Predictive biosignature of major depressive disorder derived from physiological measurements of outpatients using machine learning

Abstract Major Depressive Disorder (MDD) has heterogeneous manifestations, leading to difficulties in predicting the evolution of the disease and in patient's follow-up. We aimed to develop a machine learning algorithm that identifies a biosignature to provide a clinical score of depressive symptoms using individual physiological data. We performed a prospective, multicenter clinical trial where outpatients diagnosed with MDD were enrolled and wore a passive monitoring device constantly for 6 months. A total of 101 physiological measures related to physical activity, heart rate, heart rate variability, breathing rate, and sleep were acquired. For each patient, the algorithm was trained on daily physiological features over the first 3 months as well as corresponding standardized clinical evaluations performed at baseline and months 1, 2 and 3. The ability of the algorithm to predict the patient's clinical state was tested using the data from the remaining 3 months. The algorithm was composed of 3 interconnected steps: label detrending, feature selection, and a regression predicting the detrended labels from the selected features. Across our cohort, the algorithm predicted the daily mood status with 86% accuracy, outperforming the baseline prediction using MADRS alone. These findings suggest the existence of a predictive biosignature of depressive symptoms with at least 62 physiological features involved for each patient. Predicting clinical states through an objective biosignature could lead to a new categorization of MDD phenotypes.