 0
and 1. In an embodiment, the trained ensemble model includes a
pre-determined threshold, and the score being greater than the
pre-determined threshold indicates the subject has a higher risk
for developing the substance use disorder, and the score being less
than or equal to than the pre-determined threshold indicates the
subject has a lower risk for developing the substance use disorder.
The computing device in this method can include a SNP analyzer to
analyze the sample of the subject to obtain the one or more SNP
profiles of each of a set of specified allelic variants. In an
embodiment, the method further comprises the step of generating,
via the processor, a report, the report including the treatment
regimen recommendation and the score. In an embodiment, the method
further comprises the step of transmitting, via the processor, the
report to a state prescription drug monitoring program (PDMP)
database. In an embodiment, the method further includes the
following steps: prior to receipt of the sample, determining, via
the processor, whether a prior determined score indicating the
subject's risk for developing a substance use disorder is available
for the subject; and in response to the determination that the
prior determined score is not available for the subject,
transmitting, via the processor of the computing device, a request
for a sample of the subject; and in response to the determination
that the prior determined score is available for the subject,
determining, via the processor, a treatment regimen recommendation
based on the prior determined score. In an embodiment, the prior
determined score is stored in a patient record database containing
electronic health records of the subject. In an embodiment, the set
of specified allelic variants includes one or more of allelic
variants of the following genes: serotonin 2A receptor, galanin,
ATP binding cassette transporter 1, catechol-O-methyltransferase,
dopamine transporter, dopamine D2 receptor, dopamine D1 receptor,
methylene tetrahydrofolate reductase, dopamine beta hydroxylase,
delta opioid receptor, a first mu opioid receptor (OPRM1), dopamine
D4 receptor, gamma-aminobutyric acid, kappa opioid receptor, and a
second mu opioid receptor (MUOR). In an embodiment, the set of
specified allelic variants includes allelic variants of the
following genes: serotonin 2A receptor, galanin, ATP binding
cassette transporter 1, catechol-O-methyltransferase, dopamine
transporter, dopamine D2 receptor, dopamine D1 receptor, methylene
tetrahydrofolate reductase, dopamine beta hydroxylase, delta opioid
receptor, a first mu opioid receptor (OPRM1), dopamine D4 receptor,
gamma-aminobutyric acid, kappa opioid receptor, and a second mu
opioid receptor (MUOR). In an embodiment, the determination of the
score is further based on a subject's clinical data.
[0009] Embodiments also include a non-transitory machine-readable
storage medium storing processor-executable instructions that, when
executed by at least one processor, cause the at least one
processor to perform the following steps: in response to receipt of
a sample from a patient, determine if the sample contains a
specified set of allelic variants; in response to a determination
that the sample does not include the specified set of allelic
variants, transmit a response to a user indicating that a score
indicative of risk to a substance use disorder is not available.
The instructions further cause the processor to perform the
following: in response to a determination that the sample includes
the specified set of allelic variants, determine a first plurality
of SNP profiles for each of the specified set of allelic variants
in the sample; and determine, using a trained ensemble model, a
score indicative of risk to a substance use disorder for the
patient based on the first plurality of SNP profiles for each of
the specified set of allelic variants in the sample. The trained
ensemble model is trained with inputs comprising a second plurality
of SNP profiles associated with the specified set of allelic
variants from a first plurality of test subjects and a third
plurality of SNP profiles associated with the specified set of
allelic variants from a second plurality of test subjects and
outputs comprising a first plurality of substance use indicators
specifying that the first plurality of test subjects have been
diagnosed with the substance use disorder and a second plurality of
substance use indicators specifying that the second plurality of
test subjects have not been diagnosed with the substance use
disorder. In an embodiment, the trained ensemble model includes two
or more trained models. The trained ensemble model can be generated
by at least one of a random tree model and a support vector machine
model. In an embodiment, the sample is pre-processed prior to using
the trained ensemble model. In an embodiment, the pre-processing
includes ordinal encoding or one-hot encoding.
[0010] In an embodiment, the substance use indicators used in the
trained ensemble model are a value between 0 and 1. In an
embodiment, the instructions cause the processor to perform: in
response to the score based on the first plurality of SNP profiles
for each of the specified set of allelic variants in the sample
being greater than 0.33, return an output indicating a
predisposition to develop the substance use disorder. In an
embodiment, the instructions cause the processor to perform: in
response to the score based on the first plurality of SNP profiles
for each of the specified set of allelic variants in the sample is
less than or equal to 0.33, return an output indicating a lower
likelihood to develop the substance use disorder. In an embodiment,
the SUD includes one or more of an opioid use disorder, an alcohol
use disorder, a cannabinoid use disorder, and a cocaine use
disorder. In an embodiment, the clinical data of the patient is
provided along with the sample. In an embodiment, the clinical data
includes age, sex, race, and ethnicity. In an embodiment, the
clinical data includes demographic data, socioeconomic data, and
any data that about a subject that can be obtained by observation
or oral or written communication.
[0011] Other aspects and features of the present disclosure will
become apparent to those of ordinary skill in the art after reading
the detailed description herein and the accompanying figures.
BRIEF DESCRIPTION OF DRAWINGS
[0012] The foregoing aspects, features, and advantages of the
present disclosure will be further appreciated when considered with
reference to the following drawings:
[0013] FIGS. 1A through 1E are diagrammatic representations of a
stacked machine learning model, according to an embodiment of the
present disclosure;
[0014] FIGS. 2A through 2C are diagrammatic representations of
systems containing a stacked machine learning model, according to
an embodiment of the present disclosure;
[0015] FIGS. 3A through 3E are flowcharts of different embodiments
of methods of predicting risk of substance use disorders;
[0016] FIG. 4 is a representation of the receiver operating
characteristic curves using (i) genetic data only, (ii) genetic
data and clinical data, and (iii) genetic data and a subset of
clinical data, according to an embodiment of the present
disclosure;
[0017] FIGS. 5A and 5B are graphical representations of the
prediction distribution based on genetic data only and a
combination of genetic data and clinical data, respectively,
according to an embodiment of the present disclosure; and
[0018] FIGS. 6A and 6B are lift charts of the ensemble predictive
model developed by using genetic data only and a combination of
genetic data and clinical data, respectively, according to an
embodiment of the present disclosure.
DETAILED DESCRIPTION
[0019] The foregoing aspects, features, and advantages of the
present disclosure will be further appreciated when considered with
reference to the following description of the embodiments and
accompanying drawings. In describing the embodiments of the
disclosure illustrated in the appended drawings, specific
terminology will be used for the sake of clarity. The disclosure,
however, is not intended to be limited to the specific terms used,
and it is to be understood that each specific term includes
equivalents that operate in a similar manner to accomplish a
similar purpose. Numerous specific details, examples, and
embodiments are set forth and described to provide a thorough
understanding of various embodiments of the present invention.
However, in certain instances, well-known or conventional details
are not described in order to provide a concise discussion of
embodiments of the present disclosure.
[0020] When introducing elements of various embodiments of the
present disclosure, the articles "a," "an," "the," and "said" are
intended to mean that there are one or more of the elements. The
terms "comprising," "including," and "having" are intended to be
inclusive and mean that there may be additional elements other than
the listed elements. Any examples of operating parameters and/or
environmental conditions are not exclusive of other
parameters/conditions of the disclosed embodiments. Additionally,
it should be understood that references to "one embodiment", "an
embodiment," "certain embodiments," or "other embodiments" of the
present disclosure are not intended to be interpreted as excluding
the existence of additional embodiments that also incorporate the
recited features. Furthermore, reference to terms such as "above,"
"below," "upper," "lower," "side," "front," "back," or other terms
regarding orientation are made with reference to the illustrated
embodiments and are not intended to be limiting or exclude other
orientations.
[0021] Embodiments include computer implemented methods for
predicting risk of substance use disorders in a patient using a
computer system having one or more processors coupled to a memory
storing one or more computer readable instructions for execution by
the one or more processors. One such method includes the steps of
storing a set of data containing a plurality of subject records.
The subject records can include clinical data of each of the
subject, such as age, sex, race, prior medical history, and
ethnicity. In an embodiment, the clinical data includes demographic
data, socioeconomic data, and any data that about a subject that
can be obtained by observation or oral or written communication.
The subject records can also include data about health outcomes and
healthcare resource utilization, such as quantification or
description of the use of services by a subject for the purpose of
preventing and curing substance use disorders or underlying
disease, promoting maintenance of health and well-being, or
obtaining information about one's health status and prognosis. For
example, an increased risk of substance use disorder may include
poorer health outcomes and greater utilization of healthcare
resources. Each subject record includes a SNP (single nucleotide
polymorphism) profile for each individual of a plurality of test
subjects and subjects not diagnosed with a substance use disorder.
