Ordinary smartphone photos of hospital patients’ faces revealed clues to their risk of dying that standard clinical checks sometimes missed.
Indra Heckenbach, an aging researcher at the University of Copenhagen in Denmark, and coauthors analyzed photographs from 27,660 adults admitted through Danish emergency departments. Their median age was 73. The team reports the findings November 19 in the journal Nature Communications.
The researchers trained an artificial intelligence system on earlier patients’ photos and survival records, then tested it on patients photographed in later years. Images were taken during routine care, without standardized lighting or camera positions.
The face scores separated patients who died within 30 days from survivors better than standard early-warning scores based on vital signs. They also identified elevated risk among some patients whose warning scores were low. Combining facial scores with blood tests and vital signs worked better still.
The signal wasn’t simply a measure of looking older: associations with mortality remained after accounting for age. Exactly what the system detected remains uncertain. Color information and features around the eyes appeared to contribute, but the researchers couldn’t rule out influences from lighting, posture or medical equipment.
The photos were taken an average of about six days after admission. The study therefore doesn’t establish how well the system would work when patients first arrive. It also hasn’t shown that using the scores improves care, and broader testing across populations and hospitals is needed.
The authors envision face scores “not as standalone tools,” but alongside vital signs, laboratory measurements and medical records. Two authors have filed a provisional patent related to risk-prediction methods.
The research
Facial photographs reveal mortality risk beyond triage and clinical risk scores ↗
Indra Heckenbach, Kasper Karmark Iversen, Michael Ben Ezra, Mikkel Porsborg Andersen, Henning Bundgaard, Morten Scheibye-Knudsen and Rasmus Bo Hasselbalch
Nature Communications · Accepted
DOI: 10.1038/s41467-026-79581-8