Supervised machine learning to predict reduced depression severity in people with epilepsy through epilepsy self-management intervention.

Objective: To develop a classifier that predicts reductions in depression severity in people with epilepsy after participation in an epilepsy self-management intervention.

Methods: Ninety-three people with epilepsy from three epilepsy self-management randomized controlled trials from the Managing Epilepsy Well (MWE) Network integrated research database met the inclusion criteria. Supervised machine learning algorithms were utilized to develop prediction models for changes in self-reported depression symptom severity. Features considered by the machine learning classifiers include age, gender, race, ethnicity, education, study type, baseline quality of life, and baseline depression symptom severity. The models were trained and evaluated on their ability to predict clinically meaningful improvement (i.e., a reduction of greater than three points on the nine-item Patient Health Questionnaire (PHQ-9)) between baseline and follow-up (<=12 weeks) depression scores. Models tested were a Multilayer Perceptron (ML), Random Forest (RF), Support Vector Machine (SVM), Logistic Regression with Stochastic Gradient Descent (SGD), K-nearest Neighbors (KNN), and Gradient Boosting (GB). A separate, outside dataset of 41 people with epilepsy was used in a validation exercise to examine the top-performing model's generalizability and performance with external data.

Results: All six classifiers performed better than our baseline mode classifier. Support Vector Machine had the best overall performance (average area under the curve [AUC] = 0.754, highest subpopulation AUC = 0.963). Our analysis of the SVM features revealed that higher baseline depression symptom severity, study type (i.e., intervention program goals), higher baseline quality of life, and race had the strongest influence on increasing the likelihood that a subject would experience a clinically meaningful improvement in depression scores. From the validation exercise, our top-performing SVM model performed similarly or better than the average SVM model with the outside dataset (average AUC = 0.887).

Significance: We trained an SVM classifier that offers novel insight into subject-specific features that are important for predicting a clinically meaningful improvement in subjective depression scores after enrollment in a self-management program. We provide evidence for machine learning to select subjects that may benefit most from a self-management program and indicate important factors that self-management programs should collect to develop improved digital tools.