Each subject record can also include a plurality of physical
characteristics for each individual of a plurality of test subjects
and subjects not diagnosed with a substance use disorder. The
subject record for each of the plurality of test subjects also
includes a first substance use indicator indicating that the test
subject has been diagnosed with a substance use disorder. The
subject record for each of the plurality of subjects not diagnosed
with a substance use disorder includes a second substance use
indicator indicating the subject has not been diagnosed with a
substance use disorder. The method further includes the step of
selecting a first subset of the subject records for the plurality
of test subjects and a second subset of the subject records for the
plurality of subjects not diagnosed with a substance use disorder.
The first and second subsets do not include the first substance use
indicator and the second substance use indicator, respectively. The
first and second subsets serve as inputs into one or more initial
machine learning models. The method further includes the step of
determining whether the one or more initial machine learning models
meet a predetermined sensitivity, a predetermined specificity, and
a predetermined accuracy for predicting the first substance use
indicator for the test subjects and the second substance use
indicator of the subjects not diagnosed with a substance use
disorder. The method further includes the step of generating an
ensemble machine learning model responsive to the one or more
initial machine learning models that meet the predetermined
sensitivity, the predetermined specificity, and the predetermined
accuracy. The method further includes the step of supplying a SNP
profile of a patient to the ensemble machine learning model as an
input to obtain a score indicative of the patient's risk of a
substance use disorder. The method further includes the step of
presenting the score or other indicator indicating the patient's
risk of developing a substance use disorder to a healthcare
professional for decision support to implement a treatment regimen
and mitigate the patient's risk of substance abuse or substance
dependence.
[0022] In another embodiment, a non-transitory machine-readable
storage medium may be encoded with instructions executable by a
processing resource. The non-transitory machine-readable storage
medium includes instructions to store one or more sets of data.
Each set of data may include one or more allelic variants. Each of
the one or more allelic variants are associated with a subject.
Each of the one or more allelic variants includes one or more SNP
profiles of the subject and a value indicating whether the subject
is diagnosed with a substance use disorder (SUD). The
non-transitory machine-readable storage medium can include
instructions to generate one or more subsets of data based on the
sets of data. Each subset of data may include a plurality of SNP
profiles. Each of the plurality of SNP profiles may include a same
value indicating whether the subjects of the plurality of SNP
profiles are diagnosed with SUD. The non-transitory
machine-readable storage medium can include instructions to train a
stacked model with training data generated via creation of the
subsets of data. The stacked model may include, at least, a random
forest model and a support vector machine model. Each training data
may correspond to one of the one or more subsets of data. Each of
the one or more subsets of data may correspond to the same value
indicating whether the subject is diagnosed with SUD, such that the
training of the stacked model provides a score or other indicator
that informs a patient's risk of a substance use disorder.
[0023] In another embodiment, a non-transitory machine-readable
storage medium may be encoded with instructions executable by a
processing resource. The non-transitory machine-readable storage
medium can include instructions to retrieve from memory, a trained
stacked model. The stacked model may include, at least, a random
forest model and a support vector machine model. The stacked model
may be trained with one or more inputs and outputs derived from one
or more sets of data. The one or more inputs may include one or
more SNP profiles. Each of the one or more SNP profiles may be
associated with one or more allelic variants of the subject. The
one or more outputs may include a value indicating whether the
subject has a risk of developing a SUD. The non-transitory
machine-readable storage medium can include instructions to, in
response to receipt of an input of a sample, determine if the
sample includes the one or more alleles associated with the SNP
profiles. The non-transitory machine-readable storage medium can
include instructions to, in response to a determination that the
sample does not include one of the one or more alleles associated
with the SNP profiles, send a response to a user indicating that a
prediction is not available. The non-transitory machine-readable
storage medium can include instructions to, in response to a
determination that the sample does include the one or more alleles
associated with the SNP profiles, determine, using the retrieved
trained stacked model and an average of the results of the random
tree model and support vector machine model, a prediction of
disposition to SUD relative to SNP profiles of the one or more
alleles. A plurality of inputs to the retrieved trained stacked
model may include all possible SNP profiles associated to the
plurality of allelic variants.
[0024] Another embodiment of the disclosure is directed to a method
of determining a subject's risk of developing a substance use
disorder. This risk may be expressed as a score or a probability or
other indicator that informs a patient's risk of a substance use
disorder. In certain embodiments, the risk is presented as a binary
determination, such as one of Yes or No or one of Low Risk or High
Risk. In certain embodiments, the risk is presented as one of small
set of possibilities, such as one of low risk, intermediate risk,
or high risk. In certain embodiments, the risk is presented as a
score ranging from 0 to 1 or from 0 to 100. The method includes the
steps of receiving a sample of the subject; analyzing the sample,
via a processor of a computing device, to obtain a SNP profile of
each of a set of specified allelic variants. In response to the SNP
profiles of each of the set of specified allelic variants, the
method further includes the step of determining, via the processor
and a trained stacked model of the computing device, a score
indicating the subject's risk of developing a substance use
disorder based on the SNP profiles of each of the set of specified
allelic variants. The method further includes the step of
determining, via the processor, a treatment regimen recommendation
for the subject based on the score; and transmitting, via the
processor, the score and treatment regimen recommendation to a user
device.
[0025] In an embodiment, the score is a value output from the
trained stacked model between 0 and 1 and the trained stacked model
includes a pre-determined threshold. In an embodiment, the score
with a value greater than the threshold indicates the subject has a
high risk of developing a substance use disorder, and a score with
a value less than or equal to the threshold indicates the subject
has a low risk of developing a substance use disorder. In an
embodiment, the set of specified allelic variants includes two or
more of fifteen allelic variants. The fifteen allelic variants
include allelic variants of the following genes: serotonin 2A
receptor, galanin, ATP binding cassette transporter 1,
catechol-O-methyltransferase, dopamine transporter, dopamine D2
receptor, dopamine D1 receptor, methylene tetrahydrofolate
reductase, dopamine beta hydroxylase, delta opioid receptor, a
first mu opioid receptor (OPRM1), dopamine D4 receptor,
gamma-aminobutyric acid, kappa opioid receptor, and a second mu
opioid receptor (MUOR). In an embodiment, the set of specified
allelic variants includes all fifteen allelic variants.
[0026] The method may include transmitting, via one or more
processors of a computing device, a request for a sample of the
subject. The method may include determining, via the one or more
processors, a treatment regimen recommendation for the subject
based on the score and transmitting, via the one or more
processors, the score and treatment regimen recommendation to a
user device and to the subject's electronic medical records (EMR).
In an embodiment, the score is a value output from the trained
stacked model between 0 and 1 and the score with a value greater
than the pre-determined threshold indicates the subject has a high
risk of developing a substance use disorder, and a score with a
value less than or equal to the pre-determined threshold indicates
the subject has a low risk of developing a substance use
disorder.
[0027] Another embodiment of the disclosure is directed to a system
to determine a subject's risk of developing a substance use
disorder. The system can include an analyzer with an input for a
sample. In an embodiment, the analyzer conducts an allelic analysis
of the sample received at the input to generate the SNP profiles
for the trained ensemble model. The analysis, based on the received
sample, may provide SNP profiles of each of a set of specified
allelic variants. The system can include a trained ensemble model,
such as a trained stacked model. The trained ensemble model may
include one or more inputs and an output. The one or more inputs
may accept one or more SNP profiles from the analyzer. Based on the
SNP profiles, the trained ensemble model may provide, at the
output, a score or other indicator that informs a patient's risk of
a substance use disorder. The system may include a user interface
input/output in communication with a processor and machine-readable
storage medium. The machine-readable storage medium may store
instructions and the instructions may be executable by the
processor. The executed instruction may, in response to a risk
output from the trained ensemble model, determine an indicator to
indicate a subject's risk of developing substance use disorder, the
indicator based on the risk or probability. Further, the
instructions may include transmission of the indicator to a user
interface associated with a user device.
[0028] The embodiments described herein may utilize a trained
ensemble or trained stacked model to determine a predicted
predisposition to a substance use disorder of a subject. This
approach may allow for prevention or modification of prescription
of specific medications to individuals predisposed to a substance
use disorder, thus preventing potential dependence or addiction
through prescription. For example, upon receipt of a sample, a
system may provide a prediction of the likelihood of a patient
developing a substance use disorder within, depending on various
factors, hours or days. Certain embodiments include methods and
systems for determining health outcomes and/or healthcare resource
utilization based on evaluating the risk of a substance use
disorder using machine learning models. For example, an increased
risk of substance use disorder may include poorer health outcomes
and greater utilization of healthcare resources. Certain
embodiments include methods and systems for determining health or
wellness of the subject based on the analysis of the allelic
variants and/or clinical data of the subject. Certain embodiments
include methods and systems for determining health outcomes of the
subject based on the analysis of the allelic variants and/or
clinical data of the subject. Certain embodiments include methods
and systems for determining healthcare resource utilization by a
subject based on the analysis of the allelic variants and/or
clinical data of the subject.
