Researchers build AI that recognizes pain from brain scans — 10x smaller than standard models, opening doors for real-time pain monitoring in hospitals

A research team created a lightweight transformer (a type of AI architecture that processes multiple data streams together) that analyzes fNIRS brain scans—a non-invasive imaging technique that measures blood flow in the brain—to detect whether a patient is in pain. The model combines multiple views of the same brain data without requiring separate AI components for each view, keeping the overall system small enough to run on standard computers.

Unlike heavier AI models that process each data format separately, this system converts all brain signals into a shared digital language first, then analyzes them together. This approach reduces computing power needed by roughly 90% compared to standard methods while maintaining the same accuracy—meaning hospitals could run real-time pain detection on existing hardware instead of expensive servers.

For patients in intensive care, operating rooms, or unable to communicate (infants, sedated patients, dementia patients), this matters: doctors could get objective, automated pain readings directly from the brain rather than relying on guesswork or patient reports. For hospital IT departments, it means deploying AI pain monitoring without infrastructure upgrades. The model was tested on the AI4Pain dataset and is available as open research.