AI Reads Messy Hospital Notes to Guess Recovery Scores

Doctors check how well rehab patients are doing only once in a while. They use a standard score sheet called the FIM. But therapists write notes every single day. Those notes are full of little details about what a patient can actually do. A team in Japan wondered if a computer could read those messy notes and guess the scores.

They took notes from three kinds of therapy: physical, occupational, and speech. The notes came from one hospital over five years. Patients had strokes, broken hips, or broken backs. The researchers fed the notes to an AI that ran on local servers. They asked it to predict every single FIM item without giving it any examples first.

The AI tried to guess scores for nearly three million note-item pairs. It only gave valid answers about a third of the time. When it did guess, it was pretty good. Over three-quarters of its guesses were within one point of the real score. Notes from occupational therapy worked better than notes from physical therapy.

This could mean therapists get a daily snapshot of progress without extra testing. But the AI still stays silent most of the time. It also only worked at one hospital with Japanese notes. We don't know if it would work elsewhere or in other languages.

The study shows promise but also big gaps. The AI misses two-thirds of chances to help. It needs more training and testing before it can be trusted in real clinics. For now, human judgment still leads the way.