A machine-learning framework called DEGAS sorted early-stage Parkinson’s disease patients into clinically distinct subtypes with balanced accuracy reaching 0.93, according to a study.
DEGAS — short for disease endophenotyping through graph and adjacency metrics — was developed by researchers at the Fixel Institute for Neurological Diseases at the University of Florida. It applies machine learning to functional connectivity patterns drawn from baseline functional MRI scans, aiming to solve a problem: Parkinson’s disease progresses heterogeneously, yet no reliable biomarker currently stratifies patients for individualized care.
That gap matters because patients with the same diagnosis can follow very different trajectories. One person may lose autonomic function and develop hallucinations within a few years; another may remain stable far longer. Until now, clinicians have had no objective, scan-based way to tell them apart at the start.
Computer model sorts Parkinson's patients by brain wiring
- Researchers built a computer model that reads brain-scan “wiring diagrams” and sorts Parkinson’s patients into subgroups — one of which, linked to a sleep disorder, tends to worsen faster.
- The model is a research tool only. It is not available in clinics, and the study’s authors state external validation is required before any clinical use is considered.
- No treatment decision should change based on this study. Its value today is for trial design and research, not diagnosis.
Graph math outperformed raw connectivity
The study drew on 336 early-stage patients from the Parkinson’s Progression Marker Initiative, a public-private partnership funded by the Michael J. Fox Foundation for Parkinson’s Research and dozens of industry and philanthropic partners (PPMI). Patients were stratified by suspected REM sleep behavior disorder, or sRBD — acting out dreams during sleep — which the researchers treat as a marker of more aggressive disease.
The striking finding was what did not work. Region-of-interest analyses showed no group differences in connectivity or topology between the subtypes when the raw scan data were examined directly. Only when the team applied graph theory — a branch of mathematics that describes networks as sets of nodes and connections — did the subtypes separate.
Models incorporating graph theory features achieved a balanced accuracy of 0.91 to 0.93, with a Matthews correlation coefficient (MCC) — a statistical measure of classification quality that penalizes errors in both directions — of 0.61 to 0.78. Adjacency-only models, which used the raw connection matrices without graph-derived summaries, managed a balanced accuracy of 0.79 and an MCC of 0.52.
| Model type | Balanced accuracy | MCC |
|---|---|---|
| Graph theory features | 0.91–0.93 | 0.61–0.78 |
| Adjacency-only models | 0.79 | 0.52 |
In plain terms, the architecture of the brain’s networks — how strongly circuits are organized, not just how strongly individual regions talk to each other — carried the diagnostic signal. The strongest predictors were metrics representing the basal ganglia-thalamic circuits, which govern movement; the temporal-limbic circuits, tied to memory and emotion; and the cerebello-cortical circuits, which coordinate motor control.
A sleep disorder flagged faster decline
The clinical payoff came when predictions were validated against three-year outcomes. Patients the model classified as sRBD-positive showed worse motor outcomes, accelerated autonomic dysfunction — degeneration of the nerves regulating blood pressure, digestion and other involuntary functions — and higher rates of hallucinations, dyskinesias (involuntary, writhing movements often caused by long-term levodopa treatment) and depression.

This reframes a popular misconception. Parkinson’s is often imagined as purely a movement disorder of dopamine-producing neurons in the substantia nigra. The sRBD-linked subgroup’s pattern of autonomic, psychiatric and sleep-related decline is consistent with the established view that Parkinson’s is a whole-network condition, spreading across multiple brain circuits before and alongside motor symptoms.
Who paid for the model
Funding transparency matters here because biomarker research often draws on industry data. A. Enrique Martinez-Nunez of the Fixel Institute — listed alongside Michael S. in the acknowledgments — was funded by a research fellowship from the Napier family, the study’s acknowledgments state. The study itself, the acknowledgments note, received no specific funding source.
The data, however, come from PPMI, whose funding partners span the full Parkinson’s drug-development pipeline, from Abbvie and Biogen to Denali, Roche, Takeda and Verily. That public-private structure is what made a 336-patient, multi-year imaging dataset possible, and it is also why readers should weigh the work as a collaboration between academia and industry-backed data infrastructure.
Validation comes before the clinic
For patients, the distance between this study and the exam room remains considerable. No fMRI-based network classifier is currently an approved diagnostic for Parkinson’s disease in the United States or Europe, and the study’s authors state the constraint directly: external validation is required before clinical use is considered.
That means the model must now be tested on datasets beyond PPMI — different scanners, different populations, different health systems — before any regulator or guideline body would weigh it.
Tool not ready for patients; see neurologist for symptoms
DEGAS involves no drug, device or intervention, so there is nothing to “take” and no contraindication in the pharmacological sense. The relevant risks are interpretive: no patient can be reliably classified by this tool today, and commercial brain scans or consumer wearables cannot substitute for the research-grade functional MRI and validated algorithms used here.
Consult a neurologist promptly if you or a family member experiences any of the features this study tracked: acting out dreams during sleep, new or worsening hallucinations, unexplained falls or blood-pressure instability, or involuntary movements after starting Parkinson’s medication. These symptoms warrant professional evaluation regardless of any biomarker research, and a movement-disorder specialist can assess them properly.
References
- DEGAS, disease endophenotyping through graph and adjacency metrics: network biomarkers for Parkinson’s disease progression (study publication)
- Parkinson’s Progression Marker Initiative (PPMI), funded by the Michael J. Fox Foundation for Parkinson’s Research and partners
Decisions about diagnosis or treatment of Parkinson’s disease should be made with a qualified neurologist.