AI-based personalized real-time risk prediction for behavioral management in psychiatric wards using multimodal data.

Area of Science:

• Psychiatric care
• Machine learning in healthcare
• Behavioral risk prediction

Background:

• Suicide and harmful behaviors pose significant global health challenges.
• Psychiatric wards face high patient-to-staff ratios and workload issues, hindering real-time risk prediction.
• Existing models based on demographic data lack the precision for timely interventions.

Purpose of the Study:

• To introduce and evaluate the Temporal Fusion Transformer (TFT) model for predicting harmful behaviors in psychiatric settings.
• To integrate diverse data sources including sensor, location, and clinical data for enhanced prediction.
• To improve real-time risk assessment and patient safety in psychiatric wards.

Main Methods:

• Collected hourly data from 145 patients using wearable sensors (heart rate, movement, location).
• Developed and evaluated a binary classification model using the Temporal Fusion Transformer (TFT).
• Employed Bayesian optimization for hyperparameter tuning and 5-fold cross-validation for generalizability.

Main Results:

• The TFT model achieved 95.1% accuracy, 74.9% recall, 78.1 F1 score, and 0.863 AUC, outperforming other models.
• Variable Selection Network identified key predictors like daily entropy and heart rate variability, enhancing interpretability.
• Integration of location and biometric data improved prediction accuracy for real-time risk assessment.

Conclusions:

• This is the first study to apply the TFT model for predicting behavioral risks in psychiatric wards.
• The TFT model effectively integrates diverse data and captures temporal dependencies, suitable for psychiatric environments.
• Future work includes expanding datasets and developing a multimodal Common Data Model (CDM) to enhance clinical utility.