This disclosure relates to a practical, computer-implemented testing method and related system for determining, in advance, whether a particular patient would likely benefit from having a pharmacogenomics (PGx) test performed.
Pharmacogenomics is the study of how genetic variations in a person's genes affect their response to drugs. This relatively new field combines pharmacology (the science of drugs) and genomics (the study of genes and their functions) with the objective to develop and prescribe effective, safe medications and doses that will be tailored to a person's genetic makeup.
It is common knowledge that many drugs that are currently available are “one size fits all,” but in practice they don't work the same way for everyone. It can be difficult to predict who will benefit from a medication, who will not respond at all, and who will experience negative side effects.
While conducting PGx testing would inevitably benefit virtually all patients, the associated cost and throughput of PGx testing makes it infeasible to test every patient, for example all the patients currently under the care of a healthcare provider such as a hospital. Instead, a pragmatic approach would be to perform what is essentially a “pre-testing” method on patients in order to determine which patients are deemed to be most likely to benefit from the information provided by PGx testing. The system and method of this disclosure enables this approach to be realized.
In particular, while 90% of people have one or more variants in a PGx related gene, not all will actually be prescribed a medication that is PGx-relevant (i.e., those medications that have been firmly identified to have a link to a particular genetic variant). In other words, while a PGx test may identify that a patient has certain genetic variations that may predict they would benefit from certain drugs linked to such genetic variations (or have no benefit or could be harmed if they took other drugs linked to their genetic variations), in retrospect, there would little or no point in conducting the PGx test for that patient if their particular health or disease status indicates that these drugs are completely irrelevant to their particular health situation or disease state.
It is with this goal in mind that we have developed the predictive testing method and system of this disclosure, which we have called “MAPPeR” (Meaningful, Actionable
Pharmacogenomic Patient Results). MAPPeR provides a framework to determine the likelihood that a given patient will be prescribed a PGx-relevant medication, and therefore, would likely obtain benefit from a subsequent PGx test. This, in turn, results in more informed decision making for the clinical staff treating the patient and potentially providing healthcare cost savings. The method of this disclosure can also be performed widely on a set of (or all) patients at a particular healthcare provider, e.g., hospital, VA administration or subdivision thereof, and thereby enable prioritization of those patients for which for PGx testing is likely to result in substantial benefit.
In one aspect of this disclosure, a computer-implemented method for predicting whether a patient will likely benefit from a pharmacogenomics test is disclosed. The term “likely benefit from a pharmacogenomics test” means that the patient is likely to be prescribed a medication that is PGx-relevant. The method includes step (a) obtaining an input data in the form of data associated with disease status of the patient (i.e., a set of one or more ICD-10 codes, the common name of the diseases, or the equivalent), and/or data associated with a list of currently prescribed medications for the patient. This input data set could also take the form of a number, name of the patient or other information that can be linked with the patient's electronic medical record (EMR) in which case the data representing disease status and/or medications can be extracted from the EMR.
The method includes a step (b) of implementing in the computer a Bayesian network representable as a tripartite graph having links between three partitions: (1) a disease status partition having as elements representing one or more independent diseases of the patient; (2) a medications partition have as elements medications associated with the elements of the disease status partition or the medications prescribed for the patient; and (3) a genetics partition having as elements particular genetic variations which have an established pharmacogenomics relationship with one or more elements in the medications partition. The nomenclature assigned to the partitions is not particularly important, for example the genetics partition could be characterized as “biomarker”, “actionable alleles”, or the like.
The weights of links between the disease status partition and the medications partition in the Bayesian network are based on an analysis of a corpus of patient data including prescribed medications and disease diagnosis, and will have numerical values of between 0 and 1, where the higher the number the more probable the medication is prescribed for the particular disease. The weights of links between the medications partition and the genetics partition are a binary value of 1 or 0 depending on whether a pharmacogenomics relationship has been established between the elements of the medications partition and the elements of the genetics partition.
The method further includes a step (c) of generating (i.e., calculating) from the Bayesian network a probability of the patient being prescribed a medication having a pharmacogenomics relationship with one of the genetic variations in the genetics partition. This probability, referred to as P(M) in the following discussion, is based on the input data and the weights and links between the elements of the three partitions of the Bayesian network for the input data. Equation (1) below is one example of a procedure for generating the probability. A prediction of whether a patient will likely benefit from a pharmacogenomics test can be made based on the probability P(M) generated in step c). For example, a threshold may be determined based on analysis of the receiver operating characteristic curve (ROC) for the network to determine the optimum predictive performance (sensitivity and specificity). In one embodiment, the threshold is 0.01. If the probability is above the threshold the patient is recommended to have PGx testing performed.
