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

BACKGROUND

Suicide is a major global health issue, with approximately 700,000 deaths annually (WHO). In psychiatric wards, managing harmful behaviors such as suicide, self-harm, and aggression is essential to ensure patient and staff safety. However, psychiatric wards in South Korea face challenges due to high patient-to-psychiatrist ratios and heavy workloads. Current models relying on demographic data struggle to provide real-time predictions. This study introduces the Temporal Fusion Transformer (TFT) model to address these limitations by integrating sensor, location, and clinical data for predicting harmful behaviors. The TFT model's advanced features, such as Variable Selection Networks and temporal attention mechanisms, make it particularly suitable for capturing complex time-series patterns and providing interpretable results in psychiatric settings.

METHODS

Data from 145 patients across three hospitals were collected using wearable devices that tracked heart rate, movement, and location. The data were aggregated hourly, preprocessed to handle missing values, and standardized. A binary classification model using TFT was developed and evaluated with accuracy, recall, F1 score, and AUC. Bayesian optimization was employed for hyperparameter tuning, and 5-fold cross-validation was performed to ensure generalizability.

RESULTS

The TFT model outperformed Multi-LSTM and Multi-GRU models, achieving 95.1% accuracy, 74.9% recall, an F1 score of 78.1, and an AUC of 0.863. The Variable Selection Network effectively identified key predictive factors, such as daily entropy and heart rate variability, improving interpretability. Incorporating location and biometric data enhanced prediction accuracy and enabled real-time risk assessments.

CONCLUSION

This study is the first to use the TFT model for predicting behavioral risks in psychiatric wards. The model's ability to integrate diverse data sources, prioritize cirtical variables, and capture temporal dependencies make it highly suitable for psychiatric environments. While the TFT model performed well, challenges remain with recall due to the limited dataset. Future research will focus on expanding datasets, optimizing variable selection, and standardizing data through a multimodal Common Data Model (CDM) to further improve performance and clinical utility.