Finding Precise Causal Multi-Drug-Drug Interactions for Adverse Drug Reaction Analysis

Abstract

Mechanisms are provided for implementing a framework to learn multiple drug-adverse drug reaction associations. The mechanisms receive and analyze patient electronic medical record data and adverse drug reaction data to identify co-occurrences of references to drugs with references to adverse drug reactions (ADRs) to thereby generate candidate rules specifying multiple drug-ADR relationships. The mechanisms filter the candidate rules to remove a subset of one or more rules having confounder drugs specified in the subset of one or more candidate rules, and thereby generate a filtered set of candidate rules. The mechanisms further generate a causal model based on the filtered set of candidate rules. The causal model comprises, for each ADR in a set of ADRs, a corresponding set of one or more rules, each rule specifying a combination of drugs having a causal relationship with the ADR.

1 . A method, in a data processing system comprising at least one processor and at least one memory, the at least one memory comprising instructions executed by the at least one processor to cause the at least one processor to implement a framework for learning drug-adverse drug reaction (ADR) associations, the method comprising:
 generating, by the framework, a set of candidate rules, each candidate rule specifying a relationship between a corresponding drug pattern and a corresponding ADR;   generating, by the framework, a computer executed causal model based on at least a subset of candidate rules in the set of candidate rules, wherein the causal model comprises, for each ADR in a set of ADRs, a corresponding set of one or more causal model rules, each causal model rule specifying a at least one drug having a causal relationship with the ADR, wherein generating the causal model further comprises, for each ADR in the set of ADRs:
 generating a directed acyclic graph data structure that represents each drug pattern of each of the candidate rules, in the set of candidate rules associated with the ADR, as a variable of the directed acyclic graph data structure; 
 determining transition probabilities for transitions between variables specified in the directed acyclic graph data structure; 
 determining causal relationships among one or more drug patterns, represented by the variables, and the ADR based on the directed acyclic graph data structure and the transition probabilities; and 
 generating, for the ADR, the corresponding set of one or more causal model rules based on the determined causal relationships; and 
   executing the computer executed causal model on patient electronic medical records of at least one patient to identify probabilities of the at least one patient encountering one or more ADRs in the set of ADRs.   
     
     
         2 . The method of  claim 1 , wherein generating the set of candidate rules comprises analyzing real world evidence data at least by:
 identifying co-occurrences of references to drugs with references to ADRs in the real world evidence data;   generating the set of candidate rules based on the identified co-occurrences, wherein each candidate rule specifies a drug pattern identifying a plurality of drugs, and a corresponding ADR; and   selecting a sub-set of the candidate rules as a basis for generating the set of one or more causal model rules based on at least one of a support metric and a confidence metric associated with each candidate rule, wherein the support metric measures a number of instances of the co-occurrence, corresponding to the associated candidate rule, in the real world evidence data, and wherein the confidence metric measures, for the associated candidate rule, a probability of the ADR given the drugs in the drug pattern of the associated candidate rule.   
     
     
         3 . The method of  claim 2 , wherein identifying co-occurrences of references to drugs with reference to ADRs in the real world evidence data comprises:
 performing natural language processing of the real world evidence data to identify at least one of terms, phrases, or medical codes identifying references to drugs and references to ADRs;   evaluating relative distances,. within the real world evidence data, between each identified term, phrase or medical code identifying references to drugs and references to ADRs; and   identifying co-occurrences based on the relative distances.   
     
     
         4 . The method of  claim 2 , wherein selecting the sub-set of candidate rules comprises, for each candidate rule:
 generating a contingency table data structure, where each entry in the contingency table data structure comprises a number of patient electronic medical records, that satisfy a condition of the row and column of the contingency table data structure corresponding to the entry, wherein:   a first row of the contingency table data structure corresponds to patient electronic medical records that contain all drugs in the drug pattern of the candidate rule,   a second row of the contingency table data structure corresponds to patient electronic medical records that contain none of the drugs in the drug pattern of the candidate rule,   a first column of the contingency table data structure corresponds to the patient electronic medical records that contain the ADR in the candidate rule, and   a second column of the contingency table data structure corresponds to the patient electronic medical records that do not contain the ADR in the candidate rule.   
     
     
         5 . The method of  claim 1 , wherein generating the computer executed causal model further comprises filtering the set of candidate rules to remove a subset of one or more candidate rules having confounder drugs specified in the drug pattern of the subset of one or more candidate rules, and wherein the filtering further comprises calculating, for each candidate rule, an association score metric that measures a strength of association between the drug pattern represented in the candidate rule and the ADR represented in the candidate rule. 
     
