Accuracy of Machine Learning in Differential Diagnosis of Multiple System Atrophy: A Systematic Review and Meta-Analysis.

Background

Multiple system atrophy (MSA) is frequently misdiagnosed as Parkinson's disease (PD) or progressive supranuclear palsy (PSP), posing significant diagnostic challenges.

Objective

The aim of the study was to systematically evaluate the efficacy of machine learning (ML) in differentiating MSA from other neurodegenerative disorders.

Methods

We conducted a systematic search in PubMed, Embase, Cochrane Library, and Web of Science (up to July 5, 2025), including ML-based MSA diagnostic studies. QUADAS-2 assessed bias risk, and a bivariate mixed-effects model synthesized diagnostic metrics.

Results

A total of 51 studies (comprising 3595 MSA patients and 18,189 total cases) were included in the systematic review. Of these, 38 studies contributed to the quantitative synthesis, whereas the remaining 13 studies were summarized qualitatively. In the primary analysis comparing MSA with PD, the pooled sensitivity was 0.86 (95% confidence interval [CI]: 0.81-0.90), specificity 0.87 (95% CI: 0.83-0.90), and the area under the summary receiver operating characteristic (SROC) curve (AUC) 0.93 (95% CI: 0.90-0.95). Overall, pooled sensitivity was 0.85 (95% CI: 0.82-0.88), specificity 0.92 (95% CI: 0.89-0.94), and AUC 0.94 (95% CI: 0.92-0.96). Positron emission tomography (PET)-based models achieved optimal performance (sensitivity: 0.90, specificity: 0.96, AUC: 0.97), whereas deep learning models outperformed others (sensitivity: 0.93, specificity: 0.98, AUC: 0.99).

Conclusion

ML models integrating PET/magnetic resonance imaging (MRI) data demonstrate high diagnostic accuracy for MSA. Future work should prioritize prospective multicenter validation and standardized external testing to facilitate clinical implementation.