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

Falls are a leading cause of injury in older adults, making risk prediction a clinical priority. While many machine learning (ML) models exist, they typically provide a static assessment, predicting risk for a single, fixed timeframe. This approach overlooks how the clinical drivers of risk evolve over time, failing to distinguish between acute factors that signal an imminent fall and chronic conditions that confer long-term vulnerability. This study uses an explainable ML framework across multiple time horizons to unlock deeper clinical insights into the temporal nature of fall risk. We conducted a retrospective, matched case-control study using electronic health record (EHR) data from 99,078 patients who fell and 99,078 matched controls. Seven ML models were trained to predict fall risk across seven distinct prediction windows (3, 6, 12, 24, 36, 48, and 60 months). The best-performing model for each horizon, consistently XGBoost, was interpreted using SHAP (SHapley Additive exPlanations) to identify how the importance of clinical and demographic predictors changed over time. A clear performance trade-off emerged across time horizons. Short-term models (3-12 months) delivered balanced discrimination (best model XGBoost, AUC ≈ 0.75), while long-term models became progressively better at identifying eventual fallers (recall ≈ 80% at 60 months) at the cost of lower specificity (≈ 46%). SHAP analysis revealed distinct temporal patterns: short-term risk was driven by acute conditions like syncope, respiratory symptoms, and urinary tract infections, while long-term risk was predicted by chronic, cumulative factors such as spondylopathies, nutritional deficiencies, and benign neoplasms. Three primary risk trajectories (increasing, steady, and decreasing) were identified, each corresponding to distinct underlying clinical profiles. Fall risk is a dynamic process, not a static state. By analyzing risk across multiple timeframes, we can distinguish between acute triggers requiring immediate intervention and chronic vulnerabilities demanding long-term management. This multi-horizon framework provides a data-driven foundation for a new paradigm in fall prevention: moving beyond generic "high-risk" labels to personalized, temporally aware strategies that align the type and timing of interventions with the specific nature of a patient's evolving risk.