Doctors have long said some patients simply look sick, and a study published Nov. 19 in Nature Communications puts a number on that instinct. Researchers in Copenhagen trained a neural network on smartphone photos of 27,660 patients admitted through emergency departments in eastern Denmark between 2016 and 2024, and it ranked who would die within 30 days more accurately than the hospital's early warning score, the triage category, vital signs or blood tests.
The photos were the ordinary identification pictures nurses take and attach to a patient's electronic record, with no standard lighting or pose. The team trained the model on photos taken before 2021 and tested it on the 38% taken afterward, which it had never seen. Given two test patients, one who died within 30 days and one who survived, the model put the patient who died higher on its risk list 77% of the time, compared with 50% for a coin flip. For deaths within three days that rose to 80%, and within a day to 82%.
Better than the warning score

Danish hospitals use an early warning score that adds points for abnormal breathing rate, oxygen level, blood pressure, pulse, temperature and alertness. It ranked patients less accurately than the photo did, and so did the five-level triage category, the strongest single blood test (albumin) and the patient's age. Combining everything, the face with age, blood work and vital signs, raised the 30-day score only from 0.77 to 0.82, which suggests the face already carries much of what the lab and the monitor would add.
The patients were old and sick. Their median age was 73, and 6.9% died within 30 days, so deaths were rare in any single day's waiting room even in this group.
What the ranking means for a waiting room
A ranking that is right 77% of the time still flags many people who will be fine. The tool below sorts 1,000 imaginary patients whose scores separate those who die from those who live by the same margin the study reported, and lets you choose how many of the highest-scoring patients to flag for a closer look. It does not look at any face.
Flagging the top quarter of the room catches roughly 60% of the 30-day deaths, but more than eight in 10 flagged patients survive. The study's sharpest numbers come from the patients the usual tools would pass over. Among those whose early warning score looked reassuring, patients in the top quarter of face scores had 13.4 times the odds of dying within 30 days of those in the bottom quarter. When a high warning score and a top-quarter face score appeared together, 30-day mortality was about 30 times higher than among the lowest-risk patients.
What the model seems to be seeing

Deep networks do not explain themselves, so the researchers probed this one. Covering the eyes cut its accuracy more than covering the nose or the mouth. Heat maps of where the network focused moved from the whole image for the lowest-risk quarter to the center of the face (forehead, eyes, nose and the folds beside the mouth) for the highest-risk quarter. Nudging a photo's color toward blue raised the predicted risk more than any other color change, which the authors say may reflect blood flow or oxygen levels, a link they have not tested. Face scores also rose with the number of chronic illnesses and medications a patient had, which fits the idea that the model reads accumulated illness as well as the emergency at hand.
Reasons for caution
The photos were not taken at the front door. On average they were taken about six days after admission, and fewer than one in 10 on the first day, so the model may be seeing patients who were already visibly declining in a hospital bed. The authors acknowledge that the network could be picking up cues other than the face itself, such as image quality, posture, hospital equipment or where the picture was taken. All of the patients came from one region of Denmark, and the model has not been tested outside emergency care. Two of the seven authors have filed a provisional patent on the method.

The researchers present the score as a complement to triage, a second look that might catch the patient whose numbers seem fine but who looks worse than they measure. Putting it into practice would take photos taken at arrival, testing in hospitals with different patients and cameras, and a rule for what a nurse should do when a photo, rather than a symptom, says a patient is in danger.
How we counted: All figures are from Heckenbach, Iversen, Ben Ezra, Porsborg Andersen, Bundgaard, Scheibye-Knudsen and Hasselbalch, "Facial photographs reveal mortality risk beyond triage and clinical risk scores," Nature Communications, Nov. 19, 2026 (accepted version). The study covered 27,660 patients photographed within 30 days of emergency admission in eastern Denmark from May 2016 to May 2024; results are for the 2021 to 2024 test set. The model is an ensemble of 10 EfficientNet-B3 networks trained on survival time. Charts were drawn by The News in Data from numbers reported in the paper; the paper's own figures are at the link below. The waiting-room tool is our illustration, not a result from the study: it draws two bell curves of scores for patients who die and patients who survive, set apart so that the share of correctly ranked pairs matches the reported 0.77 (30 days) or 0.80 (3 days), at the study's mortality rates. Nature Communications paper. Download our data (CSV).