GRU-Based Temporal Modeling of Multimodal Signals for Early Dropout in Psychotherapy

Dropout in psychotherapy remains a major clinical challenge, especially in vulnerable populations such as hikikomori youth. We propose a longitudinal, speech-driven early-warning framework to monitor dropout risk in 46 hikikomori patients undergoing ICBT in a real-world clinical setting. Rather than treating dropout as a static endpoint, disengagement is modeled as a discrete-time forecasting task across sessions, integrating self-supervised speech representations, acoustic descriptors, and engagement indicators within a GRU-based temporal architecture. Results show that multimodal modeling captures early disengagement signals, achieving a positive-class recall of 0.73 with specificity of 0.50, and detecting risk up to 2 sessions before dropout. These findings support the potential of longitudinal speech-based monitoring for proactive intervention and improved retention in digital psychotherapy.