New AI Reads Doctor's Notes to Predict Which Patients Will Get Sicker

Most hospital prediction software works off tidy, structured data: lab values, heart-rate readings, billing codes. But a huge amount of what doctors actually know lives in messy free-text notes — "patient seemed confused this morning," "stopped responding to the antibiotic." That context usually gets thrown away. HERMES is a research system that uses AI to read those notes and keep the connections between events intact, rather than flattening the story into a single blob of text.

The trick is turning each patient's notes into a kind of relationship map — a knowledge graph, which is just a web of facts and how they link together. If a treatment failed and the patient got worse, that link is captured explicitly. A second AI layer then studies these maps and produces a risk score. The team tested it on MIMIC-III and MIMIC-IV, two large public databases of real intensive-care records, looking at in-hospital deaths and 30-day readmissions. HERMES consistently outperformed AI models that only read the raw text.

Why should you care? Readmissions are expensive and often preventable, and hospitals are financially penalized for them. A tool that flags a patient as high-risk before discharge could mean extra follow-up calls, a home visit, or a medication change — the kind of small intervention that keeps people out of the hospital. Earlier warning also helps doctors escalate care sooner when someone is quietly declining.

The catch: this is a paper, not a product. It was tested on historical records from one hospital system, not rolled out in live wards, and results on past data don't always transfer to new hospitals with different note-taking habits. There are also real privacy questions about feeding patient notes into AI, and no algorithm replaces a clinician's judgment. Expect this to shape hospital software in a few years, not next month.