- A machine learning model differentiated Parkinson’s disease, multiple system atrophy, and progressive supranuclear palsy.
- AUROCs and predictive values for distinguishing Parkinson’s from mimics were high.
- The algorithm predicted postmortem neuropathology in 93.8% of cases.
An artificial intelligence (AI)-based algorithm differentiated various forms of parkinsonian syndromes based on 3-T MRI.
Across two cohorts, the Automated Imaging Differentiation for Parkinsonism (AIDP) machine learning model distinguished Parkinson’s disease, multiple system atrophy (MSA) parkinsonian variant, and progressive supranuclear palsy (PSP), reported Michael Okun, MD, of the University of Florida Health in Gainesville, and co-authors in JAMA Neurology.
Reporting area under the receiver operating characteristic curve (AUROC), positive predictive value (PPV), and negative predictive value (NPV), the researchers showed that AIDP differentiated:
- Parkinson’s disease from atypical parkinsonism: AUROC 0.96, PPV 0.91, NPV 0.83
- MSA from PSP: AUROC 0.98, PPV 0.98, NPV 0.81
- Parkinson’s disease from MSA: AUROC 0.98, PPV 0.97, NPV 0.97
- Parkinson’s disease from PSP: AUROC 0.98, PPV 0.92, NPV 0.98
AIDP machine learning also predicted postmortem neuropathology in 93.8% of autopsy cases.
“The workup for neurodegenerative diseases commonly includes an MRI. With the AIDP AI-based algorithm approach, we can provide clarity for treating clinicians on the potential diagnosis, inclusive of two Parkinson’s disease mimics,” Okun said.
“This study is a game-changer for using a common MRI scan to differentiate Parkinson’s disease, multiple system atrophy and PSP,” Okun told MedPage Today. “The method was tested on a 21-center NIH and Parkinson’s study group cohort, and was simple, fast, accurate, and effective when applied on many different types of MRI scanners.”
The findings come on the heels of two recent studies about Parkinson’s biomarkers. In 2024, researchers reported that skin biopsies could detect Parkinson’s and other alpha-synucleinopathies. In 2023, a large analysis showed that an alpha-synuclein seed amplification assay could classify people with Parkinson’s disease with high sensitivity and specificity.
The Parkinson’s disease field is considering whether to use a biological classification and staging framework similar to the one for Alzheimer’s disease, Okun and colleagues noted.
“The proposed framework relies on evidence of pathological neuronal alpha-synuclein and dopaminergic neuron degeneration as core biological anchors, regardless of the clinical syndrome,” they wrote. “It is possible that the future application of AIDP, in combination with other neuronal alpha-synuclein biomarkers, may be a useful component of the Parkinson’s classification and staging system.”
The AIDP study included 249 participants from the prospective Parkinson Study Group at sites across the U.S. and Canada, and 396 participants from an auxiliary retrospective cohort.
A total of 500 people were assigned to the training set: all 296 retrospective cohort participants plus 104 from the prospective study. The independent testing set had 145 people from the prospective study, including 60 with Parkinson’s disease, 27 with MSA, and 58 with PSP. Mean age of the testing set was 67.4 and 65.5% were men.
A clinical diagnosis was confirmed by three independent, blinded neurologists who specialized in movement disorders. Diffusion MRI scans were obtained using a variety of scanners. Free water and fractional anisotropy from brain regions of interest, age, and sex were input variables.
Over 3 years, the researchers also collected 49 brains for a pathological investigation. The AIDP diagnosis was confirmed neuropathologically in 46 of the 49 brains, including five of five Parkinson’s brains, five of five MSA brains, and 36 of 39 PSP brains.
Future studies will analyze cases with clinical ambiguity and rater disagreement, Okun and co-authors said. “Although the majority of postmortem brains were from patients with PSP, we anticipate more MSA and Parkinson’s disease brains will become available to expand the pathological validation of AIDP,” the researchers noted.
New studies should consider prodromal cases, cases of dementia with Lewy bodies and corticobasal syndrome, and cases from other clinical settings besides movement disorders centers, they suggested.
Disclosures
This work was supported by a grant from the NIH.
Okun reported relationships with the NIH, Michael J. Fox Foundation, Parkinson’s Foundation, Parkinson Alliance, Smallwood Foundation, UF Foundation, Tourette Association of America, Hachette Book Group, Demos, Manson, Amazon, Smashwords, Books4Patients, Perseus, Robert Rose, Oxford, and Cambridge. He is an associate editor for the New England Journal of Medicine Journal Watch Neurology and JAMA Neurology and participated in educational activities on movement disorders sponsored by WebMD/Medscape, RMEI Medical Education, American Academy of Neurology, Movement Disorders Society, Mediflix, and Vanderbilt University.
Several co-authors reported relationships with Neuropacs. Other authors reported relationships with pharmaceutical companies, nonprofit groups, and other organizations.
Primary Source
JAMA Neurology
Source Reference: Vaillancourt DE, et al “Automated imaging differentiation for parkinsonism” JAMA Neurol 2025; DOI: 10.1001/jamaneurol.2025.0112.
Please enable JavaScript to view the