Supplementary file 1_Machine learning methods for the detection and prediction of cognitive impairment in Parkinson’s disease: a systematic review and meta-analysis.docx

BackgroundCognitive impairment in Parkinson’s disease (PD-CI) is a prevalent non-motor symptom, significantly diminishing quality of life and imposing a substantial family burden. Effective predictive tools are currently scarce, and the diagnostic pathway is intricate. With the growing use of artificial intelligence in healthcare, machine learning (ML) methodologies have been explored for the diagnosis and early risk prediction of PD-CI; however, their efficacy and accuracy necessitate systematic evaluation. Consequently, this investigation undertook a systematic review and meta-analysis.
MethodA comprehensive literature retrieval was conducted across Web of Science, PubMed, Embase, and Cochrane Library, encompassing studies published from database inception to August 10, 2025. The PROBAST tool facilitated quality appraisal, ultimately incorporating 52 publications, of which 25 addressed diagnosis and 27 focused on risk prediction.
ResultsFindings indicated that within the validation cohorts, ML models for PD-CI diagnosis achieved a c-index of 0.82, with a sensitivity of 0.57 and specificity of 0.77. For PD-CI risk prediction, the c-index reached 0.83, accompanied by a sensitivity of 0.77 and specificity of 0.76. These results suggest that ML exhibits considerable accuracy in both the diagnosis and risk prediction of PD-CI. The models primarily incorporated variables such as clinical data, genetic characteristics, biomarkers, neuroimaging, and radiomics, and no overt signs of overfitting were detected.
ConclusionThis research provides an evidence-based foundation for the future development of PD-CI risk prediction and intelligent diagnostic tools, thereby promoting the advancement and application of ML within Parkinson’s disease and related domains.
Systematic review registrationhttps://www.crd.york.ac.uk/PROSPERO/, ID: CRD42023453586.