AI Finds Four Parkinson's Subtypes, Opening Door to Personalized Treatment

Parkinson's disease looks different in every patient. Some people shake first, others struggle with balance or memory, and the disease moves through the brain at different speeds. That variety has been a nightmare for researchers, because a drug that works for one patient may do nothing for another. Now a team of researchers has built a computer model that treats the brain like a road network, tracking how damage spreads from one connected region to the next. Using repeated brain scans from the Parkinson's Progression Markers Initiative — a large long-running patient study — the model grouped patients into four clear patterns of decline.

The key result isn't just that it found four groups. It's that those groups mean something. In this study, the researchers compared their approach head-to-head against SuStaIn, a popular existing method for sorting patients by disease stage. Only their model produced groups that matched up with patients' actual movement symptoms and with known gene variants linked to Parkinson's. In other words, the categories weren't just statistical noise — they reflected real biology sitting inside real people.

Why should you care? Parkinson's is currently diagnosed by symptoms and treated largely by trial and error, with patients often cycling through medications to find something that helps. If doctors can tell early which pattern a patient is on, they can better predict what's coming, enroll the right people in the right drug trials, and eventually prescribe treatments aimed at that specific subtype. Faster, better-targeted trials also mean new drugs could reach patients sooner and cost less to develop.

There's an honest catch. The model was built and tested on data from a few hundred patients in one research cohort, mostly in the US and Europe, so it needs to prove itself in larger and more diverse groups before any doctor uses it. It also depends on repeated brain scans, which are expensive and not always available. And an algorithm that sorts patients into groups is not the same as a treatment that helps them — that step is still ahead.