AI Reads Nurses' Notes to Predict When Breathing Tubes Come Out Safely

When a patient is too sick to breathe on their own, hospitals put in a breathing tube and a machine does the work for them. The tricky part is knowing when to take it out. Pull it too early and the patient fails, needing an emergency tube put back in — a frightening, risky ordeal. Wait too long and the patient stays hooked to a machine longer than needed, raising the risk of infection and lengthening a costly hospital stay. That moment of removal is called extubation, and predicting failure is genuinely hard.

A team at the University of Washington tried a new angle. Instead of relying only on tidy numbers — oxygen levels, age, how many days on the machine — they used a large language model (an AI that reads and writes human text) to comb through respiratory therapists' written notes. Those notes are free text, full of observations like how a patient looked, coughed, or responded during breathing trials. The AI turned that messy prose into useful signals, which were then fed into a simple statistical scoring method alongside the standard numbers. Adding the note-based signals improved the prediction.

There's a catch, and the researchers say so themselves. Comparing this study to earlier ones is messy, because different studies define "failure" differently and enroll different kinds of patients. That means a model that looks sharp in one hospital may not travel well to another. The work is also a research result, not a tool you'll see in an ICU tomorrow — it needs validation on far more patients first.

Still, the direction matters. Hospitals already type enormous amounts of text about every patient, and most of it sits unused by prediction tools. Teaching AI to mine that text could make care more personal and better timed — with the honest caveat that it must be tested carefully before anyone's breathing depends on it.