[0029] In an embodiment, a system 201 (as illustrated in FIGS. 2B,
and 2C) for predicting risk of substance use disorders and/or for
providing decision support to healthcare professionals to implement
a treatment regimen and mitigate substance use includes a user
device 220, an electronic communications network, a SNP analyzer
222, a patient record database 224, a processor 204 to execute
instructions stored on a non-transitory machine readable storage
medium 206, the instructions to, when executed by the processor
204, may retrieve and/or utilize a trained ensemble and/or trained
stacked model 100 stored in a memory 214 (e.g., the computing
device 202 may receive or retrieve the trained stacked model 100
and store the trained stacked model 100 in the machine readable
storage medium 206). In an embodiment, the system 201 for
predicting risk of substance use disorders can include decision
support modules (for example, included as machine readable
instructions stored in machine readable storage medium 206) coupled
to a patient record database 224 and/or one or more of a
prescription drug database, the patient's electronic health
records, and drug utilization records. Decision support modules may
include providing evidence-based information to the healthcare
professional regarding drug selection and dosing, potential adverse
drug reactions and drug allergies, duration or form of prescribed
drugs, or stratification of treatments. The computing device 202
may include an input/output to connect to a user device 220 or user
interface. The user device 220 or user interface may include a
display 216. The display 216 may be configured to display a
graphical user interface (GUI) for receiving a SNP profile and one
or more electronic medical records (EMR) or electronic health
records (EHR) associated with a patient and displaying a risk of a
substance use disorder. The GUI may display one or more
recommendations for treatment regimens for the patient. The
electronic communications network is in communication with the user
device 220, the SNP analyzer 222, the patient record database 224,
and the processor 204 that executes the trained ensemble and/or
trained stacked model 100. In another example as shown in FIG. 2C,
the computing device 202 may be or include a SNP analyzer 222 or
another genetic information analyzer. Further, rather than connect
to the trained stacked model 100, the computing device 202 can
include the trained stacked model 100 (e.g., the trained stacked
model 100 may be stored in the machine-readable storage medium
206).
[0030] In an embodiment, a system 200 (as illustrated in FIG. 2A)
may include a computing device 203 including a machine-readable
storage medium 207 and a processor 205. The machine readable
storage medium 207 includes instructions, when executed by the
processor 205, to receive a plurality of SNP profiles and/or
substance use disorder profiles associated to the plurality of test
subjects and subjects not diagnosed with a substance use disorder
from the user device 221 or patient record database 224, train one
or more predictive machine learning models (for example, stacked
model 210 or a first random forest model 105, second random forest
model 107, and a SVM model 109) using the plurality of physical
characteristics and associated SNP profiles, thereby generating an
ensemble predictive model (for example, a trained stacked model
100) for determining risk of substance use disorder. The
machine-readable storage medium 207 may include instructions, when
executed by the processor 205, to receive a plurality of physical
and/or clinical data from a plurality of EMR associated to a
plurality of test subjects and subjects not diagnosed with a
substance use disorder from a user device 221 or patient record
database 224.
[0031] The machine-readable storage medium 207 may also include
instructions to retrieve one or more treatment regimen
recommendations associated with a particular risk of substance use
disorder. The data (as in, the training data 102) used to train the
one or more predictive machine learning models (e.g., a stacked
model 210) may include a substance use disorder profile. The
substance use disorder profile may include one or more of a
plurality of physical or clinical data from a plurality of EMR
associated to a plurality of test subjects and subjects not
diagnosed with a substance use disorder, a plurality of SNP
profiles associated to the plurality of test subjects and subjects
not diagnosed with a substance use disorder, a plurality of medical
diagnoses regarding substance use disorders associated to the
plurality of test subjects and subjects not diagnosed with a
substance use disorder, and/or a plurality of treatment regimen
recommendations.
[0032] The machine-readable storage medium 207 may include
instructions to assign a plurality of weight values to the training
data 102 within a plurality of substance use disorder profiles,
such as to a plurality of associated SNP profiles of the training
data, determine the one or more weighted data inputs of the
training data, thereby to form an initial set of decision making
rules. The machine-readable storage medium 207 may include
instructions to compare the initial set of decision-making rules to
the one or more weighted data inputs of the training data and
re-weight the initial set of decision-making rules based on the
comparison to create a re-weighted set of decision-making rules.
The re-weighted set of decision-making rules and the one or more
weighted data inputs of the training data may be used to update the
initial predictive model to generate an ensemble predictive model
for determining risk of substance use disorder.
[0033] For example, the training data 102 may be comprised of a set
of data (as illustrated in FIG. 1B) including data associated with
a group of individuals or subjects known to have a substance use
disorder and a group of individuals or subjects with no known
substance use disorders. Various models may be utilized to create a
classification, prediction, or probability of developing substance
use disorder, based on SNP profiles and/or physical and/or clinical
data. Models and methods may include decision trees, random forest
models, random forests utilizing bagging or boosting (as in,
gradient boosting), neural network methods, support vector machines
(SVM), other supervised learning models, other semi-supervised
learning models, and/or other unsupervised learning models, as will
be readily understood by one having ordinary skill in the art.
Further, a stacked model 210 or an ensemble machine learning model
may be utilized. A stacked model 210 may include two or more
learning models, for example, a first random forest model 105, a
second random forest model 107, a SVM model 109, and/or other
models as will be readily understood by one having ordinary skill
in the art.
[0034] As illustrated in FIGS. 1A and 1C through 1E, the trained
stacked models 100 may include a first random forest classifier 106
based on results from a first random forest model 105, a second
random classifier 108 based on a second random forest model 107,
and a SVM classifier 110 based on a SVM model 109. The first random
forest model 105 and second random forest model 107 may be trained
utilizing a re-sampling or bootstrapping technique. In other words,
samples from the training data 102 may be re-used or left out for
new decision trees or sets of decision trees. As illustrated in
FIG. 1B, the training data 102 may include an 80% training 116 and
20% testing 118 split (other split percentages, such as 70/30 may
be utilized, as will be readily understood by one having ordinary
skill in the art). In other words, the random forest may be trained
with the 80% 116 of the training data 102, while the 20% 118 of the
training data 102 may be held out for testing the newly trained
model. In another example, the stacked model 210 may include a
gradient boosted tree model. The gradient boosted tree model may
include similar training methods. The gradient boosted tree model
may include logic that measures error based on small changes to
predictions. In other words, the target outcome for each case may
depend on how changing a case's prediction impacts prediction
error. Other models may be utilized, for example, linear regression
models, logistic regression models, naive Bayes models, kNN models,
k-Means models, dimensionality reduction models, other gradient
boosting models, and/or neural networks, as will be readily
understood by one having ordinary skill in the art. In another
example, ensemble model training may utilize n fold
cross-validation. In other words, the set of data may be split n
times (for example, 5 fold or 20 fold cross-validation). While each
split trains a model, another split may be leftover. The trained
model may test against the leftover split. The model may be trained
again, but a different split may be held out for testing. This may
occur until all the training data 102 is utilized for training and
testing. Such a training and testing method may prevent or lessen
the likelihood of over-fitting. Over-fitting may occur when a model
exhibits random error or noise instead of an underlying
relationship. As noted, cross-validation may prevent or lessen the
likelihood of over-fitting. Other methods to avoid over-fitting may
be utilized, as will be readily understood by one having ordinary
skill in the art.
[0035] Training a model, may result in a classifier, predictor,
score, or probability. A sample 101 (for example, a SNP profile
obtained from a biological material, such as from a buccal swab or
a blood sample (or other method of obtaining DNA or RNA), other
allelic variations, RNA expression data, and/or physical or
clinical data may be entered via an input device 218 and/or a SNP
analyzer 222 into a computing device) may be applied to the
predictor to produce a probability or prediction score 114 between
0 and 1 or some other arbitrary numerical range. Multiple models or
ensemble models may be stacked to further prevent over-fitting or
over confident predictions. The average 112 of multiple models may,
when those models disagree, tend to average out towards a mean of
the original dataset.
[0036] As noted, the machine readable storage medium 206 may
include instructions to determine risk or probability of
development of substance use disorder for a particular patient
using the trained ensemble and/or trained stacked model 100
responsive to the plurality of physical or clinical data and/or the
associated SNP profile of the particular patient, identify the one
or more associated recommendations responsive to the risk of
substance use disorder, and transmit the risk of substance use
disorder and the one or more associated treatment regimen
recommendations to the user device 220 via the electronic
communication network to be displayed on the GUI. This risk may be
expressed as a score or a probability or other indicator that
informs a patient's risk of a substance use disorder. In certain
embodiments, the risk is presented as a binary determination, such
as one of Yes or No or one of Low Risk or High Risk. In certain
embodiments, the risk is presented as one of small set of
possibilities, such as one of low risk, intermediate risk, or high
risk. In certain embodiments, the risk is presented as a score
ranging from 0 to 1 or from 0 to 100. The user device 220 may be
configured to display a GUI for displaying the risk of substance
use disorder and the one or more associated treatment regimen
recommendations. For example, the user device 220 may include a
mobile device (e.g. smart phone, tablet, etc.), a desktop computer,
a laptop, a wearable computing device, or other type of computing
device, as will be readily understood by one having ordinary skill
in the art. The user device 220 may further include a mobile
computer application.
