FedTFT: Federated Temporal Fusion Transformer for Interpretable Multi-Horizon Psychiatric Risk Prediction Across...

Area of Science:

• Artificial Intelligence
• Clinical Informatics
• Psychiatry

Background:

• Psychiatric inpatient monitoring yields rich multimodal data.
• Privacy concerns and data variability across hospitals hinder centralized machine learning.
• Existing methods struggle with heterogeneous, decentralized psychiatric data.

Purpose of the Study:

• To develop a privacy-preserving federated learning framework for multi-horizon psychiatric risk prediction.
• To address data heterogeneity and privacy constraints in clinical settings.
• To improve the accuracy and interpretability of psychiatric risk forecasting.

Main Methods:

• Federated Temporal Fusion Transformer (FedTFT) model utilizing horizon-decoupled prediction heads.
• Area Under the Receiver Operating Characteristic Curve (AUROC)-weighted server aggregation for non-IID data.
• Local training with proximal updates and gradient SHapley Additive exPlanations (SHAP) for interpretability.

Main Results:

• FedTFT achieved 93.9% accuracy, AUROC 0.9054, event-F1 0.8242, and Brier score 0.0680 on a global holdout set.
• Significantly outperformed competing federated baselines in event-F1 (19.58 pp improvement) and AUROC (2.74 pp improvement).
• SHAP analysis identified key predictors like treatment time, heart-rate, and mobility changes.

Conclusions:

• FedTFT offers an accurate, calibrated, and interpretable solution for privacy-preserving psychiatric risk forecasting.
• The horizon-decoupled design and AUROC-weighted aggregation are crucial for performance.
• Enables proactive interventions through reliable, decentralized risk prediction.