     
         6 . The method of  claim 5 , wherein filtering the candidate rules further comprises:
 calculating, for each first candidate rule, an improvement metric specifying an amount of improvement of a corresponding association score for the first candidate rule over an association score for another second candidate rule specifying a sub-pattern of the corresponding drug pattern of the first candidate rule, and the corresponding adverse drug reaction of the first candidate rule; and   determining, for each first candidate rule, whether to maintain the first candidate rule or remove the first candidate rule based on a value of the improvement metric.   
     
     
         7 . The method of  claim 6 , wherein determining, for each first candidate rule, whether to maintain the first candidate rule or remove the first candidate rule based on a value of the improvement metric comprises:
 comparing the improvement metric corresponding to the first candidate rule to an improvement metric threshold value;   in response to the improvement metric corresponding to the first candidate rule not being equal to or greater than the improvement metric threshold value, determining that a confounder drug is present in the corresponding drug pattern of the first candidate rule; and   identifying the confounder drug in the corresponding drug pattern based on a difference between the corresponding drug pattern and the sub-pattern.   
     
     
         8 . The method of  claim 1 , wherein executing the computer executed causal model further comprises:
 evaluating the patient electronic medical records by applying the causal model to drug history data present in the epatient electronic medical records to identify the probabilities of the at least one patient encountering one or more ADRs in the set of ADRs; and   generating, for each patient in the at least one patient, a patient model based on the identified probabilities of the patient encountering one or more ADRs in the set of ADRs.   
     
     
         9 . The method of  claim 8 , further comprising:
 inputting the patient model into a cognitive system which implements the patient model to perform a cognitive operation based on the patient model.   
     
     
         10 . The method of  claim 9 , wherein the cognitive operation is a treatment recommendation operation that provides a treatment recommendation to a medical practitioner based on an evaluation of the other patient electronic medical record data by the cognitive system, and the patient model. 
     
     
         11 . A computer program product comprising a computer readable storage medium having a computer readable program stored therein, wherein the computer readable program, when executed on a computing device, causes the computing device to:
 receive real world evidence comprising patient electronic medical record data and adverse drug reaction data;   generate, by the computing device, a set of candidate rules, each candidate rule specifying a relationship between a corresponding drug pattern and a corresponding adverse drug reaction (ADR);   generate, by the computing device, a computer executed causal model based on at least a subset of candidate rules in the set of candidate rules, wherein the causal model comprises, for each ADR in a set of ADRs, a corresponding set of one or more causal model rules, each causal model rule specifying at least one drug having a causal relationship with the ADR, wherein generating the causal model further comprises, for each ADR in the set of ADRs:
 generating a directed acyclic graph data structure that represents each drug pattern of each of the candidate rules, in the set of candidate rules associated with the ADR, as a variable of the directed acyclic graph data structure; 
 determining transition probabilities for transitions between variables specified in the directed acyclic graph data structure; 
 determining causal relationships among one or more drug patterns, represented by the variables, and the ADR based on the directed acyclic graph data structure and the transition probabilities; and 
 generating, for the ADR, the corresponding set of one or more causal model rules based on the determined causal relationships; and 
   execute the computer executed causal model on patient electronic medical records of at least one patient to identify probabilities of the at least one patient encountering one or more ADRs in the set of ADRs.   
     
     
         12 . The computer program product of  claim 11 , wherein the computer readable program further causes the computing device to generate the set of candidate rules by analyzing real world evidence data at least by:
 identifying co-occurrences of references to drugs with references to ADRs in the real world evidence data;   generating the set of candidate rules based on the identified co-occurrences, wherein each candidate rule specifies a drug pattern identifying a plurality of drugs, and a corresponding ADR; and   selecting a sub-set of the candidate rules as a basis for generating the set of one or more causal model rules based on at least one of a support metric and a confidence metric associated with each candidate rule, wherein the support metric measures a number of instances of the co-occurrence, corresponding to the associated candidate rule, in the real world evidence data, and wherein the confidence metric measures, for the associated candidate rule, a probability of the ADR given the drugs in the drug pattern of the associated candidate rule.   
     
     
         13 . The computer program product of  claim 12 , wherein the computer readable program further causes the computing device to identify co-occurrences of references to drugs with reference to ADRs in the real world evidence data at least by:
 performing natural language processing of the real world evidence data to identify at least one of terms, phrases, or medical codes identifying references to drugs and references to ADRs;   evaluating relative distances within the real world evidence data, between each identified term, phrase or medical code identifying references to drugs and references to ADRs; and   identifying co-occurrences based on the relative distances.   
     
