パーキンソン病患者における骨粗鬆症予測のための機械学習モデルの開発と検証

To develop and externally validate machine-learning models for predicting osteoporosis in patients with Parkinson's disease (PD) using routinely collected clinical and treatment-related variables.

We assembled a multicenter retrospective cohort of 3,935 adults with PD, of whom 907 (23.1%) had osteoporosis. Candidate predictors included demographics, lifestyle factors, comorbidities, PD duration and severity, and medication exposures. Features were selected with least absolute shrinkage and selection operator, and nine classifiers were trained in the development cohort. Discrimination was evaluated by area under the receiver operating characteristic curve (AUC), calibration by decile-based plots and logistic recalibration, and clinical utility by decision-curve analysis. Models were assessed in an internal testing cohort and an independent external cohort processed with the same pipeline. Kernel SHAP was used for model interpretability.

Across development, internal testing, and external validation cohorts, models yielded moderate-to-high discrimination with generally favorable calibration. Neural network (AUC = 0.937) achieved the best performance, with support vector machine and random forest also showing strong discrimination (both AUC = 0.935). Decision-curve analysis showed higher net benefit than "treat all" or "treat none" across clinically relevant thresholds. SHAP analyses indicated that lower body mass index, higher Hoehn-Yahr stage, longer disease duration, prior fracture, family history of osteoporosis, sarcopenia, and steroid or proton-pump inhibitor use were major contributors to predicted risk.

Machine-learning models based on routinely available clinical data provide transportable risk stratification for osteoporosis in PD, demonstrating good calibration and meaningful clinical utility. These tools may support targeted screening and individualized management to mitigate fracture risk in this vulnerable population.