Explainable AI reveals temporal risk pathways in fall prediction: Extracting clinical insights from multi-horizon...

Area of Science:

• Gerontology
• Medical Informatics
• Machine Learning

Background:

• Falls represent a significant cause of injury and clinical concern in the elderly population.
• Current machine learning (ML) models often provide static fall risk predictions, failing to capture the evolving nature of risk factors over time.

Purpose of the Study:

• To develop and evaluate an explainable ML framework for predicting fall risk across multiple time horizons.
• To differentiate between acute and chronic clinical drivers of fall risk at various temporal scales.

Main Methods:

• A retrospective, matched case-control study utilizing electronic health record (EHR) data from nearly 200,000 patients.
• Trained seven ML models (XGBoost performing best) to predict fall risk across seven distinct time windows (3-60 months).
• Employed SHAP (SHapley Additive exPlanations) for model interpretability to analyze temporal predictor importance.

Main Results:

• A performance trade-off was observed: short-term models (3-12 months) offered balanced discrimination (AUC ≈ 0.75), while long-term models (up to 60 months) improved recall (≈ 80%) at the cost of specificity (≈ 46%).
• SHAP analysis revealed distinct temporal drivers: acute conditions (e.g., syncope, UTI) dominated short-term risk, whereas chronic factors (e.g., spondylopathies, nutritional deficiencies) predicted long-term risk.
• Identified three distinct fall risk trajectories (increasing, steady, decreasing) linked to specific clinical profiles.

Conclusions:

• Fall risk is a dynamic process, necessitating temporally aware prediction models.
• Distinguishing between acute triggers and chronic vulnerabilities allows for personalized, time-aligned fall prevention strategies.
• This multi-horizon approach offers a data-driven foundation for a new paradigm in fall prevention, moving beyond static risk labels.