AI Analyzes Faces to Measure Pain Levels

https://spectrum.ieee.org/machine-learning-measure-pain-surgery

Publish Date: 2026-01-12 11:52:21

Summary:

This innovative study, published in the IEEE Open Journal of Engineering in Medicine and Biology, aims to address the significant challenge of monitoring pain levels in patients who are conscious but unable to communicate their pain, such as infants or those with dementia. A team of researchers, led by Bianca Reichard from the Institute for Applied Informatics in Leipzig, Germany, has developed a contactless, camera-based pain monitoring technique that analyzes patients’ facial expressions and heart rate data to estimate their pain levels. The approach utilizes a machine learning algorithm that scans facial expressions and remotely detects heart rate via photoplethysmogram (rPPG), leveraging a light-shining method that measures changes in blood volume. Contrary to other methods that use brief, ideal video clips, this system was trained on longer, more realistic footage from surgical scenarios, reflecting real-world conditions that include partial obstructions and other challenges. The resulting model shows around 45 percent pain prediction accuracy, a notable achievement considering the video footage’s realistic disturbances. Though the accuracy could likely improve with more complex models, this innovative and practical method stands out for its non-intrusive nature and realistic training data.

Key Points:

Researchers developed a contactless camera-based method to monitor patients’ pain by analyzing facial expressions and remote heart rate data.
The machine learning system uses remote photoplethysmogram (rPPG) to detect heart rate changes through reflected light on the skin, avoiding wired sensors.
The model was trained using both realistic, extended surgical videos and a newly developed dataset from patients undergoing heart procedures.
This approach achieved an accuracy rate of 45 percent in pain prediction, significantly valid given the realistic, often challenging footage conditions.
The simple statistical model used could potentially be improved using more complex techniques such as neural networks.