Integrating ecological momentary and passive sensing data to improve depression severity prediction: insights from the WARN-D study

Accurately predicting depression severity remains a major challenge due to limited samples, as well as a lack of dynamic, multimodal data. Here we leveraged data from 1,212 young adults collected over 85 days as part of the WARN-D study, combining smartphone-based ecological momentary assessments (EMA) and smartwatch-based passive sensing. We employed Long Short-Term Memory (LSTM) neural networks to predict (1) daily PHQ-2 and (2) weekly PHQ-9 depression scores for 3 months in a preregistered holdout set. Models using EMA data alone achieved the best predictions (daily R2=0.49, weekly R2=0.40), comparable to models with combined EMA and passive sensing. Passive sensing models alone showed no predictive power (daily and weekly R2≈0.0). Variable importance assessment via SHAP analyses highlighted the importance of momentary affect, contextual factors such as self-reported sleep quality, and negative as well as positive experiences throughout the day as EMA key features. Passive sensing features like sleep and physical activity were less informative. However, in combined EMA+passive sensing models, one of the top 10 features related to smartwatch derived step-counts. Our results suggest that EMA features alone may be more predictive for depression severity than passive sensing, but that assessing mood in addition to supplementary environmental information related to context is necessary for gaining better predictions.