Bone Reports · Oct 28, 2026
Hip fracture risk prediction in older adults: Effects of sampling ratio and calibration in random survival forests
By Chida K, Hananoe A, Tetsuo I et al.
Background Hip fracture (HF) is common among older adults. Predicting HF risk may support preventive screening, but rare-event survival modeling presents calibration challenges. Building on prior evidence favoring random survival forests (RSF), we evaluated case-control sampling ratios and post-hoc calibration in RSF, benchmarked against standard Cox regression. Methods In this retrospective claims-based cohort of 2,388,375 individuals (19,463 incident HFs), RSF and Cox models were trained across sampling ratios (1:1 to 1:50) with 100 ID-level bootstrap repetitions and post-hoc reweighted isotonic survival regression (RW-ISR). Performance over 6 years was evaluated among individuals aged ≥75 years using C-index and integrated Brier Skill Score (IBSS). Results For HF prediction among individuals aged ≥75 years, RW-ISR consistently improved calibration while preserving discrimination. The highest median C-index was 0.770 for both RSF and clinical Cox, substantially exceeding demographic baselines (0.728). Calibration was highest under 1:1 sampling with RW-ISR, where clinical Cox achieved superior calibration skill compared with RSF (median IBSS: 0.214 vs. 0.159). Partial dependence in RSF showed clinically plausible associations. Conclusions RW-ISR restored calibration without compromising discrimination. Notably, clinical Cox achieved discrimination comparable to RSF (median C-index 0.770 for both) and superior calibration, indicating that the more complex machine-learning model did not confer additional predictive benefit over standard regression in this older population. Across both models within the experimental conditions examined, optimal sampling ratios differed between discrimination and calibration objectives. Routine claims and checkup data enabled effective population-level risk stratification.