[0037] As noted, the output of each of the models may be a value
from 0 to 1 or some other arbitrary number. As noted, an average of
the outputs may be taken, as a correction for potential
disagreements in the models. In a further example, an allocated
threshold may indicate predisposition to develop substance use
disorder, such as an output greater than 0.33, while a value less
than or equal to 0.33 may indicate no or lower likelihood or
predisposition to develop substance use disorder. Further, the
allocated threshold may be set during the training of the stacked
model 210. For example, the machine-readable storage medium 207 may
include instructions to automatically set the allocated threshold
or include a predetermined value for a threshold. In another
example, various indicators may be associated with various
thresholds. For example, an indicator may include, "YES", "NO",
"LOW", "HIGH", "INTERMEDIATE", some other indicator to indicate
level of risk associated with substance use disorder, and/or some
combination thereof. Further, "YES" may indicate that a patient is
predisposed to develop substance use disorder, while "NO" may
indicate that the patient is not predisposed to develop substance
use disorder. "LOW" may indicate that a patient has a low risk of
substance use disorder, "INTERMEDIATE" may indicate that a patient
has a level of risk somewhere in between "HIGH" and "LOW" (but not
past a threshold to indicate a high risk or low risk), and "HIGH"
may indicate that a patient has a high level of risk of substance
use disorder. In another example, the indicator may be presented,
via the user device 220, with or without the percentage of risk (in
other words, the output of the average of the trained stacked model
100).
[0038] In an embodiment, the machine-readable storage medium
further can include instructions the processor to: in response to
the score that informs a patient's risk of a substance use disorder
being greater than the pre-determined threshold, retrieve the
non-opioid based treatment regimen recommendation from the
database; and transmit to the user interface an output indicating
the predisposition of the patient to develop the substance use
disorder and the non-opioid based treatment regimen recommendation.
In an embodiment, the machine-readable storage medium further can
include instructions to: in response to the score that informs a
patient's risk of a substance use disorder being less than or equal
to the pre-determined threshold, retrieve the opioid based
treatment regimen recommendation from the database; and transmit to
the user interface an output indicating the predisposition of the
patient to develop the substance use disorder and the opioid based
treatment regimen recommendation. In an embodiment, a treatment
regimen recommendation can include a reduced dose of the opioid. In
an embodiment, a treatment regimen recommendation can include a
reduced dose of the opioid in combination with the non-opioid drug
or other interventions. In an embodiment, a treatment regimen
recommendation can include a limited time period for use of the
opioid drug followed by extended use of the non-opioid drug or
other interventions. For example, a treatment regimen
recommendation can include use of an opioid drug for no more than 3
days, then use of a non-opioid drug.
[0039] The present disclosure also discloses embodiments directed
to a method for predicting a patient's risk of developing a
substance use disorder or for providing decision support to
healthcare professionals to implement a treatment regimen on a
graphical user interface (GUI) of a computing system using an
ensemble or stacked machine learning model. In an example, training
data 102 or the sample 101 may be pre-processed by a pre-processing
module 104 or pre-processing instructions stored on the computing
device 203 and/or computing device 202, respectively, or at the
memory 215 where the stacked model 210 is stored or the memory 214
where the trained stacked model 100 is stored. In one such method,
a plurality of physical or clinical data and associated SNP profile
of a particular patient may be processed to a format for
processing, such as normalization, one-hot encoding, and/or ordinal
encoding, by one or more machine learning models, and storing the
plurality of physical data and associated SNP profile of the
particular patient to a data structure in a patient records
database 224. The method may further include training one or more
initial predictive models for predicting a patient's risk of
developing a substance use disorder using the plurality of physical
or clinical data and/or associated SNP profile of a plurality of
subjects (with and without diagnoses of substance use disorder)
constituting the training data 102, thereby generating an updated
predictive model for predicting risk of substance use disorders or
for providing decision support to healthcare professionals to
implement a treatment regimen. The method may additionally include
assigning a plurality of weight values to the one or more data
inputs of the training data 102, determining the one or more
weighted data inputs of the training data 102, thereby to form an
initial set of decision making rules, and comparing the initial set
of decision making rules to the one or more weighted data inputs of
the training data 102. In an embodiment, the clinical data includes
age, sex, race, and ethnicity. In an embodiment, the clinical data
includes demographic data, socioeconomic data, and any data that
about a subject that can be obtained by observation or oral or
written communication.
[0040] Further, the method may include, for example, re-weighting
the initial set of decision-making rules based on the comparison to
create a re-weighted set of decision-making rules. The re-weighted
set of decision-making rules and the one or more weighted data
inputs of the training data may be used to update the initial
predictive model to generate the updated predictive model for
predicting risk of substance use disorders or for providing
decision support to healthcare professionals to implement a
treatment regimen.
[0041] An embodiment of the method includes the step of checking a
PDMP to determine if a patient's risk of developing a substance use
disorder has been determined by the methods described herein. In an
embodiment, the PDMP is consulted by a user, such as a doctor or
pharmacist or other healthcare professional, for a patient who is
prescribed a particular drug (e.g., a pre-op team for a patient who
has chosen an elective surgery or a doctor prescribing an opioid).
If the PDMP records indicate that the patient has been evaluated
for his/her risk of developing a substance use disorder, then the
drug prescription is provided in response to his/her risk of
developing a substance use disorder. For example, a non-opioid
option is provided to the patient if his/her risk of developing a
substance use disorder is high. Or, an opioid regimen is provided
to an individual who would benefit from the opioid regimen and has
a low risk of developing a substance use disorder. If the patient
has not been evaluated for his/her risk of developing a substance
use disorder, then the methods include the steps of collecting a
sample from the patient and evaluating the patient's risk of
developing a substance use disorder. If the patient's risk is high,
then a healthcare provider or a pharmacist reviews the results with
the patient and a treatment regimen is developed. The patient is
offered a limited opioid prescription, such as one to last less
than three days, or is provided with non-opioid alternatives. The
method also includes updating the PDMP records with the patient's
risk of developing a substance use disorder and the resulting
treatment regimen. An opioid regimen includes one or more of a
prescription opioid, such as hydrocodone, oxycodone, oxymorphone,
morphine, codeine, or fentanyl. Non-opioid alternatives or
non-opioid treatment regimens can include drug-based and/or
non-drug-based options. Drug-based options include disease-specific
treatments when available (e.g., triptans for migraines or
gabapentin/pregabalin/duloxetine for neuropathic pain),
interventional therapies (e.g., corticosteroid injections or nerve
blocks), acetaminophen, non-steroidal anti-inflammatory drugs,
tricyclic antidepressants, serotonin/norephinephrine reuptake
inhibitors, or topical agents (lidocaine or capsaicin). Examples of
non-drug-based options, include acupuncture, chiropractic
approaches, osteopathic manipulative treatment, massage therapy,
electrical nerve stimulation, occupational therapy, and physical
therapy.
[0042] In an embodiment, an insurance payor can access a PDMP to
determine if a patient's risk of developing a substance use
disorder has been determined by the methods described herein. If
the patient has been evaluated for his/her risk of developing a
substance use disorder, then insurer authorizes coverage of a drug
prescription in response to his/her risk of developing a substance
use disorder. For example, a non-opioid option is provided to the
patient if his/her risk of developing a substance use disorder is
high. Or, an opioid regimen is provided to an individual who would
benefit from the opioid regimen and has a low risk of developing a
substance use disorder. If the patient has not been evaluated for
his/her risk of developing a substance use disorder, then the
insurance payor requests processing a sample from the patient and
evaluating the patient's risk of developing a substance use
disorder. If the patient's risk is high, then the payor informs the
patient's healthcare provider or pharmacist to review the results
with the patient and develop a treatment regimen. The patient is
offered a limited opioid prescription, such as one to last less
than three days, or is provided with non-opioid alternatives. The
method also includes updating the PDMP records with the patient's
risk of developing a substance use disorder and the resulting
treatment regimen. In an embodiment, the insurance company requests
an evaluation of the patient's risk of developing a substance use
disorder be performed in the context of a coverage determination
(i.e., whether the insurance company will pay for whatever
medication therapy is selected).
[0043] FIGS. 3A through 3E illustrate flow diagrams, implemented in
a computing device, to predict a subject's or patient's
predisposition to substance use disorder (in other words, a
patient's risk of developing a substance use disorder), according
to an embodiment. The method is detailed with reference to the
computing device 202. The actions of methods 300, 322, and 342, may
be completed by the computing device 202. Specifically, methods
300, 322, and 342 may be included in one or more programs,
protocols, or instructions loaded into the machine-readable storage
medium 206 of the computing device 202 and executed on the
processor 204 or one or more processors of the computing device
202. The order in which the operations are described is not
intended to be construed as a limitation, and any number of the
described steps may be combined in any order and/or in parallel to
implement the methods.
