Researchers trained deep neural networks on 27,660 smartphone photographs of patients admitted to Danish emergency departments to test whether facial images could help predict short-term mortality. The patients had a median age of 73, and 52% were male. The central claim is that a face photograph captures prognostic signals not fully represented by standard measures collected during emergency assessment. The model’s face-derived risk score discriminated mortality better than triage category, early warning scores, vital signs, blood chemistry and estimated facial age.
The score achieved an area under the curve of 0.80 for death within three days and 0.77 for death within 30 days. Its value was especially notable when combined with existing assessment: patients with both high early warning scores and a face score in the highest quartile had about 30 times the 30-day mortality risk of low-risk patients. Among patients with low early warning scores, the highest facial-risk quartile still identified markedly increased risk, with an odds ratio of 13.39. The findings suggest that automated analysis of routine facial photographs could complement triage and monitoring, including by flagging risk among people whose conventional scores appear reassuring.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.