Baseline clinical features outperform structural MRI in predicting rapid cognitive and motor decline in Parkinson's Disease

Cognitive and motor decline affect more than 80% of individuals with Parkinson's disease within a decade of diagnosis, yet their trajectories remain largely unpredictable at the individual level. Identifying reliable early-stage prognostic markers could define disease-modification windows and improve trial enrichment. The objective of this study was to determine whether structural MRI-derived atrophy rates or baseline clinical features provide superior prediction of rapid cognitive and motor decline in Parkinson’s disease, and to formally validate the resulting models and their clinical utility. Classification models were developed to predict rapid decline, defined as a decrease of ≥ 5 points on the MoCA and an increase of ≥ 10 points on the MDS-UPDRS3 from baseline to any timepoint 3–5 years post-baseline. Models were trained using longitudinal MRI-derived regional atrophy rates and baseline clinical features independently. The PDBP cohort served as an independent external validation set, and the University of Miami cohort was used for risk stratification. Model performance was evaluated using AUROC and complementary classification metrics; calibration, performance ceilings, and clinical utility were further assessed using calibration curves and decision curve analysis. Feature analysis was performed to identify clinically informative predictors. Structural MRI alone demonstrated limited prognostic utility for individual-level prediction. In contrast, baseline clinical features yielded substantially stronger discrimination for both outcomes. Notably, cognitive prediction remained robust after removal of baseline MoCA, whereas motor prediction was strongly dependent on baseline UPDRS3. External validation in the PDBP cohort (N = 541) preserved high NPVs (cognitive: 0.908 [0.872–0.940]; motor: 0.863 [0.818–0.901]). Incorporating first-year trajectory slopes improved AUROC by approximately 5%, supporting a single follow-up visit as a practical refinement timepoint. Performance gains rapidly plateaued after inclusion of a small number of high-value features, indicating an early ceiling effect. Decision curve analysis confirmed net clinical benefit across a wide range of threshold probabilities.