An explainable hybrid framework combining latent growth modeling and stacked machine learning for Parkinson’s disease stage prediction

Parkinson’s disease (PD) is characterized by highly heterogeneous motor and non-motor symptoms, and the rate of progression differs across patients, which makes accurate staging difficult and a major challenge. Prediction of PD stage is further complicated by the limitations of cross-sectional methods, which capture only a single time point and fail to reflect longitudinal symptom evolution. To overcome this limitation, in this study motor and non motor symptom domains are analyzed by applying Latent Growth Curve Modeling (LGCM) to measure baseline levels and longitudinal changes in motor function (Movement Disorder Society-Unified PD Rating Scale, MDS-UPDRS-III), impulsivity (Questionnaire for Impulsive-Compulsive Disorders, QUIP), anxiety (State-Trait Anxiety Inventory, STAI), and sleep (Epworth Sleepiness Scale). The coefficients obtained from the LGCM are used as inputs to a stacking machine learning (SML) framework to predict the Hoehn & Yahr stage at two future visits, V n (V10) and V n + 1 (V12). This hybrid approach, referred to as SML β , integrates multiple diverse base learners with a meta-modeling strategy, thereby enhancing prediction accuracy, capturing complex non-linear interactions, and reducing overfitting. Model validity is confirmed using 5-fold cross-validation. Using data from 571 participants in the Parkinson’s Progression Markers Initiative (PPMI), the SML β approach demonstrates strong predictive performance at visit V n (MAE = 0.1850, MSE = 0.0756, RMSE = 0.2750, R 2 = 0.8580 ) and at visit V n + 1 (MAE = 0.2050, MSE = 0.0864, RMSE = 0.2940, R 2 = 0.8080 ), outperforming individual learners. Explainable AI analysis using SHAP identifies motor and sleep trajectories, combined with demographic factors, as the most influential predictors. These findings demonstrate that SML β provides insight into variability in PD progression and enables robust, interpretable prediction of patient staging at future visits.