[0044] FIG. 3A illustrates a flowchart of a method 300 to utilize a
stacked machine learning model to predict a subject's or patient's
predisposition to substance use disorder (in other words, a
patient's risk of developing a substance use disorder), according
to an embodiment. The method 300, at step 302, may include
receiving a sample for allelic analysis (e.g., such as from a
buccal swab, a blood sample and/or any other sample containing
genetic material). The sample is analyzed for one or more SNP
profiles associated to a specified set of allelic variants of the
subject or patient. In an example, the sample is analyzed for one
or more allelic variants (for example, the 15 allelic variants
shown in Table 2 below). Information from a biological material or
a sample 101 may be entered into a computing device 202 via an
input device 218. The input device 218 may be a keyboard,
touchscreen, and/or mouse. In another example, a SNP analyzer 222
may send the SNP profiles to the computing device 202 (for example,
from the analysis of a buccal swab). In yet another example, the
computing device 202 may include the SNP analyzer 222 or other
analyzer to obtain genetic information from a biological
material.
[0045] In such examples, in response to the receipt of a sample, at
step 303, the computing device 202, via the included SNP analyzer
222, may analyze the sample to obtain the one or more SNP profiles
of the allelic variants (or other genetic information) from the
sample. In an embodiment, the SNP analyzer can include any
automated DNA or RNA sequencing system coupled to a software module
to call SNPs in genome or transcriptome samples. In an embodiment,
data from a next generation sequencing system or high density SNP
arrays are aggregated and aligned to call the SNPS of the set of
specified allelic variants.
[0046] The method 300, at step 304, may include, in response to
receipt of the sample, determining whether the computing device 202
is communicatively connected to the trained ensemble or trained
stacked model 100. If the computing device 202 is not connected to
the trained ensemble or trained stacked model 100, the computing
device 202 may, at step 306, send a prompt to a healthcare
professional that an initial risk, prediction, or score of
substance use disorder is not available. After sending the prompt,
the computing device 202 may, at step 308, attempt to re-establish
connection with the trained ensemble or trained stacked model 100.
In another example, the computing device 202 may not send a prompt,
but rather attempt to re-establish connection with the trained
ensemble or trained stacked model 100, which may take a short
period of time (e.g., seconds or minutes). After attempting to
re-establish the connection, the method, at step 310, may restart
the process with the same sample (for example, by resubmitting the
one or more allelic variants for analysis to the computing device
202). In another example and as illustrated in FIG. 2C, the
computing device 202, rather than connecting to the trained stacked
model 100, may include the trained stacked model (e.g., the trained
stacked model 100 may be stored in either the machine-readable
storage medium 206 or another memory of the computing device 202).
Thus, the computing device 202 may not check for a communicative
connection.
[0047] The method 300, at step 312, may include determining, via
the computing device 202, whether the sample includes the one or
more allelic variants used to train the trained ensemble or trained
stacked model 100 (e.g., specified allelic variants). In response
to a determination, at step 312, that not all the one or more
allelic variants are present (for example, due to an indeterminate
allele or no call), the method 300, at step 314, may include
prompting, via the computing device 202 to the user device 220 or
display 216, a healthcare professional that an initial risk,
prediction, or score of a likelihood to develop a substance use
disorder is not available, at which point, the healthcare
professional may decide to input another sample for analysis.
[0048] In response to a determination that the sample includes the
one or more allelic variants, the method 300, at step 316, may
further determine, via the trained stacked model 100, a score or
other indicator that informs a patient's risk of a substance use
disorder. In other words, the SNP profiles of the allelic variants
may be input into the trained stacked model 201 or applied to the
model classifiers, resulting in an output (e.g., 0 to 1 or some
other range of numbers). Further, the score or prediction may be
compared to a threshold. As noted, the threshold may be
pre-determined or input into the computing device 202. If the score
or prediction is less than the threshold, at step 318, a score,
prediction, and/or indicator indicating no or low potential or
likelihood for development of a substance use disorder by a patient
may be returned or transmitted to the user device 220.
Additionally, a recommendation for a treatment regimen may be sent
to a healthcare professional, based on the determination that there
is no or low risk for substance use disorder.
[0049] In response to a determination that the score indicator is
greater than a threshold, at step 320, a score, prediction, and/or
indicator indicating a potential or likelihood for development of a
substance use disorder by a patient may be returned or transmitted
to the user device 220. Additionally, the method 300 may include
identifying one or more associated recommendations related to the
determined score, prediction, and/or indicator, and automatically
transmitting (along with the score or prediction) the determined
initial risk and the one or more associated treatment regimen
recommendations (e.g., a recommendation to utilize an alternate
drug or prescription, such as a non-opioid, or a reduced dose
opioid drug-based treatment regimen) to the user device 220 or
display 216 of a healthcare professional via the electronic
communication network in real time or in a timeframe of hours to
days. Thus, the healthcare professional may have relevant
information on treatment regimens to discuss with the patient, and
the treatment regimens would be based on a potential for or risk of
developing a substance use disorder and/or the treatment regimen
recommendation. The method 300 may further include updating the EMR
of the patient with further clinical data or patient history and
updating the prediction of the initial risk of substance use
disorder to reflect the additional clinical data or patient history
(e.g., updating the EMR at the patient record database 224).
[0050] FIGS. 3B-3E illustrate a flowchart of a method 322 to
utilize a stacked machine learning model to predict a subject's or
patient's predisposition to substance use disorder (in other words,
a patient's risk of developing a substance use disorder), according
to another embodiment. In such embodiments, at step 324, a
healthcare professional may initiate the method 322 prior to an
operation (e.g., pre-operation) to determine a pain management
regimen for the patient post-surgery. In another example, method
322 may be initiated by a healthcare professional at an appointment
or while performing an annual wellness check (e.g., a wellness
check for employment or work-related insurance or pharmaceutical
insurance purposes or at the request of a subject/patient's
insurer). Further, the initiation may be in response to a
determination that the patient requires a prescription of a
particular drug or that the patient or a healthcare professional
has requested the determination of the patient's risk of developing
a substance use disorder. In an embodiment, a healthcare
professional may initiate the method 322 when a patient requests an
elective procedure.
[0051] In response to an initiation of method 322, the computing
device 202 may determine whether any prior scores for a particular
patient have been determined. In other words, the computing device
202 may determine whether a patient's genetic data has been
utilized in the process to determine the risk of developing a
substance use disorder. In such examples, the computing device 202
include an input/output connected to a patient record database 224.
The computing device 202 may determine whether the prior scores are
stored in the patient record database 224 (e.g., in a patients
EMR), in a prescription drug monitoring program (PDMP) database or
computing device, and/or in another data storage location.
[0052] If a patient does not have prior scores available, the
computing device 202, at step 328, requests a biological sample,
genetic data, and/or physical or clinical data of the patient for
analysis or data input at the user device 220. The request may be
transmitted to a healthcare professional. The prompt may be
displayed via the display 216 of a user device 220. In another
example, the sample (e.g., genetic information from a DNA or RNA
from a buccal swab) may be input into a SNP analyzer 222 (for
example, via a biofilm chip) and the data output from the SNP
analyzer 222 may be transmitted to the computing device 202. In yet
another example, the computing device 202 may include the
functionality of the SNP analyzer 222. In another example, the
computing device 202 may prompt or initiate shipment of a home swab
kit to the patient. The computing device 202, at step 330, analyzes
the sample to determine the SNP profiles of the set of specified
allelic variants. The computing device 202 may analyze the
biological sample or other genetic material or pre-process the
physical or clinical data. In response to the analysis of the SNP
profiles of the set of specified allelic variants or receipt of
analytical data from the swab or other genetic material and/or in
response to the pre-processing of the physical or clinical data,
the computing device 202 may initiate the determination of the
score indicating the risk of developing use disorder, as described
in FIGS. 3A and 3B (e.g., at steps 312 through 320 or steps 304
through 320).
[0053] In response to a determination that the score indicating the
risk of developing substance use disorder is less than the
threshold, at step 332, the computing device 202 may generate a
report.
[0054] The report may include the allelic variants of the patient,
the score determined from the trained ensemble or trained stacked
model 201, the level of risk associated with developing substance
use disorder based on the score, a treatment regimen
recommendation, and/or physical or clinical data. In another
example, in response to a determination that the score indicating
the risk of developing use disorder is more than the threshold, at
step 334, the computing device 202 may generate a report as well,
with similar information as described above for step 332.
[0055] In response to the generation of the report indicating a
high risk of developing substance use disorder, the computing
device 202 may transmit the report to the PDMP as in step 336. In
such examples, a patient's physical or clinical data, along with
the patient's predisposition to developing a substance use
disorder, may be uploaded from the report to a state or a federal
PDMP. Further, the current prescription may be uploaded to the
state or a federal PDMP. Thus, doctors, nurses, or other users who
may not utilize methods 300, 322, and 342, may determine whether a
patient is more or less likely to develop a substance abuse
disorder based on the data in the PDMP.
