Machine learning models trained on routine clinical metrics forecast rapid motor or cognitive decline in Parkinson's disease over three to five years

The study trained models on demographic, motor, cognitive, and imaging variables collected during routine visits. Researchers found that non-imaging clinical measures, such as baseline UPDRS scores and cognitive test results, contributed most to accurate forecasts of rapid progression. This pattern aligns with prior longitudinal cohorts in the PPMI and DATATOP studies, where early motor and executive-function deficits similarly stratified trajectories. The approach did not rely on novel biomarkers, instead leveraging data already captured in standard neurology practice.

These findings extend earlier observational work showing that motor phenotype and mild cognitive impairment at diagnosis predict faster decline, yet add prospective validation across a larger single-center sample. Because the models were developed retrospectively, they cannot yet establish whether early risk stratification alters treatment response or reduces complications. Confounding by unmeasured comorbidities and variable follow-up intervals remains possible, as in most electronic-health-record analyses.

Next steps include external validation in independent registries and randomized trials testing whether model-guided interventions, such as earlier initiation of cognitive rehabilitation or advanced therapies, improve patient-centered outcomes. Regulatory consideration for clinical decision support tools will require prospective performance data and assessment of workflow integration.

Evidence quality is moderate: retrospective single-center design with internal validation demonstrates association but cannot prove causation or generalizability. A multicenter prospective cohort with pre-specified endpoints and blinded outcome assessment is required before routine adoption.