Acute psychiatric inpatient wards provide essential therapeutic functions that community-based services cannot replace, yet they face significant challenges including inadequate institutional support and staff burnout. Wearable sensors and artificial ...
Acute psychiatric inpatient wards provide essential therapeutic functions that community-based services cannot replace, yet they face significant challenges including inadequate institutional support and staff burnout. Wearable sensors and artificial intelligence hold the potential to overcome these constraints by enabling objective and continuous measurement of patients' behaviors and biosignals. This study aimed to develop and validate artificial intelligence-based clinical prediction models for clinical crises in acute psychiatric wards using wearable sensor data.
A multicenter cohort was established across four acute psychiatric wards in three medical institutions in South Korea. A total of 275 inpatients aged 13 years or older diagnosed with schizophrenia, schizoaffective disorder, or mood disorders were prospectively recruited. Comprehensive clinical information and wearable device data, including heart rate, acceleration, location, and activity information, were collected, yielding 5,658 person-days of data.
Study 1 developed and validated a deep learning model to predict multidimensional psychiatric symptoms—psychosis, anxiety, depression, and mania—measured by clinician-rated scales at approximately weekly intervals. Using a time-series deep learning architecture with wearable sensor data as input, the model performed two prediction tasks: (1) binary classification of symptom deterioration compared to the previous week, and (2) regression prediction of symptom severity scores. The multitask learning model, which simultaneously predicted multiple symptoms, achieved a mean area under the receiver operating characteristic curve (AUROC) of 0.73 (range across symptoms 0.69–0.77) for predicting symptom deterioration in external validation, comparable to single-task models that predicted individual symptoms separately (mean AUROC 0.71; range 0.67–0.78). For symptom score prediction, the multitask model achieved a mean R² of 0.82 (range 0.70–0.87), outperforming single-task models (mean R² 0.62; range 0.42–0.77). Feature importance analysis identified age and other basic demographic information, the number of unique heart rate and acceleration values, and location entropy as key predictive factors.
Study 2 developed and validated a real-time prediction model for seclusion, restraint, and injections. Using an attention-based time-series prediction architecture, the model predicted the occurrence of these events at 15-minute intervals. Six prediction horizons were defined, ranging from within 1 hour to within 24 hours, and multitask learning models that simultaneously predicted all horizons were compared with single-task models that predicted individual horizons. In external validation, the highest performance was achieved by the single-task model for prediction within 24 hours (AUROC 0.83). Mean AUROC across horizons was 0.76 (95% confidence interval [CI] 0.74–0.77) for the multitask model and 0.78 (95% CI 0.77–0.79) for single-task models, with the multitask model showing significantly lower performance (adjusted P=0.01). However, mean F1 score was 0.41 (95% CI 0.40–0.43) for the multitask model and 0.31 (95% CI 0.30–0.33) for single-task models, with the multitask model demonstrating significantly superior performance (adjusted P<0.001). For shorter prediction horizons within 6 hours, the multitask model showed more than a three-fold improvement in area under the precision-recall curve (AUPRC) relative to the baseline event rate. Feature importance analysis identified age, location-based distance traveled, circadian rhythm, and location entropy as key predictive factors.
This study demonstrated that psychiatric symptoms and clinical crises can be predicted using wearable sensor data without extensive clinical information input, in a multicenter, transdiagnostic acute psychiatric ward cohort. Multitask learning-based deep learning enabled robust and efficient prediction even in a limited data environment. Age and sex served as contextual information for interpreting wearable sensor signals, while circadian rhythms and indoor location patterns were confirmed as important features for predicting mental health outcomes. Considerable inter-ward heterogeneity was observed in clinical characteristics and wearable sensor data, suggesting that site-specific models may be more appropriate than universal models for future clinical prediction in acute psychiatric wards. This study has limitations including limited sample size, selection bias, and restricted external validity. Future work requires large-scale consortium-based integrated data collection, integration and standardization with electronic health records, and prospective clinical validation built upon ethical and institutional foundations. This study aimed to establish the groundwork for a human-centered artificial intelligence-based clinical decision support system that enhances the well-being of both patients and staff, carrying forward the tradition of systematic observation in psychiatric wards into the 21st century through digital technology.