[0056] At step 338, the computing device 202 may transmit a
treatment regimen recommendation based on the report or based on
the score indicative of the level of risk associated with
developing substance use disorder. In an embodiment, the report may
include the original prescription and the score or indicator for a
patient's likelihood or predisposition to develop a substance use
disorder. Based on the returned score indicating a high risk of
developing a substance use disorder, the computing device 202 may
provide suggestions or recommendations for a treatment regimen or
possible alterations to the prescription to a drug that may reduce
the patient's likelihood to develop the substance use disorder
(e.g., suggesting an alteration to a prescription from an opioid
based drug to a non-opioid based drug based on the determined
likelihood a patient may develop a SUD).
[0057] At step 340, the computing device 202 may transmit the
report and/or treatment regimen recommendation to patient data
storage or a patient's EMR (e.g., patient record database 224). In
another example, a field associated with likelihood to develop
substance abuse disorder may be altered or updated, based on the
results described above. For example, the computing device 202 may
include or add a note or indication of the high risk of developing
a substance use disorder in the patient's EMR (e.g., stored medical
records at a patient record database 224 or other data storage
locations).
[0058] In FIGS. 3D through 3E, the method 342 may be initiated, at
step 344, in addition to what is described above for method 322, by
a pharmacy benefit manager (PBM) or by an insurer (e.g.,
prescription drug insurance provider). In such examples, prior to
funding or paying for a portion of or all of a certain prescription
drug, a PBM or insurer may request that a patient provide a report
or score indicating the risk of the patient developing a substance
use disorder (as described above). As such a determination may take
time, in some examples, the computing device 202, PBM, or insurer
may provide a 3-day maximum prescription of the drug prescribed, at
step 346. At step 330, in response to receipt of the sample from
the patient, the computing device 202 may perform an analysis to
obtain one or more SNP profiles of the allelic variants or other
patient related information and then initiate the determination of
the score indicating the risk of developing use disorder, as
described in FIG. 3A (e.g., at steps 312 through 318 or steps 302
through 318). Further, once the score is determined, the computing
device 202 may generate reports, transmit the report or update the
PDMP as in step 336, recommend possible alterations to a
prescription, generate a treatment regimen recommendation, and/or
transmit or store the report and prescription data to the patient
data storage (e.g., patient record database 224) as described above
for FIGS. 3B through 3C (e.g., at steps 332 through 340).
[0059] As used herein, an electronic communications network may be
any type of network configured to provide communications between
components of system 200--the user device 220 or user interface,
the SNP analyzer 222, the patient record database 224, and the
processor 204 that executes the trained ensemble or trained stacked
model 100. For example, the electronic communications network may
include a network infrastructure that facilitates the exchange of
information and provides communication, such as the Internet, a
Local Area network (LAN), wired network, cable, Wireless Local Area
Network (WLAN), cellular, satellite, or other suitable connections
that enable the exchange of information between the components of
systems, as will be readily understood by one having ordinary skill
in the art. In some other embodiments, however, the components of
the system 200 may be connected through a dedicated direct
communication link.
[0060] As used herein, a "machine readable storage medium" may be
any electronic, magnetic, optical, or other physical storage
apparatus to contain or store information such as executable
instructions, data, and the like. For example, any machine-readable
storage medium described herein may be any of random access memory
(RAM), volatile memory, non-volatile memory, flash memory, a
storage drive (e.g., hard drive), a solid state drive, any type of
storage disc, and the like, or a combination thereof. As noted, the
machine-readable storage medium 206 may store or include
instructions executable by the processor 204.
[0061] The processor 204 or processing resource that executes the
trained ensemble or trained stacked model 100 can include a server,
storage services, cloud provided by an entity such Microsoft
Azure.TM. or Amazon Web Services.RTM. cloud, among others, as will
be readily understood by one having ordinary skill in the art. The
processor or processing resource may also include a plurality of
computing devices in communication with the electronic
communications network. For example, in some embodiments, the
processor may be a plurality of processors connected together in
communication with the electronic communications network. In other
embodiments, the processor may be a group of graphical processing
units configured to work in parallel as a GPU cluster. A processor
may include a single processor device and/or a plurality of
processor devices (e.g., distributed processors). A processor 204
may be any suitable processor capable of executing/performing
instructions. A processor 204 may include a central processing unit
(CPU) that carries out program instructions to perform the basic
arithmetical, logical, and input/output operations required to
execute the method of predicting risk of substance use disorders or
for providing decision support to healthcare professionals to
implement a treatment regimen. A processor 204 may include code
(e.g., processor firmware, a protocol stack, a database management
system, an operating system, or a combination thereof) that creates
an execution environment for program instructions. Processes and
logic flows described herein may be performed by one or more
programmable processors executing one or more computer programs to
perform functions by operating on input data and generating
corresponding output. The processor is communicatively connected to
one or more patient record databases through the electronic
communication network. Patient record databases 224 may include one
or more memory devices that store information and are accessed and
managed through or by the processor 204. For example, databases may
include Oracle.TM. database or other relational databases or
non-relational databases such as Hadoop. Databases can include
computing components such as database server configured to receive
and process requests for data stored in memory devices of databases
and to provide data from the databases. The patient record
databases 224 may be located remotely, such as in a different
geographical location or the cloud. The patient record databases
224 may include EMR of the patients, training data, and/or initial
and updated predictive models. The patient record databases 224 may
include other clinical data such as age, sex, race, and ethnicity.
In an embodiment, the clinical data includes demographic data,
socioeconomic data, and any data that about a subject that can be
obtained by observation or oral or written communication.
[0062] Data encryption may be used to comply with business or
regulatory rules, such as Health Insurance Portability and
Accountability Act (HIPAA) privacy rules or the like. Such data
encryption may ensure a secured communication of the determined
initial diagnosis and the one or more associated recommendations
between the user device and the third parties, thereby preventing
the privacy of the communicated data to be compromised.
[0063] Disclosed herein are embodiments of methods of evaluating
various substance use disorder risks of a subject, such as opioid
use disorder (OUD), alcohol use disorder (AUD), cocaine use
disorder (CUD), cannabinoid use disorder (CAUD), and/or nicotine
use disorder (NUD). An embodiment includes a multi-variant genetic
test of evaluating an OUD, AUD, CUD, NUD, and/or CAUD risk of a
subject or patient. Opioids used by a subject can be one or more of
any chemicals that interact with opioid receptors on nerve cells in
the body, including but not limited to illegal drugs such as
heroin, synthetic opioids such as fentanyl, and prescription pain
relievers, such as oxycodone (OxyContin.RTM.), hydrocodone
(Vicodin.RTM.), codeine, morphine, and other. Alcohol used by a
subject can be one or more of any chemicals that interact with
alcohol receptors on nerve cells in the body, including but not
limited to legal drugs such as wine, beer, malt liquor, and liquor.
Cocaine used by a subject can be one or more of any chemicals that
interact with various receptors on nerve cells in the body,
including but not limited to illegal drugs such as cocaine and
cocaine derivatives. Cannabis used by a subject can be one or more
of any chemicals that interact with various receptors on nerve
cells in the body, including but not limited to illegal and legal
drugs such as cannabis or other substances including
tetrahydrocannabinol. Nicotine used by a subject may be one or more
of any chemicals that interact with various receptors on nerve
cells in the body, including but not limited to legal drugs such as
tobacco (included in products, such as cigarettes, cigars, chewing
tobacco, or snuff), liquid nicotine (included in products, such as
electronic cigarettes), or other forms of nicotine (for example,
nicotine patches, lozenges, gum, or other dissolvable
products).
[0064] To address the role of genetic variability as a screening
method for OUD risk, a novel multi-variant genetic test was
previously evaluated in a small case cohort study that included 37
patients with a known addiction to prescription opioids or heroin,
30 age- and gender-matched subjects with no known addiction to
prescription opioids or heroin, and 138 additional subjects
assessed for generalizability. Using a panel of single nucleotide
polymorphisms (SNPs) that have been associated with brain pathways
that regulate reward together with a computer learning model, a
predictive algorithm was developed to estimate OUD risk. The model
yielded results suggesting that the algorithm could be used with
97% certainty to identify individuals with low likelihood of
developing opioid addiction, and 88% certainty to identify
individuals with high likelihood of developing opioid
addiction.
