According to media reports, a international research team recently introduced an AI tool that can “predict life outcomes from a facial photo” – it can estimate the survival chances of cancer patients based solely on a facial image. It is not intended to replace doctors, but rather to serve as a supplement to clinical experience: especially valuable in situations where medical information is scarce and comprehensive assessments are difficult to carry out.
This study was published in The Lancet Digital Health, and the authors described it as a first of its kind. After training, the model fulfilled two roles: it could estimate a person’s age and predict the survival outcomes for cancer patients. The team provided a large amount of data—over 40 million face images, along with photos of more than 30,000 cancer patients, many of whom have already passed away. Finally, it generates a risk score for the patients. Those with the highest scores are more likely to die within 90 days. The tool has an age estimation error of about 5 years, and it can accurately divide the population into high- and low-risk groups. There is a significant difference in mortality rates between the two groups, and it is effective across different ages, races, genders, and types of cancer.
The algorithm focuses on the lower part of the nose bridge, and experts pour cold water on it
Interestingly, it is the position of the algorithm “looking”. The team found that when evaluating the model, the focus is mainly on the bridge of the nose and the area below the eyes. One of the paper authors, Mark, a radiologist at Brigham Hospital in Massachusetts, realized after reviewing literature that Hippocrates described over two thousand years ago: sunken eye sockets, sunken temples, and thin nose tips are early external signs of life approaching its end. In other words, AI has inadvertently quantified ancient medical observations into calculable signals. Mark admitted in an interview that doctors are already judging patients' conditions based on appearance during training. However, “it’s usually very difficult to accurately estimate the remaining survival period of a dying patient.” He believes that a face might be able to reveal some aspects of physiological reserves and the risk of death.
However, there are also voices of caution beyond the applause. Bobbel, a professor of video processing and machine learning at Surrey University, pointed out that the study does not prove that the prediction offers additional clinical value beyond doctors’ assessments of the patient’s overall health condition—it even fails to measure the doctor’s subjective judgment. He added that the tool still needs to be compared with doctors’ evaluations and further verified in multiple hospitals across different patient groups.
From “looking at appearances” to “divination,” AI has turned the dying signs that humans have always intuitively perceived into quantitative indicators. However, whether scores can truly influence treatment decisions remains to be answered by more rigorous clinical trials. The technology has captured the signals, but it does not understand the causes and effects—the machine is still far from achieving this.