From Psychosocial State Inference to Temporally Grounded Decision Points for Intervention Timing in JITAIs

Psychosocial states such as stress, focus, and relaxation fluctuate rapidly during cognitive tasks, yet most detection systems rely on snapshot-based inference, identifying states at isolated time points without accounting for their temporal evolution. This limits the ability to anticipate transitions, hindering the anticipatory support required for effective intervention timing. This study presents a temporal modelling approach that infers psychosocial states from multimodal digital biomarkers, including EEG, electrodermal activity, cardiovascular, and behavioural signals, collected across 20 participants performing n-back working memory tasks. Combining unsupervised clustering, Markov modelling, sequential pattern mining, and DTW-based trajectory clustering, the pipeline constructs dynamic psychosocial user models capturing individual regulation profiles and state transition dynamics. Three behavioural clusters emerged: Stress-Dominant, Focused Oscillators, and Volatile Switchers, with stress showing strong temporal persistence. A temporal early warning detection and JITAI validation block identifies 205 biomarker-supported warning events across 19 of 20 participants, each with sufficient lead time for intervention. These results demonstrate that temporal user modelling from digital biomarkers can bridge real-time state inference with anticipatory support by identifying decision points for intervention timing in just-in-time adaptive interventions.