AI and Neurological Rare & Neuroimmune Diseases ── Parkinson's ── 2026-11-12
Parkinson's disease and parkinsonism: treatment, trials and AI
- Abstract
Systematic review and meta-analysis of supervised machine-learning and deep-learning models that predict a future diagnosis of Parkinson's disease in people without the disease, analysed in two strata: at-risk cohorts (mainly isolated REM sleep behaviour disorder, B1) and general or clinical populations (B2). The protocol and search strategy were frozen and approved by the supervisor on 13 October 2026, before the searches were run on 21-22 October 2026. This registration is retrospective; the frozen files are attached unchanged, and all later amendments are listed in the attached protocol document.
Journal IF: unknown (could not be matched)Reference: [No authors listed]. Machine learning models for predicting future Parkinson's disease: systematic review and meta-analysis. Open Science Framework. [Epub ahead of print]. doi:10.17605/osf.io/qgazv.Checked: Abstract only - Abstract
BACKGROUND: To develop and validate a T1-weighted imaging (T1WI)-based radiomic signature and integrate it with clinical features to construct a combined model for the early identification of cognitive decline in Parkinson's disease (PD). METHODS: We retrospectively analyzed data from 233 PD patients, including 182 from the Parkinson's Progression Markers Initiative (PPMI) and 51 from the National Alzheimer's Coordinating Center (NACC). Based on baseline Montreal Cognitive Assessment (MoCA) scores, patients were classified into those with normal cognition (PD-NC) and those with mild cognitive impairment (PD-MCI). The PPMI cohort was randomly divided into training (n = 127) and test (n = 55) sets at a 7:3 ratio, while the NACC cohort served as an external validation set (n = 51). Radiomic features were extracted from T1WI of the bilateral basal ganglia (globus pallidus, putamen) and bilateral thalamus. After dimensionality reduction, a radiomic model was constructed. Independent clinical predictors were identified and integrated with the radiomic signature to establish a combined model. Model performance was evaluated using receiver operating characteristic (ROC) curves, confusion matrices, and F1-scores. SHapley Additive exPlanations (SHAP) was employed to interpret model outputs. RESULTS: The radiomic model achieved area under the curve (AUC) values of 0.809, 0.777, and 0.776 in the training, test, and external validation sets, respectively. Multivariable regression identified the Geriatric Depression Scale (GDS) score (odds ratio [OR] = 1.41, P = 0.005) and resting tremor (OR = 4.11, P = 0.036) as independent predictors. The combined model incorporating these clinical factors with the radiomic signature yielded significantly improved AUCs of 0.848 (95% CI: 0.774-0.906), 0.796 (95% CI: 0.666-0.893), and 0.812 (95% CI: 0.628-0.925) across the three datasets, outperforming either model alone. Among the 13 retained radiomic features, texture features were predominant (69.2%). SHAP analysis confirmed that the Rad-score was the most influential predictor. CONCLUSION: The combined model integrating T1WI-based radiomic features from the bilateral basal ganglia (globus pallidus, putamen)-thalamus with clinical characteristics demonstrates robust predictive capability for cognitive decline in PD. This approach offers valuable support for early intervention and precision management in clinical practice.
Journal IF-equivalent: 3.8 (OpenAlex 2-year mean citedness, value as of 2026-10-15, retrieved 2026-10-16; not the official Clarivate IF)Reference: Jian Y, Shi H, Hu T, Zhou S, He X, Tao D, et al. Interpretable radiomics signature of the basal ganglia-thalamic circuitry for predicting PD-MCI: a multi-center study with external validation. BMC Med Imaging. 2026 Aug 12;26(1):593. doi:10.1186/s12880-026-02668-7.Checked: Abstract only