SYSTEMS, METHODS, AND ARTICLES FOR IMPUTING DIRECTED TEMPORAL MEASUREMENTS
The present disclosure relates to predicting a data element in an electronic health record (EHR) for a subject using a trained machine learning model including an attention module. An example method includes obtaining a query for the prediction of the data element, obtaining a plurality of observations about the subject, processing the query and observations with the trained machine learning model having an attention module to generate a prediction of the subject characteristic, and providing the prediction of the data element as an output.
1 . A method for imputing a value associated with a subject within an electronic health record (EHR) system, the method comprising:
receiving a request to impute the value associated with the subject at a temporal instance;
retrieving a subset of data associated with the subject from the EHR system, the subset of data comprising a plurality of stored values associated with one or more temporal instances;
providing the temporal instance indicated in the request and the subset of data to a trained artificial intelligence engine, the trained artificial intelligence engine configured to perform actions, comprising:
determining relationships between the stored values, the relationships being determined by calculating a set of scores for multiple subsets of features of the stored values that represent interdependencies between the stored values;
adjusting the calculated scores based on a temporal proximity of the one or more temporal instances of the stored values relative to the temporal instance of the value being imputed to generate time-adjusted scores; and
generating an imputed value by creating a weighted combination of the stored values, wherein the weights are determined based on the time-adjusted scores for the multiple subsets of features of the stored values; and
providing the imputed value in response to the request.
2 . The method of claim 1 , wherein the subset of data are associated with the subject.
3 . The method of claim 1 , wherein the request includes a unit of measure.
4 . The method of claim 1 , wherein adjusting the calculated scores based on the temporal proximity of the one or more temporal instances of the stored values relative to the temporal instance of the value being imputed comprises applying a time decay function to the calculated scores that is dependent on differences between the temporal instance of the value being imputed and the one or more temporal instances of the stored values in the EHR.
5 . The method of claim 4 , wherein the time decay function comprises an exponential time decay function or a linear time decay function.
6 . The method of claim 1 , wherein adjusting the calculated scores based on the temporal proximity of the one or more temporal instances of the stored values relative to the temporal instance of the value being imputed comprises applying more weight to stored values that are relatively nearer in time to the temporal instance of the value being imputed.
7 . The method of claim 1 , wherein determining the relationships between the stored values in the EHR comprises use of a multi-head attention module.
8 . The method of claim 1 , wherein generating an imputed value includes:
applying a weight matrix to the time-adjusted scores to combine them into a single representation; and
processing the single representation using a classifier to generate the imputed value.
9 . The method of claim 8 , wherein applying a weight matrix to the time-adjusted scores comprises using a Hadamard product module.
10 . The method of claim 8 , wherein processing the single representation to generate the imputed value comprises using a multilayer perceptron module.
11 . The method of claim 1 , further comprising predicting an occurrence of an adverse event based on the imputed value.
12 . The method of claim 1 , further comprising assessing a predicted eligibility for a clinical trial based on the imputed value.
13 . The method of claim 1 , further comprising predicting a gap in care based on the imputed value.
14 . The method of claim 1 , wherein the imputed value is a lab test value.
15 . The method of claim 1 , wherein the imputed value is a characteristic evaluated by a clinical assessment.
16 . The method of claim 1 , wherein the stored values include at least one prior lab test result.
17 . The method of claim 1 , wherein the stored values include at least one prior clinical assessment result.
18 . The method of claim 1 , wherein the temporal instance of the value being imputed comprises a date.
19 . A computing system for imputing a value associated with a subject within a structured electronic health record (EHR) system, the computing system comprising:
one or more processors; and
one or more non-transitory computer-readable media collectively storing instructions that, when collectively executed by the one or more processors, cause the one or more processors to perform actions, the actions comprising:
receiving a request to impute the value associated with the subject at a temporal instance;
retrieving a subset of data associated with the subject from the EHR system, the subset of data comprising a plurality of stored values associated with one or more temporal instances;
providing the temporal instance indicated in the request and the subset of data to a trained artificial intelligence engine, the trained artificial intelligence engine configured to perform actions, comprising:
determining relationships between the stored values, the relationships being determined by calculating a set of scores for multiple subsets of features of the stored values that represent interdependencies between the stored values;
adjusting the calculated scores based on a temporal proximity of the one or more temporal instances of the stored values relative to the temporal instance of the value being imputed to generate time-adjusted scores; and
generating an imputed value by creating a weighted combination of the stored values, wherein the weights are determined based on the time-adjusted scores for the multiple subsets of features of the stored values; and
providing the imputed value in response to the request.
20 . One or more non-transitory computer readable media collectively storing a computer program thereon, the program, when collectively executed by one or more processors, implements operations for imputing a value associated with a subject within a structured electronic health record (EHR) system, the operations comprising:
receiving a request to impute the value associated with the subject at a temporal instance;
retrieving a subset of data associated with the subject from the EHR system, the subset of data comprising a plurality of stored values associated with one or more temporal instances;
providing the temporal instance indicated in the request and the subset of data to a trained artificial intelligence engine, the trained artificial intelligence engine configured to perform actions, comprising:
determining relationships between the stored values, the relationships being determined by calculating a set of scores for multiple subsets of features of the stored values that represent interdependencies between the stored values;
adjusting the calculated scores based on a temporal proximity of the one or more temporal instances of the stored values relative to the temporal instance of the value being imputed to generate time-adjusted scores; and
generating an imputed value by creating a weighted combination of the stored values, wherein the weights are determined based on the time-adjusted scores for the multiple subsets of features of the stored values; and
providing the imputed value in response to the request.