[0065] A subsequent study was conducted using genetic samples from
a large cohort of subjects previously studied with respect to OUD
that were provided by the National Institute on Drug Abuse (NIDA)
Genetics Consortium (NGS), both to substantiate the earlier
findings and develop an ensemble model. In this case study, the
National Institute on Drug Abuse provided genetic samples from a
large cohort of subjects previously studied for opioid use disorder
(OUD). This case study utilized a larger data set and advanced
machine learning to build a model that more accurately predicts OUD
risk based on genotype alone (i.e., expression of alleles
associated with reward, self-control, and affect) or in combination
with clinical data (i.e., sex, age, and other substance abuse or
dependence). Genetic data was modeled with and without clinical
data using a machine learning platform that utilized a random
forest model 105, a gradient boosted tree model 107, an elastic net
model, a SVM model 109, or some combination thereof. Training and
20-fold cross-validation were performed with 80% 116 of randomly
selected samples of the training data 102. Predictive accuracy of
the model was assessed in a holdout analysis of 20% 118 of the
training data 102. A microarray of 15 single nucleotide
polymorphisms (SNPs) associated with risk for opioid addiction was
developed.
[0066] The machine learning model (or stacked model 210) was
trained using data from 77 patients with prescription opioid or
heroin addiction and 76 age- and gender-matched subjects with no
known addiction to prescription opioids or heroin who had
participated in the previous study. The resulting model was then
tested on genomic DNA samples from human subjects who had
participated in two genome-wide association studies (GWAS) (the
Lachman study,
nidagenetics.org/studies/study-14-genome-wide-analysis-addiction-suscepti-
bility-genes, and the Nelson study, nidagenetics.
org/studies/study-18-opioid-dependence-candidate-genes-and-g-x-e-effects)-
. The NIDA Center for Genetic Studies provided the DNA samples and
the subjects' baseline demographic and clinical data, including
information regarding substance abuse or dependence on non-opioid
substances. Subjects in the Lachman study were recruited from
methadone maintenance treatment programs in the New York City
metropolitan area and met DSM-IV criteria for opioid dependence, as
confirmed by the Structured Clinical Interview for DSM-IV.
Subjects, with no known addiction to prescription opioids or
heroin, in the Nelson study were recruited from geographic areas in
proximity to clinics providing opioid replacement therapy in the
greater Sydney, Australia, region, but had themselves used opioids
recreationally fewer than 10 times. Participants in both studies
provided written informed consent (which allowed for subsequent use
of genetic samples), and all procedures in this study were
performed in accordance with the informed consent.
[0067] Genomic DNA samples were genotyped by AutoGenomics, Inc.
(Carlsbad, Calif., USA) using their INIFINITI.RTM. PLUS microarray
platform and Neural Response Panel of 15 SNPs associated with risk
for opioid addiction. Sanger sequencing confirmed >99.9%
concordance with SNP genotyping. The potential importance of these
SNPs on predicting opioid addiction was analyzed using a
probabilistic model, with normalization based on the SNP with the
highest relative predictive importance.
[0068] Data from the two NIDA studies were modeled using a
proprietary platform comprising a random forest model 105, a
gradient boosted tree model 107, and elastic net classifier (for
example, a SVM model 109). Each trained model produced a
classifier. When information from a biological sample 101 is
applied to each classifier, a prediction score or other indicator
is the output (for example, a value between 0 and 1). The
predictions from each classifier were combined by ensemble learning
techniques, where multiple sets of predictions from the classifiers
have the same dependent variable and the same or similar
independent variables. In another example, the predictions or
output from the classifiers may be averaged 112 to produce a
prediction score 114.
[0069] All data processing was conducted under fully blinded
conditions, with One-Hot Encoding or Ordinal Encoding used to
transform continuous variables into discrete categories. The only
major assumption was that model training data are representative of
the future scoring data.
[0070] Various combinations of the following data were considered
and preliminarily tested: genetics, age, sex, and abuse or
dependence on other substances (i.e., nicotine, alcohol,
cannabinoids). Based on preliminary findings, the ensemble (blended
or stacked) model was developed and deployed using genetic data
only and genetic plus clinical data. As OUD cases are rare relative
to the overall general population, an under-sampling method was
used in which the number of samples in the rare class was matched
to an approximately similar number of samples from subjects with no
known addictions. Twenty-fold cross-validation was performed with
80% 116 of randomly selected samples used for training/learning and
20% 118 of the samples were used for holdout analysis.
[0071] Confusion matrix outputs were as follows: (i) F1 score,
which measures accuracy based on precision and recall
(F1=2*precision*recall/(precision+recall), in which precision is
the number of correct positive results divided by the number of all
positive results returned by the model, and recall is the number of
correct positive results divided by the number of samples that
should have been identified as positive; (ii) true positive rate
(sensitivity); (iii) false positive rate; (iv) true negative rate
(specificity); (v) false negative rate; (vi) positive predictive
value (percentage of all positive predictions that were correct);
(vii) negative predictive value (percentage of all negative
predictions that were correct); (viii) accuracy (percentage of all
correctly classified instances), and (v) Matthews Correlation
Coefficient, which measures model quality when the data set is
unbalanced. Additional outputs included: prediction distribution
graphs, which illustrate how well the model discriminates between
prediction classes; lift charts, which depict how well a model
segments the target population and predicts the target outcome; and
receiver operating characteristic (ROC) curves, which plot true
positives against false positives with greater area under the curve
(AUC) indicating better accuracy.
[0072] Genomic samples and clinical data from 458 cases and 508
cases with no known addiction to prescription opioids or heroin
were analyzed. GWAS included genomic samples from 1181 subjects,
663 cases from the Lachman study and 518 cases with no known
addiction to prescription opioids or heroin from the Nelson study.
To identify likely cases of OUD per DSM-5 criteria, 215 subjects
were excluded from analyses based on the following criteria: cases
that were identified as positive for both opioid dependence and
opioid abuse (n=136); cases with no known addiction to prescription
opioids or heroin and 1 case that were positive for opioid abuse
only (n=13); or insufficient information (n=66).
[0073] In the remaining analysis set (N=966), 458 subjects from the
Lachman study were classified as positive for OUD per DSM-5
criteria and 508 (2 from the Lachman study and 506 from the Nelson
study) were classified as negative for OUD. Baseline demographic
and clinical data for the analysis set are shown in Table 1.
TABLE-US-00001 TABLE 1 Baseline Demographics and Clinical Data Case
Subjects with No Subjects Known Addiction Variable (N = 458) (N =
508) Mean age (SD) - year 42.0 (7.6) 35.0 (10.6) Female sex - no.
(%) 201 (43.9) 283 (55.7) Race or ethnic group - no. (%) White
non-Hispanic 47 (10.3) 360 (70.9) Black non-Hispanic 110 (24.0) 1
(0.2) Hispanic 291 (63.5) 1 (0.2) Other 10 (2.2) 122 (24.0) Unknown
24 (4.7) Substance abuse or dependence - no. (%)* Alcohol abuse 45
(9.8) 106 (20.9) Alcohol dependence 194 (42.4) 134 (26.4) Nicotine
dependence NA 204 (40.2) Cannabis abuse 38 (8.3) 65 (12.8) Cannabis
dependence 249 (54.4) 135 (26.6) *Per DSM-IV criteria DSM-IV,
Diagnostic and Statistical Manual of Mental Disorders, 4th Edition;
NA, not available
[0074] Among the 15 SNPs included in the Neural Response Panel, a
genetic polymorphism in the serotonin receptor gene (rs7997012) had
the highest probability of being individually associated with
opioid addiction and was therefore assigned a normalized impact
score of 1.0 (Table 2). Other SNPs that had relatively high
individual probabilities of being associated with opioid addiction
were for polymorphisms located in genes encoding galanin (rs948854)
and the ATP-binding cassette (ABC) transporter (rs1045642), with
normalized impact scores of 0.7562 and 0.6910, respectively.
Deployment of the previously developed algorithm based on only
genetics to the 966 subjects included in the NIDA analysis set
resulted in a prediction that was 91.5% sensitive (95% CI:
88.5-93.9%), 45.1% specific (95% CI: 40.7-49.5%), and 67.1%
accurate (95% CI: 64.0-70.0%), with AUC=0.7000.
TABLE-US-00002 TABLE 2 Normalization of Genetic Variants Impact SNP
ID Gene Gene Description Unnormalized Normalized rs7997012 5-HTR2A
Serotonin 2A receptor 0.0797 1.0 rs948854 GAL Galanin 0.0603 0.7562
rs1045642 ABCB1 ATP binding cassette transporter 1 0.0551 0.6910
rs4680 COMT Catechol-O-methyltransferase 0.0446 0.5591 rs6347 DAT1
Dopamine transporter 0.0432 0.5423 rs1800497 DRD2 Dopamine D2
receptor 0.0331 0.4157 rs4532 DRD1 Dopamine D1 receptor 0.0276
0.3455 rs1801133 MTHFR Methylene tetrahydrofolate reductase 0.0232
0.2909 rs1611115 DBH Dopamine beta hydroxylase 0.0210 0.2628
rs2236861 DOR Delta opioid receptor 0.0197 0.2473 rs1799971 OPRM1
Mu opioid receptor 0.0113 0.1419 rs3758653 DRD4 Dopamine D4
receptor 0.0111 0.1397 rs211014 GABA Gamma-aminobutyric acid 0.0100
0.1250 rs1051660 OPRK1 Kappa opioid receptor 0.0079 0.0990
rs9479757 MUOR Mu opioid receptor 0.0029 0.0367
[0075] Deployment of the trained ensemble or trained stacked model
100 developed using the analysis set of 966 cases and cases with no
known addiction to prescription opioids or heroin including only
genetic data yielded a prediction that was 73.6% sensitive (95% CI:
63.4-82.3%), 75.5% specific (95% CI: 66.0-83.5%), and 74.6%
accurate (95% CI: 67.9-80.6%), with AUC=0.7840 (Table 3; FIG. 4).
