Identifying and predicting fast versus slow Parkinson's disease motor progressors using clinical and digital data

Digital health technologies offer high-frequency, objective measures for Parkinson's disease (PD) yet individual stratification is lacking. This study aimed to identify and predict fast versus slow PD motor progressors by integrating longitudinal clinical evaluation with smartphone motor testing.

In this 96-week subgroup analysis of participants in the Exenatide-PD3 trial, motor progression was measured using Movement Disorder Society-Unified Parkinson's Disease Rating Scale Part III OFF scores and smartphone testing (baseline, weeks 24, 48, 72, 96). Fast versus slow progressors were identified using data-driven subtyping; linear mixed-effects models assessed longitudinal OFF scores. Baseline clinical and digital features (in-clinic and at-home) were used to predict 96-week motor progression. Model performance was quantified by the area under the receiver operating curve (AUC).

Data-driven clustering identified 26.5% (26/98) as fast progressors, with higher baseline OFF scores (+18.16, 95% CI 13.70 to 22.62; adjusted p<0.001) and greater 96-week progression (+8.17, 95% CI 2.75 to 13.60; adjusted p=0.019), undetected by prespecified approaches. Baseline prediction models combining smartphone features, alone or with clinical scores, consistently outperformed the clinical model alone (best AUC=0.78 vs 0.53). Translational value of smartphone assessments was strengthened by high user acceptability (>96% support future in-clinic use) and predictive value in a deployment scenario (AUC 0.80, 95% CI 0.79 to 0.82).

Data-driven clustering identified one in four early-to-mid stage PD patients as fast motor progressors. Integrating smartphone-derived features with clinical scores improved baseline motor prediction. High user acceptability supports this approach for individual-level PD stratification, addressing heterogeneity in motor progression that may obscure treatment trial effects.