FEDI-CODE: A federated and causally-informed framework for dementia risk prediction using multi-site patient data

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Neuroscience

Background:

• Early dementia detection is crucial for intervention but hindered by data fragmentation and privacy concerns.
• Existing methods often require centralized data, posing privacy risks and limiting collaborative research.
• Developing privacy-preserving, collaborative solutions is essential for advancing dementia screening.

Purpose of the Study:

• To propose FEDI-CODE, a Federated and Causally Informed Dementia Estimation framework.
• To enable privacy-preserving, collaborative dementia risk prediction across distributed healthcare data.
• To integrate deep learning, federated learning, and counterfactual inference for robust dementia estimation.

Main Methods:

• Federated learning to train models across institutions without data centralization.
• Deep learning for temporal modeling of longitudinal imaging and clinical data.
• Counterfactual inference for estimating individualized treatment effects on modifiable risk factors.

Main Results:

• FEDI-CODE achieved 83.7% accuracy, 82% F1-score, and 0.86 AUC-ROC on simulated multi-site data.
• The framework demonstrated robustness and generalization on external datasets (79.2% accuracy, 0.80 AUC-ROC).
• Interpretable causal insights were generated through individual treatment effect estimation.

Conclusions:

• FEDI-CODE offers a scalable, interpretable, and privacy-aware solution for early dementia screening.
• The framework facilitates collaborative research and personalized risk assessment across institutions.
• This approach enhances timely intervention and disease management for dementia.