Explainable AI / Published research
Explainable clinical risk models
A risk score is only the beginning
Understanding which clinical factors matter, and how their importance changes over time.
01 / Inside the method
How it works.
Select a stage to explore the workflow.
Conceptual schematic · Simplified method overview
Retrospective health records supply clinical predictors for studying fall risk.
Seven prediction horizons, from 3 to 60 months, frame separate questions about short- and long-term risk.
Machine learning models estimate risk within each horizon. Comparing horizons helps reveal how the prediction task changes over time.
SHAP attributes model predictions to input features. Comparing these attributions highlights temporal changes in predictor importance; it does not establish causality.
02 / Context & sources
The research.
Looking beyond a single risk score
Our GeroScience study compares fall prediction across seven time horizons, from 3 to 60 months. SHAP analysis examines how the importance of clinical predictors changes between short- and long-term models. The results distinguish acute triggers from chronic vulnerabilities in retrospective health records.
This work evaluates research models; it does not establish a prospectively validated clinical decision tool.
Related publications
- Temporal risk pathways in fall prediction — GeroScience, 2026.
- Risk prediction and interpretation using explainable AI and large language models — ICMHI 2025.
- Explainable AI for BPPV risk assessment — Health Information Science and Systems, published online in 2024.