Personalized Early Detection of Depression Onset Using Multivariate Mobile Passive Sensing

Major depressive disorder is common, recurrent, and typically detected only after full symptom onset. Passive smartphone sensing offers a scalable approach to continuous, low-burden behavioral monitoring of depression risk, but prior work applying single-variable statistical process control (SPC) methods to passive sensing data has not yielded reliable prospective detection of depressive episodes. We examined whether multivariate SPC methods applied to passive sensing data could improve prospective detection of depressive symptom onset. We analyzed data from 82 college students (31.7% developed elevated depressive symptoms) followed for up to four years using continuous passive smartphone sensing. We applied exponentially weighted moving average (EWMA) control charts to individual EMA (self-esteem, stress, social level) and passive sensing variables (activity, sleep, location visits), a multivariate EWMA procedure, and a principal component analysis-based EWMA (PCA-EWMA) integrating the full passive sensing feature space. While univariate models performed poorly (MCC = .06–.21), PCA-EWMA showed stronger performance (MCC = .39–.42; sensitivity 81–85%), with up to 76.2% of alerts occurring before future depressive symptom onset. Variable contribution analyses revealed person-specific behavioral signatures driving alerts. These findings show that coordinated behavioral changes passively detected via smartphones can signal rising depressive symptoms.