Parkinson's AI Summarizes Years of Doctor Visits With Far Fewer Errors

Parkinson's disease is a lifelong condition, so patients see doctors many times over years. Specialists must read every past visit to spot whether symptoms are getting better, worse, or staying the same. That is slow and exhausting work, and there simply are not enough Parkinson's specialists to go around. AI could help, but earlier AI models were unreliable — they often invented facts or got dates wrong.

Researchers built a new system called MA-RAG (multi-agent retrieval-augmented generation). Think of it as a team of AI assistants instead of one rushed writer. One agent retrieves the relevant parts of medical records, another checks numbers and dates, another verifies time order, and a final agent reviews everything before producing a plain-language summary. It can answer real clinical questions, like "Has this patient's mobility declined since last year?"

The results are striking. On clinical tests, MA-RAG improved factual accuracy by 122% and reduced the rate of AI "hallucinations" — made-up information — by 98%. Human Parkinson's specialists rated its summaries as highly organized and clinically useful, much better than summaries from standard AI tools. This suggests a team-of-specialists approach can make AI trustworthy enough to help in real clinics.

There is, of course, a catch. The system was tested on research data, not yet in a working hospital. Real-world medical records are messier, and privacy rules add extra hurdles. Also, doctors should always double-check AI summaries — but this research shows that combining fact-checking agents can turn a flaky assistant into a reliable one.