Measurement fidelity, not model complexity, governs individual cognitive trajectory prediction in Parkinson's disease.

Research overview

The text below is the source article’s original abstract under a verified reuse license. cc by

Whether individual-level cognitive trajectories in Parkinson's disease are predictable remains unresolved. Here, we provide convergent evidence that measurement fidelity, rather than model complexity, governs the prediction ceiling. We evaluated ten machine learning paradigm families across 26 configurations in two independent cohorts-the Parkinson's Progression Markers Initiative (PPMI, N = 1018) and the National Alzheimer's Coordinating Center (NACC, N = 523). Motor subtype classification succeeded as a positive control (AUROC = 0.869), while cognitive trajectory prediction from the Montreal Cognitive Assessment remained uninformative (AUROC = 0.581) across all methods, including 42 genetic features. Synthetic experiments confirmed that models recover trajectories at high signal-to-noise ratios but collapse uniformly at levels present in clinical screening data. Replacing coarse screening with detailed neuropsychological tests improved prediction in both cohorts (PPMI: Symbol Digit Modalities Test AUROC = 0.725 vs. MoCA 0.596; NACC: Logical Memory AUROC = 0.684 vs. MMSE 0.648). Within-patient trajectory R² = 0.20 for MoCA domains confirmed that approximately 80% of score variance represents non-signal variance. These findings redirect the clinical AI agenda from algorithm development toward measurement innovation: higher-frequency, higher-fidelity cognitive instruments are necessary before individual-level prediction becomes feasible.

Related topics