Answers are grounded in this paper and cite their sources.
Fall vulnerability in Parkinson’s disease (PD) evolves through the longitudinal accumulation of motor and nonmotor features routinely documented in clinical narratives but inaccessible to structured electronic health record analysis at scale. We created a reusable platform for PD outcomes research using a few-shot large language model-based pipeline capable of extracting fall occurrence, recent fall frequency, and nine prespecified risk-modifying domains from outpatient neurology documentation. The pipeline was validated against expert consensus in 2557 notes and applied to 25,758 notes from 2811 patients at a tertiary movement disorders center (2012–2025). Final-model agreement closely approximated expert interrater agreement for fall occurrence (κ = 0.841 vs 0.891) and recent fall frequency (κ = 0.889 vs 0.892), while several risk-modifying domains achieved agreement comparable to or exceeding human annotators. We then used lagged generalized estimating equation models to test whether NLP-derived fall status and risk-modifying domains predicted fall occurrence at the subsequent visit. A prior fall (OR, 3.23; 95% CI, 2.89–3.61), motor progression (OR, 1.22; 95% CI, 1.11–1.34), and motor complications (OR, 1.13; 95% CI, 1.03–1.23) independently predicted next-visit falls. This study establishes a scalable framework for narrative phenotyping of PD fall vulnerability.
No takes yet. Share an insight, caveat, or question.
Vizcarra et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: