Interpretable survival prediction in Parkinson’s disease with AI-assisted web deployment

Abstract

Survival prediction in Parkinson’s disease (PD) remains limited, and few models have been translated into reproducible web-based tools. We analyzed 3148 patients with PD from the Chinese Parkinson’s Disease Registry, enrolled at 19 tertiary hospitals between 2018 and 2020 and followed through December 31, 2024. Cox regression, random survival forest, survival tree, and XGBoost survival models were compared using repeated five-fold cross-validation with multiple imputation performed within training folds. The final Cox model was fitted across 20 completed datasets, pooled using Rubin’s rules, and evaluated using split-first temporal validation. The 11 predictors retained in the final model covered demographic, genetic, treatment-related, motor, and non-motor domains. Cox, random survival forest, and XGBoost showed similar discrimination, with mean C-index values of 0.716 for all three models, whereas the survival tree performed lower, with a mean C-index of 0.690. For the deployed Cox model, the mean 2-, 4-, and 6-year AUCs were 0.713, 0.726, and 0.750. In temporal validation, the C-index was 0.708, with modest underestimation of absolute mortality risk. The selected Cox model was implemented in a web platform providing individualized survival estimates, with language-model modules limited to structured data entry and general PD education.