In another aspect, an improved computer is disclosed which is configured to facilitate recommendations for conducting pharmacogenetics testing on a patient. The computer includes a) a memory storing an input data set in the form of data reflecting disease status of the patient, and/or data reflecting a list of currently prescribed medications for the patient; and (b) a processing system configured to implement a Bayesian network representable as a tripartite graph having links between three partitions: (1) a disease status partition having as elements representing one or more independent diseases of the patient; (2) a medications partition have as elements medications associated with the elements of the disease status partition or the medications prescribed for the patient; and (3) a genetics partition having as elements particular genetic variations which have a pharmacogenomics relationship with the elements in the medications partition. The weights of links between the disease status partition and the medications partition are based on an analysis of a corpus of patient data including prescribed medications and disease diagnosis, and the weights of links between the medications partition and the genetics partition are a binary value of 1 or 0 depending on whether a pharmacogenomics relationship has been established between the elements of the medications partition and the elements of the genetics partition. The computer further includes (c) executable instructions for the processing unit generating from the Bayesian network a probability of the patient being prescribed a medication having a pharmacogenomics relationship with one of the genetic variations in the genetics partition, P(M), based on the input data and the weights and links between the elements of the three partitions for the input data. A prediction of whether a patient will likely benefit from a pharmacogenomics test can be made based on the probability P(M) generated by the executable instructions c).
In another aspect, a method of selectively conducting pharmacogenomics testing on a multitude of patients of a healthcare provider (e.g., hospital, clinic, VA, etc.) includes steps of: a) entering input data for each of the multitude of patients (current diagnoses, and/or medications); b) conducting the method for predicting patient benefit of pharmacogenomics testing as explained above for each of the multitude of patients based on the input data; c) using the predictions P (M) for each of the patients to prioritize which patients should be subject to pharmacogenomics testing, and d) conducting pharmacogenomics testing for the patients prioritized in step c). Steps a), b) and c) of the method can be performed periodically, such as daily, for most or potentially all patients of the healthcare provider. The input data can optionally take the form of a patient medical record number, patient identifying information such as name, or number associated with the patient. Such information can then be linked to disease and medication data in an EMR for the patient and the input data extracted from the record and input into the Bayesian network. In one embodiment the health care provider is a hospital or medical clinic. In one possible embodiment the healthcare provider is the U.S. Veterans Administration (VA) or subdivision thereof, e.g., particular VA hospital or clinic.
As explained above, while pre-emptive PGx testing for all patients in a particular healthcare system could bring precision medicine at a population health level, the scarcity of resources and the need for economic efficiency calls for a tool that could stratify patients by the potential utility of this type of genetic testing. The innovation represented by this disclosure enables this ability to stratify patients for potential utility of PGx testing. It has the potential to accelerate the adoption of pharmacogenomics by increasing the yield of actionable results, which in turn could increase the engagement of clinicians and patients to this emerging field and bring the benefits of PGx testing to a much wider number of patients. Further, no blood test, genomic test, or other physically invasive test need be done to conduct the method of this disclosure, rather all is needed is information from the patient's electronic medical record.
As will be explained below, the primary end user of the system and method of this disclosure (MAPPeR for shorthand) would be a healthcare provider who could use it to identify which patients have the most immediate potential to benefit from PGx testing. Specific end users could be primary care providers at hospitals and clinics and physicians participating PHASeR programs in VA medical centers and the VA in general. PHASeR (PHarmacogenomic Action for cancer SuRvivorship) is a program funded by philanthropist Denny Sanford to provide free genetic testing for up to 250,000 veterans through a partnership between Sanford Health, the assignee of this invention, and the VA. Of course, the MAPPeR tools and methods are generally applicable to other medical systems and hospitals.
