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

AbstractAccurately predicting depression severity remains a major challenge due to small samples, inaccurate out-of-sample predictions, and a lack of intensive longitudinal data from multimodal data sources. 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 on a training set of 854 participants to predict (1) daily PHQ-2 and (2) weekly PHQ-9 depression scores for 3 months in a preregistered holdout set (n=358). Model estimates are highly conservative, given that we excluded PHQ-2 and PHQ-9 as predictors. We compared EMAonly, passive sensing only, and combined models. The EMA-only model achieved the best prediction (daily R2=0.37, weekly R2=0.21), comparable to the combined model. Passive sensing models alone showed no predictive power (daily and weekly R2≈0.0). SHAP derived variable importance highlighted momentary affect, self-efficacy, and contextual factors like negative and positive experiences throughout the day as key EMA features. Passive sensing features like sleep and physical activity were less informative. Interestingly, one of the top 10 features in the combined model was smartwatch-derived physical activity intensity. In sum, EMA show strong performance predicting depression severity. To best leverage such data, it is important to go beyond affect and symptoms, and also contextual information like stressors.