Prediction models that can detect the onset of psychotic experiences are a key component of developing Just-In-Time Adaptive Interventions (JITAI). Building these models on passively collectable data could substantially reduce user burden. In this study, we developed prediction models to detect experiences of auditory verbal hallucinations (AVH) and paranoia using ambulatory sensor data and assessed their stability over 12 weeks.

Fourteen individuals diagnosed with a schizophrenia-spectrum disorder participated in a 12-day Ecological Momentary Assessment (EMA) study. They wore ambulatory sensors measuring autonomic arousal (i.e., electrodermal activity, heart rate variability) and completed questionnaires assessing the intensity/distress of AVHs and paranoia once every hour. After 12 weeks, participants repeated the EMA for four days for a follow-up assessment. We calculated prediction models to detect AVHs, paranoia, and AVH-/paranoia-related distress using random forests within nested cross-validation. Calculated prediction models were applied to the follow-up data to assess the stability of prediction models.

Prediction models calculated with physiological data achieved high accuracy both for AVH (81%) and paranoia (69%–75%). Accuracy increased by providing models with baseline information about psychotic symptom levels (AVH: 86%; paranoia: 80%–85%). During the follow-up EMA accuracy dropped slightly throughout all models but remained high (73%–84%).

Relying solely on physiological data to detect psychotic symptoms achieved substantial accuracy that remained sufficiently stable over 12 weeks. Experiences of AVHs can be predicted with higher accuracy and long-term stability than paranoia. The findings tentatively suggest that psychophysiology-based prediction models could be used to develop and enhance JITAIs for psychosis.