It's All About Specific Features: Development of Prediction Models for Key Geriatric Risks Across Three Hospital Providers.

BACKGROUND

Managing geriatric risks is challenging, as early detection of conditions like delirium, frailty, dysphagia, falls, and malnutrition relies on time-intensive manual assessments requiring geriatric expertise. ML approaches can support healthcare professionals by enabling automated risk prediction using electronic health records (EHR) data, but existing models often lack geriatric-specific features.

OBJECTIVES

This study examines whether incorporating geriatric-specific features improves EHR-based prediction models for key geriatric risks and whether a one-shot Federated Learning (FL) approach enhances performance.

METHODS

EHR data from three hospital providers were linked with acute geriatric data to develop and compare three model types: (1) a baseline model using only EHR data, (2) a model incorporating geriatric-specific features, and (3) a FL ensemble.

RESULTS

Inclusion of geriatric-specific features significantly improved model performance, while a one-shot FL approach provided only marginal, non-significant additional benefit.

CONCLUSION

Geriatric-specific features are essential for capturing the complexity of the acute geriatric population.