Prediction of First and Recurrent Falls in Parkinson's Disease

To develop and externally validate prediction models for next-year falls in Parkinson's disease (PD), and to evaluate the incremental value of imaging and fluid biomarkers.

We developed models in Parkinson's Progression Markers Initiative (PPMI) to predict next-year falls with external validation in University of Pennsylvania (UPenn) cohorts. We evaluated incident falls among visits without current or prior falls and recurrent falls among visits with current falls.

We calculated area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), and development-derived risk strata. The incident model, comprising Movement Disorder Society-Unified Parkinson's Disease Rating Scale Part III, Montreal Cognitive Assessment, and Geriatric Depression Scale (GDS), was developed in 773 visit-pairs and validated in 589 visit-pairs.

AUROC was 0.704 (95% confidence interval [CI] = 0.653-0.750) in PPMI and 0.630 (95% CI = 0.564-0.687) in UPenn. Incident fall risk increased from 7.1 to a clinically meaningful level across low- to high-risk groups in UPenn.

The recurrent model, comprising Hoehn and Yahr (H&Y), GDS, and Symbol Digit Modalities Test, was developed in 302 visit-pairs and validated in 137 visit-pairs. AUROC was 0.683 (95% CI = 0.623-0.753) in PPMI and 0.641 (95% CI = 0.538-0.732) in UPenn.

Recurrent fall rates increased from 22.2 to 65.0% across low- to high-risk groups in UPenn. Exploratory biomarker augmentation modestly improved discrimination beyond clinical models in PPMI, but not in external validation, where estimates were imprecise.

Clinical measures may identify next-year fall risk in PD with moderate discrimination, although performance attenuated in external validation.

Exploratory biomarkers provided modest incremental value in development but not in external validation. ANN NEUROL 2026.