A LARGE North American prospective cohort study has found that an MRI-based machine-learning tool, Automated Imaging Differentiation for Parkinsonism, can reliably distinguish Parkinson disease from multiple system atrophy and progressive supranuclear palsy, meeting its predefined diagnostic end points and showing strong concordance with neuropathology. The research offers important reassurance that AI-assisted imaging may soon contribute meaningfully to earlier, more accurate diagnostic pathways for complex parkinsonian disorders.
The study, conducted across 21 Parkinson Study Group sites between 2021 and 2024, enrolled 249 rigorously phenotyped patients with consensus diagnoses from three blinded movement-disorder specialists. A retrospective cohort of 396 patients was added to strengthen model training. Diffusion MRI at 3 T was analysed using support vector machine learning, and an independent testing set evaluated real-world performance.
Automated Imaging Differentiation for Parkinsonism achieved high discriminative accuracy: Parkinson disease versus atypical Parkinsonism (AUROC 0.96), multiple system atrophy versus progressive supranuclear palsy (0.98), Parkinson disease versus multiple system atrophy (0.98), and Parkinson disease versus progressive supranuclear palsy (0.98).
Predictive values were similarly strong, with PPVs up to 0.98 and NPVs up to 0.98 depending on the comparison. In a subset of autopsy cases, 46 of 49 imaging-based predictions matched postmortem neuropathology, indicating robust biological validity.
AI Performance Supports Clinical Use in Diagnostic Workups
The findings suggest that Automated Imaging Differentiation for Parkinsonism could meaningfully assist clinicians faced with overlapping motor syndromes, where early differentiation influences counselling, prognosis, and trial enrollment. The authors note that although the tool is not a standalone diagnostic, its performance in a prospective, real-world cohort strengthens the rationale for incorporating advanced imaging-AI pipelines into specialist assessment.
The study also provides one of the largest prospective validations of a machine-learning MRI biomarker in atypical parkinsonism. Future work will test generalizability in community settings and evaluate whether Automated Imaging Differentiation for Parkinsonism can support earlier identification of phenoconversion in at-risk patients.
Reference
David E et al. Automated Imaging Differentiation for Parkinsonism. JAMA Neurol. 2025;DOI:10.1001/jamaneurol.2025.0112.
