Supervised Machine Learning to Predict Follow-Up Among Adjuvant Endocrine Therapy Patients

Abstract

Long-term adjuvant endocrine therapy patients often fail to follow-up with their care providers for the recommended duration of time. We used electronic health record data, tumor registry records, and appointment logs to predict follow-up for an adjuvant endocrine therapy patient cohort. Learning predictors for follow-up may facilitate interventions that improve follow-up rates, and ultimately improve patient care in the adjuvant endocrine therapy patient population.We selected 1455 adjuvant endocrine therapy patients at Vanderbilt University Medical Center, and modeled them as a matrix of medical-related, appointment-related, and demographic related features derived from EHR data. We built and optimized a random forest classifier and neural network to differentiate between patients that follow-up, or fail to follow-up, with their care provider for at least five years. We measured follow-up three different ways: thought appointments with any care providers, appointments with an oncologist, and adjuvant endocrine therapy medication records. Classifiers make predictions at the start of adjuvant endocrine therapy, and additionally use temporal subsets of data to learn the change in accuracy as patient data accrues.Our best model is a random forest classifier combining medical-related, appointment-related, and demographic-related features to achieve an AUC of 0.74. The most predictive features for follow-up in our random forest model are total medication counts, patient age, and median income for zip code. We suggest that reliable prediction for follow-up may be correlated with amount of care received at VUMC (i.e., VUMC primary care).This study achieved moderately accurate prediction for followup in adjuvant endocrine therapy patients from electronic health record data. Predicting follow-up can facilitate interventions for improving follow-up rates and improve patient care for adjuvant endocrine therapy cohorts. This study demonstrates the ability to find opportunities for patient care improvement from EHR data.