Predicting symptom worsening in remitted depression on maintenance pharmacotherapy using digital biomarkers: A prognostic modeling study using machine learning

Background: Depression is highly recurrent, and predicting relapses in a timely manner is critical. We applied machine learning to predict the worsening of depressive symptoms. Methods: We conducted a 52-week cohort study of patients with recurrent depression on maintenance pharmacotherapy, using a smartphone app and a wearable device. Participants reported their depression level by filling in the Kessler Psychological Distress Scale (K6) every week on the app. We first classified participants based on their lifestyle characteristics. We then applied the leave-one-participant-out cross-validated (LOOCV) XGBoost to predict K6 scores. We also simulated how the model can perform, where the data of a new patient is collected for some time and then added to the existing dataset to predict the new patient's symptom worsening in the future. Results: We analyzed the data from 89 participants (49 males; median age, 44 years). We identified two distinct clusters of participants: participants in Cluster 1 had unstable sleep patterns and spent more time indoors, whereas those in Cluster 2 spent more time working/studying. The straightforward LOOCV performance showed good AUC but low kappa. When we added observations of a new patient for three months, the weighted kappa between the predicted and the observed K6 classes improved to 0.68 (95 % confidence interval: 0.55–0.81) for Cluster 1 and 0.59 (0.48–0.70) for Cluster 2. Conclusions: Subtyping patients by their behavioral patterns and applying machine learning allowed us to build prediction models for depression relapses among patients on maintenance pharmacotherapy. Funding: Shionogi & Co., Ltd.