Towards Just-In-Time Adaptive Interventions - Moment-to-moment Prediction of Negative Affect in Internalizing Disorders using Digital Phenotyping

Negative affect (NA)—encompassing heightened states of sadness, anxiety, and guilt—is a key symptom across a spectrum of internalizing disorders. Recent advancements in digital phenotyping (DP) and machine learning (ML) have raised hopes of automatically identifying short-term fluctuations in NA through digital phenotyping, a prerequisite for Just-In-Time Adaptive Interventions (JITAI). But evidence on the short-term prediction of NA states with passive DP data is sparse. In this preregistered study, we examined whether passive sensor data (heart rate, step count, GPS) and contextual information could predict short-term changes in NA in a sample of 158 outpatients with internalizing disorders. We systematically compared personalized ML approaches (i.e. mixed effect random forests) to population-based models (e.g., random forests) and two benchmark models (individual and population based average NA). We found that personalized ML approaches yielded substantial performance gains compared to population-based models. Nonetheless, the best models only marginally outperformed the benchmark predicting average NA per person. Our findings question the feasibility of detecting short term changes in NA based on completely passively collected sensor data. Future efforts could incorporate richer and/or more raw data streams or test sequential modelling approaches. This will help clarify whether DP and personalized ML can reliably deliver just-in-time, data-driven support for individuals with internalizing disorders.