FIG. 4 is a representation 400 of the receiver operating
characteristic curves using (i) genetic data only 402, (ii) genetic
and clinical data 404, and (iii) genetic and a subset of clinical
data 406, according to an embodiment of the present disclosure.
FIGS. 5A and 5B are graphical representations 500, 501 of the
prediction distributions based on genetic data only and genetic and
clinical data, respectively. The prediction distribution graph
indicated moderate discrimination between cases and cases with no
known addiction to prescription opioids or heroin (FIG. 5A).
TABLE-US-00003 TABLE 3 Performance of Ensemble Model on Holdout
Analysis Genetic Data Genetic + Subset Analysis Metric Only
Clinical.dagger. Data Genetic + Clinical.dagger-dbl. Data F1 score
0.7322 0.8962 0.8690 True positive (sensitivity) 0.7363 0.9011
0.9130 False positive 0.2451 0.0980 0.1287 True negative
(specificity) 0.7549 0.9020 0.8713 Positive predictive value 0.7283
0.8913 0.8289 Negative predictive value 0.7624 0.9109 0.9362
Accuracy 0.7461 0.9016 0.8882 Matthews correlation coefficient
0.4909 0.8026 0.7747 AUC 0.7840 0.9593 0.9354 .dagger.Clinical data
includes age, sex, nicotine and/or alcohol abuse or dependence,
cannabinoid abuse or dependence, and no illicit drug use.
.dagger-dbl.Subset analysis excludes African-American subjects.
Clinical data includes age, sex, nicotine and/or alcohol abuse or
dependence, cannabinoid abuse or dependence, and no illicit drug
use. AUC, area under the receiver operating characteristic
curve
[0076] Models that included both genetic and clinical data showed
increased predictive accuracy for OUD risk (FIG. 5B). The
prediction from the model that included age, sex, and abuse or
dependence on other substances was 87.9% sensitive (95% CI:
79.4-93.8%), 73.5% specific (95% CI: 63.9-81.8%), and 80.3%
accurate (95% CI: 74.0-85.7%), while including the additional
clinical variable of "no illicit drug use" to the model resulted in
prediction that was 90.1% sensitive (95% CI: 82.0-95.4%), 90.2%
specific (95% CI: 82.7-95.2%), and 90.2% accurate (95% CI:
85.0-94.0%), with AUC=0.9593 (Table 3; FIG. 4). The prediction
distribution graph indicated considerable discrimination between
cases and cases with no known addiction to prescription opioids or
heroin (FIG. 5B). Lift charts provide a measure of the
effectiveness of the ensemble predictive model and are calculated
as the ratio between the results obtained with and without the
ensemble predictive model. Lift charts are visual representations
of model performance. Lift charts showed a close fit between
predicted and actual values, and the area under the ROC curves
indicated that a high proportion of the predictions could be
classified as true (FIGS. 4 and 6A-6B; Table 3). FIGS. 6A and 6B
are lift charts of the ensemble predictive model developed by using
genetic data only 600 and a combination of genetic and clinical
data 601, respectively, according to an embodiment of the present
disclosure. Deployment of the model using clinical data alone
demonstrated approximately 50% predictive accuracy.
[0077] Upon observing the mismatch in age between the case and
cases with no known addiction to prescription opioids or heroin
datasets, the model was deployed using the genetic and clinical
variables without including age. In this analysis, predictive
accuracy was slightly diminished with 76.9% sensitivity (95% CI:
66.9-85.1%), 76.5% specificity (95% CI: 67.0-84.3%), and 76.7%
accuracy (95% CI: 70.1-82.5%). Similarly, to account for potential
differences in allele frequencies among different racial/ethnic
groups, the model was deployed using the genetic and clinical
variables on a subset of cases (n=348) and cases with no known
opioid (n=507) that did not include African-American subjects (the
only racial/ethnic group in which minor allele frequencies reported
in dbSNP for a number of the SNPs included in the Neural Response
Panel differ significantly from other racial ethnic groups). This
analysis resulted in prediction that was 91.3% sensitive (95% CI:
82.0-96.7%), 87.1% specific (95% CI: 79.0-93.0%), and 88.8%
accurate (95% CI: 83.1-93.1%), with AUC=0.9394 (Table 3, FIG. 4).
These findings were similar to the results observed in the full
case cohort.
[0078] Results of the current study demonstrate that risk for
opioid addiction can be predicted by a panel of SNPs associated
with the brain's reward pathways, and that prediction sensitivity,
specificity, and accuracy can be improved by including clinical
data such as age, sex, and abuse or dependence on other substances
in the predictive algorithm. The study demonstrated that use of
ensemble machine learning to search for the best set of models
(blended or stacked), based on both the characteristics of the data
and the prediction target increases the predictive accuracy of the
derived algorithm. This study confirmed the findings of an earlier
study in a much larger case cohort.
[0079] This affordable, non-invasive or minimally invasive tool for
risk assessment for opioid addiction or OUD provides decision
support for implementing treatment regimens. For individuals
prospectively identified as genetically at-risk for OUD, a
treatment regimen may avoid or reduce opioid exposure, even in the
context of perioperative pain management. The treatment regimen for
such an individual would be involve non-opioid modalities, such as
regional anesthesia and analgesic techniques, and multimodal
analgesia. Additionally, knowledge regarding the genetic
predisposition for OUD informs treatment paradigms for at-risk
individuals who have been exposed to opioids, as such individuals
may be candidates for buprenorphine-assisted dose reduction or
behavioral interventions. A highly reliable risk assessment tool
can provide improved access to opioid pain medications for
individuals with legitimate medical needs, who are at low risk for
addiction, but who currently are being denied adequate pain
management, or are having difficulty accessing primary care.
[0080] While a genetic component to the pathogenesis of opioid
addiction is generally well accepted, GWAS to identify candidate
genes and their associated SNPs have often not been replicated or
have shown inconsistent results. The SNPs selected for use in the
polygenetic panel described in this study were based on published
literature, and it may be the case that individual SNPs that fail
to achieve a statistically significant association with opioid
addiction in GWAS may not be causal but closely linked to addiction
risk. There remains the difficulty of tying the SNP panel used in
this study to the underlying neurobiology of opioid addiction,
however, the SNP panel is clearly a crucial factor in the model.
While certain clinical characteristics (e.g., older age, female
sex) appear to be associated with longer term opioid use, clinical
characteristics alone were not predictive in this model; without
the genetic component, the predictive accuracy of the model
declined to approximately 50%.
[0081] In the Lachman study approximately two-thirds of subjects
were Hispanic and one-quarter were African-American, while 70% of
subjects in the Nelson study were White. While frequencies of
alleles are slightly different between racial and ethnic groups,
the current study cannot determine whether these dissimilarities
are related to ethnic differences or are disease-related. However,
the subset analysis that excluded African-American subjects yet
showed the model retained a high level of predictive accuracy
suggests that racial/ethnic allelic frequency differences do not
affect the model described here. Nonetheless, additional studies
that control for ethnicity are in progress. Additionally, it was
observed that including age as a variable in model development
appeared to increase its predictive accuracy, however it is not
clear whether age is a true signal of risk or whether it merely
reflects increased likelihood of opioid exposure.
[0082] Results from this study demonstrate the generalizability of
a highly sensitive and specific predictive assay for risk of opioid
addiction, which can prevent opioid exposure in at-risk
individuals, while at the same time increasing access to
legitimately needed opioid medications among individuals suffering
from severe and debilitating acute and chronic pain who are at low
risk for addiction. Both patients and clinicians can make better
informed treatment decisions with knowledge of potential risks for
addiction. Risk for opioid addiction can be predicted by a panel of
SNPs associated with the brain's reward pathways, and predictive
accuracy increases with the inclusion of clinical data. Methods and
systems for predicting the risk of opioid addiction inform
treatment decisions for the patient.
[0083] While many embodiments of the present invention have been
shown and described herein, it will be obvious to those skilled in
the art that such embodiments are provided by way of example only.
Numerous variations, changes, and substitutions will now occur to
those skilled in the art without departing from the invention. It
should be understood that various alternatives to the embodiments
of the invention described herein may be employed in practicing the
invention. It is intended that the following claims define the
scope of the invention and that methods and structures within the
scope of these claims and their equivalents be covered thereby.
* * * * *