     
         14 . The computer program product of  claim 12 , wherein the computer readable program further causes the computing device to select the sub-set of candidate rules at least by, for each candidate rule:
 generating a contingency table data structure, where each entry in the contingency table data structure comprises a number of patient electronic medical records, that satisfy a condition of the row and column of the contingency table data structure corresponding to the entry, wherein:   a first row of the contingency table data structure corresponds to patient electronic medical records that contain all drugs in the drug pattern of the candidate rule,   a second row of the contingency table data structure corresponds to patient electronic medical records that contain none of the drugs in the drug pattern of the candidate rule,   a first column of the contingency table data structure corresponds to the patient electronic medical records that contain the ADR in the candidate rule, and   a second column of the contingency table data structure corresponds to the patient electronic medical records that do not contain the ADR in the candidate rule.   
     
     
         15 . The computer program product of  claim 11 , wherein the computer readable program further causes the computing device to generate the computer executed causal model at least by filtering the set of candidate rules to remove a subset of one or more candidate rules having confounder drugs specified in the drug pattern of the subset of one or more candidate rules, and wherein the filtering further comprises calculating, for each candidate rule, an association score metric that measures a strength of association between the drug pattern represented in the candidate rule and the ADR represented in the candidate rule. 
     
     
         16 . The computer program product of  claim 15 , wherein the computer readable program further causes the computing device to filter the candidate rules at least by:
 calculating, for each first candidate rule, an improvement metric specifying an amount of improvement of a corresponding association score for the first candidate rule over an association score for another second candidate rule specifying a sub-pattern of the corresponding drug pattern of the first candidate rule, and the corresponding adverse drug reaction of the first candidate rule; and   determining, for each first candidate rule, whether to maintain the first candidate rule or remove the first candidate rule based on a value of the improvement metric.   
     
     
         17 . The computer program product of  claim 16 , wherein the computer readable program further causes the computing device to determine, for each candidate rule, whether to maintain the candidate rule or remove the candidate rule based on a value of the improvement metric at least by:
 comparing the improvement metric corresponding to the first candidate rule to an improvement metric threshold value;   in response to the improvement metric corresponding to the first candidate rule not being equal to or greater than the improvement metric threshold value, determining that a confounder drug is present in the corresponding drug pattern of the first candidate rule; and   identifying the confounder drug in the corresponding drug pattern based on a difference between the corresponding drug pattern and the sub-pattern.   
     
     
         18 . The computer program product of  claim 11 , wherein the computer readable program further causes the computing device to execute the computer executed causal model at least by:
 evaluating the patient electronic medical records by applying the causal model to drug history data present in the patient electronic medical records to identify the probabilities of the at least one patient encountering one or more ADRs in the set of ADRs; and   generate, for each patient in the at least one patient, a patient model based on the identified probabilities of the patient encountering one or more ADRs in the set of ADRs.   
     
     
         19 . The computer program product of  claim 18 , wherein the computer readable program further causes the computing device to:
 input the patient model into a cognitive system which implements the patient model to perform a cognitive operation based on the patient model, wherein the cognitive operation is a treatment recommendation operation that provides a treatment recommendation to a medical practitioner based on an evaluation of the other patient electronic medical record data by the cognitive system, and the patient model.   
     
     
         20 . An apparatus comprising:
 at least one processor; and   at least one memory coupled to the at least one processor, wherein the at least one memory comprises instructions which, when executed by the at least one processor, cause the at least one processor to implement a framework for learning multiple drug-adverse drug reaction (ADR) associations, which operates to:   receive real world evidence comprising patient electronic medical record data and adverse drug reaction data;   generate, by the framework, a set of candidate rules, each candidate rule specifying a relationship between a corresponding drug pattern and a corresponding ADR;   generate, by the framework, a computer executed causal model based on at least a subset of candidate rules in the set of candidate rules, wherein the causal model comprises, for each ADR in a set of ADRs, a corresponding set of one or more causal model rules, each causal model rule specifying at least one drug having a causal relationship with the ADR, wherein generating the causal model further comprises, for each ADR in the set of ADRs:
 generating a directed acyclic graph data structure that represents each drug pattern of each of the candidate rules, in the set of candidate rules associated with the ADR, as a variable of the directed acyclic graph data structure; 
 determining transition probabilities for transitions between variables specified in the directed acyclic graph data structure; 
 determining causal relationships among one or more drug patterns, represented by the variables, and the ADR based on the directed acyclic graph data structure and the transition probabilities; and 
 generating, for the ADR, the corresponding set of one or more causal model rules based on the determined causal relationships; and 
   execute the computer executed causal model on patient electronic medical records of at least one patient to identify probabilities of the at least one patient encountering one or more ADRs in the set of ADRs.