DIAGNOSTIC ACCURACY OF MULTIMODAL AI FRAMEWORKS VS. CLINICAL ASSESSMENT FOR EARLY-STAGE PARKINSON’S DISEASE: A SYSTEMATIC REVIEW AND META-ANALYSIS

Abstract

Background: Early and accurate diagnosis of Parkinson disease (PD) remains a clinical challenge, particularly during prodromal and early symptomatic stages when motor manifestations are subtle, heterogeneous, or overlap with other movement disorders. Conventional clinical diagnosis relies largely on subjective neurological examination, which may result in delayed or inaccurate identification. Advances in artificial intelligence, especially multimodal systems that integrate complementary data sources, have created new opportunities to improve early diagnostic accuracy and reduce uncertainty in clinical decision-making.

Objective: To systematically evaluate and compare the diagnostic performance of multimodal artificial intelligence–based models with conventional clinical assessment in the detection of early-stage Parkinson disease.

Methods: A systematic review and meta-analysis were conducted in accordance with PRISMA 2020 guidelines. Electronic databases including PubMed, Embase, Web of Science, and the Cochrane Library were searched for studies published between 2010 and 2025. Eligible studies included adult participants with early-stage PD, defined as Hoehn and Yahr stages I–II or disease duration of five years or less, and compared multimodal AI-based diagnostic models with standard clinical reference methods. A random-effects meta-analysis was used to calculate pooled sensitivity, specificity, diagnostic accuracy, and area under the receiver operating characteristic curve. Heterogeneity, publication bias, subgroup analyses, and sensitivity analyses were also performed.

Results: Fourteen studies met inclusion criteria. Multimodal AI systems demonstrated high pooled diagnostic performance, with sensitivity of 0.90 (95% CI: 0.87–0.93), specificity of 0.88 (95% CI: 0.85–0.91), diagnostic accuracy of 0.89 (95% CI: 0.86–0.92), and an AUC of 0.93 (95% CI: 0.90–0.95). Conventional clinical assessment showed lower overall accuracy at 0.74, indicating an absolute improvement of approximately 15% with AI-based approaches. Subgroup analyses revealed that fully multimodal AI frameworks achieved the highest pooled accuracy of 0.92 and maintained strong performance even in Hoehn and Yahr stage I disease. Moderate heterogeneity was observed, while sensitivity analyses confirmed result stability and no significant publication bias was detected.

Conclusion: Multimodal artificial intelligence–based diagnostic models consistently outperformed conventional clinical evaluation in identifying early-stage Parkinson disease. These findings support the role of AI-assisted systems as valuable clinical decision-support tools to enhance early diagnosis and improve patient care pathways.