New AI Forecasts ICU Patients' Next 6 Hours — But Isn't Ready for Alarms

A team of researchers has created a new model called PGP-Clinical-TimeKAN that predicts the next six hours of an ICU patient's vital signs — heart rate, blood pressure, oxygen levels, and lab results — based on the previous 24 hours. Instead of forecasting each measurement separately, it treats them as one interconnected system that evolves together. They tested it on records from 6,882 intensive care patients, totaling more than 54,000 time windows.

What makes it different is that it produces probabilities, not just numbers. It can say not only that a patient's blood pressure is likely to drop, but also how likely — including a range of possible outcomes. That's useful in medicine, where knowing the uncertainty in a prediction matters. The model also accounts for missing data and uses basic medical knowledge about how organs affect each other. On several accuracy measures, it outperformed 13 existing models on the same data.

But there's an important catch. The improvement in pure forecasting was tiny — about half a percent better than a simpler existing method. And when the researchers tried to turn these forecasts into a "risk score" that flags which patients are about to deteriorate, the score was less reliable than a dedicated, simpler deterioration-detection model. In plain terms: being good at predicting future vital signs does not automatically make the model good at sounding the alarm.

The authors themselves note this limits what they can claim for clinical use. This is research, not a hospital tool yet. But it shows a promising future where AI monitors a patient's whole trajectory, not just isolated warning signs. The ultimate goal is to give care teams a few extra minutes — or hours — to intervene, potentially saving lives. For now, it's a careful step, not a breakthrough ready for the bedside.