Development and external validation of a machine learning-based model for identifying advanced Parkinson's disease

Identifying patients with advanced stage in Parkinson's disease (PD) is crucial for timely therapeutic intervention, yet current tools rely on subjective clinical scales or expensive biomarkers. We aimed to develop and validate a predictive model based on routine blood biomarkers selected by machine learning algorithms.

We retrospectively analyzed data from 536 PD patients in a discovery cohort and 80 patients in an independent external validation cohort. Patients were classified as having early or advanced stage based on Hoehn and Yahr staging scale. LASSO and Random Forest (RF) algorithms were employed to screen predictors from demographic and routine blood variables. Selected features were integrated into a multivariable logistic regression model to construct a predictive model. Model performance was evaluated using the AUC, calibration curves, and decision curve analysis (DCA).

Six biomarkers including total bilirubin (TB), indirect bilirubin (IBIL), albumin (ALB), cholinesterase (ChE), lactate dehydrogenase (LDH), and creatine kinase (CK), were identified as independent predictors based on LASSO and RF algorithms. The predictive model demonstrated excellent discriminative ability in the discovery cohort (AUC = 0.873) and maintained robust performance in the external validation cohort (AUC = 0.736). Calibration curve showed good agreement between predicted probabilities and observed outcomes. DCA confirmed the clinical net benefit of the model across a wide range of threshold probabilities. Notably, advanced-stage patients exhibited significantly higher levels of TB, IBIL, ALB, and ChE, but lower levels of LDH compared to early-stage patients.

We established a reliable, non-invasive, and economically efficient predictive model using six blood biomarkers to identify advanced PD.