Large language models as a screening tool for the qualification of patients with Parkinson's disease for device-aided therapies.
Large language models as a screening tool for the qualification of patients with Parkinson's disease for device-aided therapies.
Where did the research take place?
The study site has not been established. Author addresses may differ from where the research occurred.
Katowice, PL · Author affiliation
Department of Neurology and Stroke Subunit, Kornel Gibiński University Clinical Center of Medical University of Silesia, Katowice, Poland. p.pobudejski99@gmail.com.Location evidence
PL · Author affiliation · country only
Institute of Computer Science, Faculty of Mathematics and Computer Science, University of Wroclaw, Wroclaw, Poland.Location evidence
A plain-language reading has not been prepared for this paper yet.
Original abstract
Motor complications become increasingly prominent as Parkinson's disease (PD) progresses. Although device-aided therapies (DAT) improve symptoms control in advanced PD, referral practices remain inconsistent, partly due to unequal access to specialist centers. Screening tools such as MANAGE-PD help identify patients who may benefit from DAT, while large language models (LLMs) are emerging as scalable clinical decision-support tools. To assess the screening performance and clinical safety of LLM-based triage in identifying PD patients eligible for DAT and to compare LLMs classifications with specialist assessments and MANAGE-PD. This retrospective single-center study included 279 DAT-naïve PD patients referred for advanced therapy evaluation. Clinical data covering motor complications, treatment burden, functional impairment and non-motor symptoms were transformed into standardized narrative case summaries. Seven commercially available LLMs were evaluated in a zero-shot setting, and model outputs were compared with specialist qualification decisions used as the reference standard, and with MANAGE-PD classifications. Sensitivity, specificity, predictive values, accuracy and agreement statistics were analyzed. The best-performing model (gemini-2.5-flash) achieved 100% sensitivity with no false-negative classifications. Specificity reached 69.2%, and overall accuracy was 78.5%. Agreement with specialist decisions was similar for LLM-based screening and MANAGE-PD (κ ≈ 0.575 vs. 0.567), with no significant differences in Cohen's kappa (p = 0.81) or discordant classifications (McNemar p = 1.0). LLM-based triage achieved high-sensitivity screening performance comparable to an established clinical screening tool. These findings support a potential role for LLMs as scalable support systems within DAT referral pathways, complementing rather than replacing specialist clinical assessment.