Parkinson's disease follows no single script. Some patients stay relatively stable for years. Others lose cognitive or motor function much faster. New research published in npj Parkinson's Disease shows that machine-learning models can predict, three to five years in advance, which patients are likely to experience that steeper decline - and the strongest predictive signals came from clinical measurements neurologists already collect during routine visits.
Researchers at the University of Miami trained models on data from 1,602 participants in the Parkinson's Progression Markers Initiative, then validated their findings in an independent cohort of 541 patients from the Parkinson's Disease Biomarkers Program. The models achieved AUROC scores above 0.80, a threshold that reflects strong ability to distinguish higher-risk from lower-risk patients.
Clinical data beat brain scans
The research team expected MRI-based measures of brain atrophy to improve their predictions. That did not happen. Structural MRI data added little beyond what routine clinical evaluations already provided, and in some motor-decline models, adding imaging data actually reduced performance.
"I was very surprised that structural MRI didn't make that much of a predictive difference," said Ihtsham ul Haq, M.D., professor of neurology at the University of Miami Miller School of Medicine and senior author of the study. "It's not that it makes literally no difference, just that what we can measure clinically matters so much more for the model."
The interdisciplinary project brought together neurologists, radiologists, computer scientists and AI specialists. Yelena Yesha, Ph.D., Knight Foundation Endowed Chair of Data Science and AI, and Yusen Wu, Ph.D., a research assistant professor working across neurology and computer science, led the model development. "It finds patterns in disease progression that no single field would catch alone," Wu said. "It doesn't replace clinical judgment. It sharpens it, giving us a real shot at getting ahead of diseases like Parkinson's before the damage is done."
What the models flagged
The most informative predictors split along two tracks. For motor decline, two factors dominated: results from a synuclein seed amplification assay, which detects abnormal alpha-synuclein biology, and how quickly a patient's MDS-UPDRS motor score changed during the first year after diagnosis. For cognitive decline, the strongest predictors were the rate of early cognitive worsening and the pace of motor decline during that same first year.
Prediction performance improved when the models incorporated information about a patient's trajectory over the first 12 months, rather than relying on a single baseline snapshot. "More data isn't necessarily better data," Dr. Haq said. "The information has to be relevant and some of the most useful information in our study came from things neurologists already measure clinically."
Why earlier identification matters
Identifying higher-risk patients earlier could sharpen the design of clinical trials. Researchers currently have limited ability to account for differences in disease progression among participants. "If you are testing whether a drug slows progression, identifying people likely to progress gives you a better chance of detecting whether the treatment actually changes that trajectory," Dr. Haq said.
Growing evidence also suggests that addressing factors such as blood pressure, physical activity, hearing loss and vision problems may influence long-term brain health. Pinpointing patients at elevated risk could underscore the importance of those interventions before decline accelerates. The models may also help researchers understand why some patients deteriorate faster than others and potentially surface new therapeutic targets.
Why this matters for science and research professionals
The study's external validation across two independent datasets is notable: models developed with one large Parkinson's cohort performed comparably when tested on a completely separate group, suggesting the patterns they detect are broadly meaningful rather than artifacts of a single dataset. The counterintuitive finding that structural MRI added little predictive value beyond clinical measures also carries a practical signal - it suggests that high-quality longitudinal clinical data, systematically collected, can power useful predictive models without requiring expensive imaging infrastructure. The team plans to investigate whether similar machine-learning approaches can predict rapid decline in other neurodegenerative diseases, including Alzheimer's, and whether measures of brain connectivity offer predictive information that structural scans do not.
Your membership also unlocks:
