When my patient is not my patient : inferring primary-care relationships using machine learning

Abstract

This paper demonstrates that one can infer with respectable accuracy a physician's view of the therapeutic relationship that he or she has with a given patient, using data available in the patient's electronic medical record. In this study, we differentiate between the active primary relationship, the inactive primary or non-attending relationship, and the coverage relationship. We demonstrate that a single model built using the Averaged One-Dependence Estimator (AODE) classifier and learned with six attributes taken from patient visit history and physician practice characteristics can, for most of the 18 physicians tested, differentiate patients with a coverage relationship to a given physician from those with a primary relationship, achieving accuracies of 0.90 or greater as determined by the area under the receiver operating characteristic curve. Three of the 18 datasets had too few coverage patients to adequately characterize. We also demonstrate that, surprisingly, physicians are generally of like mind when assessing the therapeutic relationship that they have with a given patient. We find that for all physicians in our sample, a model built individually with any one physician's assessments performs statistically identically to the model built from the assessments of all other physicians combined. As a sub-goal of this research, we test the performance of different attribute selection methods on our dataset, comparing greedy vs. randomized search and wrapper vs. filter evaluators and finding no practical difference between them for our data. We also test the performance of several different classifiers, with AODE emerging as the best choice for this dataset. Lastly, we test the performance of linear vs. non-linear meta-learners for Stacked

(cont.) Generalization on our dataset, and find no increase in accuracy for the more complex meta-learners.