Back

Facial photographs reveal mortality risk beyond triage

Heckenbach, I.; Iversen, K.; Ben Ezra, M.; Andersen, M.; Bundgaard, H.; Scheibye-Knudsen, M.; Hasselbalch, R.

2026-02-15 emergency medicine
10.64898/2026.02.12.26346148 medRxiv
Show abstract

Rapid risk stratification is essential in the clinic, yet vital signs, laboratory tests, and triage scores may not fully capture risk at presentation. We investigated whether facial photographs taken after emergency admission provide an additional mortality signal. Using 27,660 smartphone facial photographs, we trained deep neural networks to identify mortality risk with a Cox proportional hazards framework. Face-derived risk scores strongly stratified short- and long-term mortality, outperforming triage levels, vital signs, blood chemistry, and facial age estimation. Face scores revealed substantial heterogeneity in risk within each triage category, including low-acuity patients with mortality risk exceeding that of high-acuity levels. Combining high-acuity triage with high face score quartiles identified patients with a sixty-fold increased mortality risk within 30 days (odds ratio 60.43 vs lowest risk and triage), while triage alone resulted in six-fold increase in mortality (odds ratio 6.05 vs lowest triage). Integrating facial risk scores with routine clinical variables further improved risk stratification. These findings indicate that unstructured facial photographs encode clinically actionable risk and may augment existing triage and monitoring strategies.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.