Healthcare providers could essentially match their patients from their clinic schedule each day to see whom they should recommend having PGx testing, by preforming the methods of this disclosure on a regular, e.g., daily basis, as will be described in conjunction with
Another potential user of a version of MAPPeR could be patients themselves. An API could be set up with already available tools like the CMS Blue Button to allow patients to enter their own information via a suitable user interface (e.g. a web browser); the information entered prompts extraction of the input data needed for the Bayesian network and a prediction is generated by the Bayesian network and reported back to the patient in substantial real time. This would allow the patient to see if PGx testing would have potential benefit for them. This would empower patients to have conversations with their healthcare providers and request PGx testing.
Referring now to
This input data set 100 is provided to the computer 120 and in particular a Bayesian network 124 is constructed and implemented in the computer. The network 123 generates an output 126 in the form of a prediction or recommendation for PGx testing as will be explained below. This Bayesian network 124 (see
(1) a disease status partition 202 having as elements 203, 205, 207 etc. representing one or more independent diseases of the patient, there being up to N such diseases, where N is some integer greater than or equal to 1;
(2) a medications partition 204 have as elements 211, 213, 215 etc. medications that are associated with one or more of the elements 203, 205, 207 etc. of the disease status partition 202, or are medications which have been prescribed for the patient; and
(3) a genetics partition 206 (which could also be called “biomarker”, “actionable alleles” etc.) having as elements 232, 234, 236 etc. particular genes with alleles 232 (variants of the gene) which may have an established pharmacogenomics relationship with one or more the elements in the medications partition 204. It will be appreciated that as pharmacogenomics research progresses the content of the genetics partition 206 may expand over time to reflect new discoveries between genetic variants and response to pharmacological products.
The weights of links 210, 212, 214, 216 etc. between the disease status partition 202 and the medications partition 204 are based on an analysis of a large corpus of patient data which includes prescribed medications and disease diagnosis. This will be discussed in more detail below. In
Our method then continues with step (c) of generating (i.e., calculating) from the Bayesian network of
In our approach shown in
Our method also provides a directed link from disease status to medication prescription to PGx biomarkers, as indicated by the three partitions of
The Bayesian network 124 that lies at the core of the MAPPeR program can be broken down into a tripartite graph as shown in
and
where P(Mi|Di) is the weight of each link between the disease status and medications partitions.
The variable Gi,j, just represents the existing of a pharmacogenomics link between Gene k and Medication i. If there is a link, this value is 1; if not, it is 0. The probability will be 0 only in the event that the particular medication is not linked to any significant PGx biomarker/gene.
An assumption is built into Equation 1 that the events of being prescribed the medication given a single disease are independent (i.e. P ((M|Di)∩(M|Dj))=P(M|Di)*P(M|Dj)).
The links between medication and PGx biomarkers (220, 222) are binary representations of a direct connection determined by a combination of the guidelines established by the Clinical Pharmacogenetic Implementation Consortium (CPIC), see https://cpicpgx.org/guidelines/, and the FDAs Table of Pharmacogenomic Biomarkers in Drug Labeling. See gttps://www.fda.gov/drugs/science-and-research-drugs/table-pharmacogenomic-biomarkers-drug-labeling.
The probabilistic disease-to-medication mapping (represented by the links 210, 212, 214, etc. and the binary linkage from medications to PGx biomarkers (220, 222) are combined to form the core Bayesian network of the MAPPeR test method, using Equation 1 to calculate the probability P(M).
The indication of PGx relevance comes from having an above-threshold probability for any of the specific medications that have PGx relevance. It is possible to generate a single value to represent their probability of benefiting from PGx testing, especially Equation (1) integrates more information such as the likelihood of allele presence. Alternatively, the value could be represented by the calculation for the probability P of medication prescription: Equation (2) P=1−(πi1−P(Mi)),Mi ϵ M, where M is the set of all medications for which the patient has an above-threshold probability estimate. P could be the patient's PGx score, essentially.
The probability of each medication is calculated given the disease profile available for that patient. The probabilities for Warfarin being prescribed and Clopidogrel being prescribed, given the disease profile of I48 and I25 are fairly straightforward, as they have only one link. The probabilities are thus the single weight of the edge connecting I48 to Warfarin and I25 to Clopidogrel. For Ticagrelor, the probability is the combination of probabilities of either I48 or I25 or both resulting in Ticagrelor being prescribed. The prescriptions are assumed to be independent, so they can be calculated through a product of the events. P(Ticagrelor|[I48, I25])=P(Ticagrelor|I48)*P(Ticagrelor|I25)+P(Ticagrelor|I48)*P(No Ticagrelor|I25)+P(No Ticagrelor|I48)*P(Ticagrelor|I25). An alternative method for performing the calculation is that the result is just the compliment of both disease diagnoses not resulting in a Ticagrelor prescription: P(Ticagrelor|[I48, I25])=1−P(No Ticagrelor|I48)*P(No Ticagrelor|I25)=1−(1−0.24)(1−0.05)=1−(0.76)(0.95)=1−0.722=0.278. If there were more disease diagnoses present that result in a Ticagrelor prescription, the complimentary probability of that linkage (p) would just be added into the product [1−(0.76)(0.95)(1−p)].
