AI and Neurological Rare & Neuroimmune Diseases ── Parkinson's ── 2026-11-01
Parkinson's disease and parkinsonism: treatment, trials and AI
- Abstract
This systematic review and meta-analysis aimed to evaluate the performance of deep learning (DL)-based methods for classifying individuals with Parkinson’s disease (PD) from handwriting and drawing tasks, following PRISMA guidelines. Google Scholar, IEEE Xplore, PubMed/MEDLINE, ScienceDirect, and Science.gov were searched from 1 July to 17 July 2026, without publication date restrictions. Eligible studies evaluated a DL architecture for PD detection using handwriting or drawing and reported at least one of five performance metrics: accuracy, sensitivity, specificity, precision, or F1-score. Two reviewers independently performed study selection, data extraction, and risk-of-bias assessment using PROBAST+AI. Performance estimates were logit-transformed and pooled separately using random-effects models. Heterogeneity was assessed using Cochran’s Q, I2, and τ2, and results were presented using forest plots with 95% confidence intervals. Potential sources of heterogeneity were investigated through subgroup analyses according to DL architecture, dataset, task, and validation procedure, together with sensitivity analyses. Publication bias was assessed using Egger’s test, and certainty of evidence was evaluated using GRADE. Of 5233 unique records identified, 30 studies were included. Overall pooled performance exceeded 92% across the evaluated metrics. No significant differences were found between fully DL and hybrid approaches or among DL algorithms. Dataset-related differences were observed but were inconsistent across performance metrics. Combined handwriting and drawing tasks consistently achieved higher performance (>95%) than individual tasks, while cross-validation was associated with higher accuracy, specificity, and sensitivity than hold-out validation. Despite these promising findings, substantial heterogeneity and very low certainty of evidence limit the current clinical applicability of these approaches. More standardized protocols, subject-level data separation, and external validation in independent populations are needed to develop robust and generalizable AI models and clarify the potential of handwriting and drawing as non-invasive tools for PD detection.
Journal IF-equivalent: 6.0 (OpenAlex 2-year mean citedness, value as of 2026-10-31, retrieved 2026-11-01; not the official Clarivate IF)Reference: Gorce P, Jacquier-Bret J. Diagnostic Performance of Artificial Intelligence for Parkinson’s Disease Detection Using Handwriting and Drawing Tasks: A Systematic Review and Meta-Analysis. Sensors. 2026 Oct 30;26(21):6893. doi:10.3390/s26216893.Checked: Abstract only - Abstract
Background: Parkinson’s disease (PD) requires both disease-specific knowledge and contextual clinical reasoning. However, most evaluations of medical large language models (LLMs) emphasize closed-ended questions and factual accuracy, providing limited insight into broader dimensions of performance on professional PD tasks.
Methods: Nine contemporary LLMs were evaluated using a 200-item English-language professional PD benchmark spanning five clinical domains, four difficulty levels, and multiple question formats. Responses were assessed for expert-rated factual accuracy, clinical reasoning, Global Quality Scale (GQS), and Flesch–Kincaid Grade Level (FKGL).
Results: The nine models generated 1,800 responses. Between-model separation was minimal for expert-rated factual accuracy (Kendall’ s W = 0.031) but greater for FKGL (W = 0.235), Clinical Reasoning (W = 0.256), and GQS (W = 0.301). Between-model separation in expert-rated factual accuracy was small for closed-ended questions (W = 0.020, p = 0.001) but greater for open-ended questions (W = 0.121, p < 0.001). Expert-rated factual accuracy remained weakly discriminative across PD domains and difficulty levels, while GQS showed the greatest domain-specific separation in Safety & Ethics (W = 0.431). Flagship variants consistently showed higher Clinical Reasoning and GQS, but not uniformly higher factual accuracy.
Conclusion: Contemporary LLMs showed limited between-model separation in expert-rated factual accuracy within this benchmark and scoring framework, whereas greater differentiation emerged in open-ended tasks, clinical reasoning, overall response quality, and readability. Multidimensional evaluation may therefore provide a more informative characterization of model performance and help inform task-specific model selection and oversight in medical applications.
Journal IF-equivalent: 3.1 (OpenAlex 2-year mean citedness, value as of 2026-10-07, retrieved 2026-10-08; not the official Clarivate IF)Reference: Xie L, Wei R, Xiong L, Lu R, Peng K, Ma W. Comparative performance of nine large language models on professional Parkinson’s disease tasks: a multidimensional benchmark study. Front Neurol. 2026 Oct 30;17:1970459. doi:10.3389/fneur.2026.1970459.Checked: Abstract only