Key Takeaways
- A face score from one smartphone photo predicted 3-day mortality (AUC 0.80) and 30-day mortality (AUC 0.77) in 27,660 Danish emergency patients
- The score outperformed triage level, early warning scores, vital signs, blood chemistry and estimated face age
- Patients flagged high-risk by face score but low-risk by standard scores still had markedly elevated 30-day mortality (OR 13.39)
The face carries information the chart doesn't
When a patient arrives at an emergency department, clinicians score them using triage levels, vital signs and blood work. But a study published in Nature Communications on November 19, 2026, asked a striking question: could a single smartphone photograph of the patient's face carry risk information those tools miss?
The answer appears to be yes. Researchers trained deep neural networks on facial photos from 27,660 patients admitted to Danish emergency departments. The median age was 73. The resulting "face score" predicted 3-day mortality with an AUC of 0.80 and 30-day mortality at 0.77. According to the authors, it outperformed triage level, early warning scores, vital signs, blood chemistry and even estimated facial age.
The signal was strongest where it matters most. Patients who looked low-risk on standard scores but landed in the highest face-score quartile still had substantially elevated 30-day mortality, with an odds ratio of 13.39. High-score patients with high warning scores had roughly 30 times the 30-day mortality of low-risk patients.
This is not the cancer-survival face-age research covered earlier this year. That involved a different country, different patients and a time horizon of years, not days. The study also carries an obvious caveat list: it is one country's emergency population, the model has not been shown to improve outcomes if doctors act on it, and two of the authors have filed a provisional patent on risk-prediction methods.
Health note: This is an early research finding from one country's emergency population. It does not mean a photograph should be used to make treatment decisions.
The Twikup Take
Clinicians have always read faces. "He looks sick" is the oldest triage tool there is. What this study does is show that intuition contains measurable signal a camera can extract, including in patients the standard scores call safe. Whether that signal should ever be used — on consent, bias and dignity grounds — is a harder question than whether it exists. On the evidence here, it exists.
Image: Illustrative image.