Primary validation of MAPPeR involved assessment of the performance of the disease-to-medication mapping (i.e., the links and weights between the disease status partition 202 and the medications partition 204,
Second, a similar binomial hypothesis-based approach was used to assess the performance of the mapping system on external data. The disease-to-medication mapping was constructed using the entire set of patient data and tested against a set of 170 records from patients in the VA Precision Oncology Cohort A, which was obtained as part of the VA's Al Tech Sprint. External consistency was determined similarly as the internal consistency, with a binomial hypothesis test performed for each of the links between disease and medication. In this assessment, 97.4% of the links were found to be consistent between the mapping and VA patient data.
Lastly, a rigorous assessment of the disease-to-medication mapping was carried out by varying the operational threshold used to define a reliable mapping from disease diagnosis to medication prescription. For this evaluation, operational thresholds of the posterior probability were varied from 0 to 1, which resulted in an Area Under the Receiver
Operating Characteristic Curve (AUC) of 0.737. An optimal operational threshold identified as that closes to perfect performance (Sensitivity=1.0, Specificity=1.0) was determined to be 0.01, corresponding to a 0.562 sensitivity, 0.889 specificity, and 10.259 Diagnostic Odds Ratio. Thus, this threshold 0.01 can be used to make the recommendation of whether or not the patient is likely to benefit from a PGx test, where if the probability P(M) is at or above the threshold the test is recommended.
It is possible to add additional layers of predictive value to our method to make it more sensitive to potential impact of PGx testing and guide the end user in a more nuanced way. For example, we could factor in not just existing ICD-10 codes for existing diseases but also disease risk calculations based on any other relevant clinical data. This would add even more pre-emptive value by noting the likelihood of needing a medication before a disease state has even arisen. Another example would be to weight the likelihood of having an actionable genetic variant based on the prevalence of that genetic variant in the population. Thus, PGx testing may be more or less likely to bring value if the genetic variant in question for a particular drug is more or less prevalent. Disease risk prediction would provide a little more predictive value to the disease status information, i.e. we could preemptively determine what diseases they may experience in a given time frame. The allele likelihoods would not be integrated into the Bayesian network, but they would also help determine the PGx recommendation, in conjunction with predicted medications. For PGx to be useful, a patient needs to be prescribed particular medications and have specific allele(s) related to medication. These two things together will make the results of a PGx test useful. The allele likelihoods would be connected to the predicted medications to make a recommendation. Essentially, a patient might be predicted to be prescribed Codeine, but unless there is some degree of certainty they will have a CYP2D6 variation, the PGx test would not be all that useful.
Our current implementation of MAPPeR uses a server constructed using a Shiny application user interface. Our planned implementation would include improving the user interface and adding connectivity to CMS Blue Button and VA Health API.
In one configuration, implementation into clinical care would come with integration into the clinical practice at a hospital or medical center, and/or the Veterans Administration (VA) or subdivision thereof. We can integrate the backend of MAPPeR into an electronic medical record and test various approaches to alerting a provider that a patient meets the PGx criteria. In addition to adding layers to make the clinical prediction model more sophisticated, interfacing with the EMR can make the whole process more automated so that a provider could be a more passive participant rather than having to affirmatively enter in a patient's data. Thus, MAPPeR would always be working in the background and only alert a provider when they are actively seeing a patient who would benefit from testing.
As noted in
The appended claims are offered as further descriptions of the disclosed inventions. All questions concerning scope are to be answered by reference to the appended claims.
This application claims priority to U.S. Provisional Application Ser. No. 62/986065 filed Mar. 6, 2020, incorporated by reference herein in its entirety.
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/US2021/020996 | 3/5/2021 | WO |
Number | Date | Country | |
---|---|---|---|
62986065 | Mar 2020 | US |