In addition to prior fall events, motor progression independently predicted next-visit falls with an odds ratio of 1.22 (95% confidence interval, 1.11 to 1.34), while documented motor complications also independently elevated fall risk with an odds ratio of 1.13 (95% confidence interval, 1.03 to 1.23).
The research addressed a longstanding barrier in tracking fall vulnerability in Parkinson’s disease, which evolves through the longitudinal accumulation of both motor and nonmotor features. While clinicians routinely record these critical clinical features within unstructured narrative notes during patient encounters, such narrative details have historically remained inaccessible to structured electronic health record analysis at scale. To overcome this limitation, the investigators created a reusable outcomes research platform driven by a few-shot large language model pipeline designed specifically to extract unstructured clinical data from outpatient neurology documentation.
The automated extraction pipeline was engineered to identify fall occurrence, recent fall frequency, and nine prespecified risk-modifying clinical domains from narrative records. To establish the system’s reliability, the researchers validated the pipeline against expert consensus across a dataset of 2,557 clinical notes. In validation testing, the final model achieved an interrater agreement metric that closely matched human expert agreement. For fall occurrence, final-model agreement reached a Cohen’s kappa of 0.841 compared to an expert interrater agreement kappa of 0.891, while recent fall frequency reached a kappa of 0.889 compared to 0.892 among experts. Furthermore, several extracted risk-modifying domains achieved agreement levels comparable to or exceeding human annotators.
Following validation, the investigators applied the automated pipeline across a comprehensive longitudinal cohort of 25,758 clinical notes from 2,811 patients managed at a tertiary movement disorders center between 2012 and 2025. The authors concluded that this large-scale implementation establishes a robust, scalable framework for narrative phenotyping of fall vulnerability in Parkinson’s disease outcomes research.
Subject of Research: Language model narrative extraction for predicting fall risk in Parkinson's disease
Article Title: Longitudinal fall patterns in Parkinson’s disease: automated narrative extraction in 2811 patients
Article References: Vizcarra, J. A., Xie, K., Super, M., Cao, Q., Bouche, E., Deliberato, R. O., Mendonca, E. A., Tachinardi, U., Siderowf, A., Litt, B., & Ellis, C. A. (2026). Longitudinal fall patterns in Parkinson’s disease: automated narrative extraction in 2811 patients. npj Parkinson's Disease. https://doi.org/10.1038/s41531-026-01625-z
Image Credits: AI Generated
DOI: 10.1038/s41531-026-01625-z
Keywords: Parkinson's disease, fall vulnerability, fall occurrence, fall frequency, clinical narratives, electronic health records, large language models, few-shot learning, natural language processing, narrative phenotyping, lagged generalized estimating equations, motor progression, motor complications, outpatient neurology documentation, tertiary movement disorders center, interrater agreement, outcomes research
News Source: Cassandra Pierce. (November 18, 2026). Language Models Extract Clinical Notes to Predict Next-Visit Parkinson’s Disease Falls. Scienmag.



