In Simple Terms
Researchers taught a computer to study face photos and tested whether it could estimate age and survival chances for people with cancer. Higher scores were linked to more deaths in the groups studied, but the tool cannot predict what will happen to one person. It is not ready to guide medical care.
From a Doctor’s Impression to a Computer-Based Measure
Doctors sometimes consider a patient’s general appearance, energy, and frailty alongside test results and medical records. These impressions may inform decisions about surgery, chemotherapy, radiation, or how much support a patient might need. But judgments can vary from one doctor to another because they depend partly on personal experience and observation.
A team led by radiation oncologist Ray Mak of the Mass General Brigham Cancer Institute explored whether computer analysis could provide a more consistent additional piece of information. The goal was not to replace doctors, but to see whether a computer could detect patterns in faces that might be hard to spot by eye and offer a measure to consider alongside other clinical information.
The team first trained the model on about 40 million face images without giving it a specific task. After it learned broad patterns in the images, the researchers adapted it for two jobs: estimating how old a face looked, and analyzing photos of cancer patients to estimate their chances of survival. This apparent biological age is based on facial appearance, not a person’s official age, and it may differ from their actual age.
What Did the Patient Tests Show?
In a test group of 3,820 cancer patients, the researchers sorted the results into different risk-score groups. Of the 122 people in the highest-scoring group, 64 died within 90 days of having their photos taken. This suggests that a high score was linked to a notable death rate over a short period in this group. It does not mean that everyone with a high score will die, or that the score explains why anyone died.
In a separate study involving more than 38,000 patients, the researchers found another link: people whose faces the model judged to look older than their actual age were, overall, more likely to die. This describes a pattern in the data, not a rule that applies to every patient. Trends across large groups do not automatically provide a reliable prediction for any one person.
The researchers reported that combining the face-age estimate with the survival score improved the clinical prediction model compared with using either measure alone. The two measures may capture different information and could be more useful when considered as part of a broader medical picture. However, the available information does not show that the tool is ready to guide routine treatment.
Which Parts of the Face Caught the Algorithm’s Attention?
The team also used a method to see which areas of the face the model focused on when making its estimates. For age estimates, the model highlighted the creases running from the nose to the corners of the mouth, as well as parts of the forehead. For death-risk scores, it focused more on the bridge of the nose and the area beneath the eyes.
But knowing where an algorithm focuses does not explain why those areas are linked to an outcome. The images may contain visible clues related to general health, reflect other overlapping factors, or include patterns the model picked up during training that have no clear biological meaning. The findings do not show that wrinkles in a particular area or the shape of someone’s nose causes risk, or that these features are reliable medical indicators on their own.
The researchers discussed the possibility that appearance may reflect certain habits or health conditions. Mak noted that smoking and heavy alcohol use have been linked to older-looking facial age estimates in some people. But these observations do not prove that the model can directly detect smoking or alcohol use, and they do not rule out other factors that may affect both a person’s appearance and health.
Can Smiling or Lighting Change the Result?
The team tested photos with different facial expressions, makeup, and lighting conditions. The researchers said these differences did not substantially change the results in the analyses they reported. The model could also handle some types of cosmetic surgery when estimating age. Rhinoplasty was an exception: in photos taken after nose surgery, the model tended to estimate a slightly older age than it did in photos of the same person taken before surgery.
This finding highlights that an algorithm’s performance depends not only on whether it can analyze a face, but also on the kinds of images it sees and on changes to a person’s features. Even when a factor does not appear to affect results in a particular test, researchers need to check across diverse groups and realistic photo-taking conditions before drawing broad conclusions.
Important Limits Before Any Medical Use
The research team does not consider the model ready for clinical use. One major limitation is that the cancer-patient data used to develop the survival tool came from the Boston area and included mostly white people. That raises a basic question about how accurately the results would apply to people from different backgrounds, especially since a single group cannot adequately represent the full range of facial features.
The researchers are working to expand the training data to include more diverse faces. This is not a minor technical detail: limited representation can lead to uneven performance across groups. If a single score is presented as objective but is less accurate for some people, it could create a misleading impression or contribute to unfair decisions. That is why the tool needs broad, independent testing, comparisons of performance across different groups, and clear information about its margin of error.
There is also a risk that people could mistake a score for a final judgment. A high score might alarm a patient or lead someone to reduce treatment or care if the result is misread. A low score, meanwhile, could falsely reassure someone who needs urgent attention. Matthew Liao, a professor of bioethics at New York University, has warned that an inaccurate estimate could lead to someone being denied needed care if the tool is trusted too much.
The concerns extend beyond the clinic. In the future, technologies that estimate health from faces could be used by insurers making coverage decisions or by employers screening job candidates. There is no evidence that the current model has been used for these purposes, but the possibility shows why privacy, consent, and fairness need to be considered in advance—including who owns the images and results, and how people can challenge them. An automated score is not automatically neutral, nor is every use of it appropriate.
The Bottom Line
This study points to a new line of research that uses face photos to identify signals that may be linked to outcomes for cancer patients. Higher scores were associated with a greater risk of death in the test groups, and combining the face-age estimate with the survival score appeared to improve predictions compared with either measure alone. But the algorithm cannot yet explain exactly what it is detecting, and the study does not show that particular facial features cause illness or can predict an individual patient’s future. Before face photos could become a tool in hospitals, researchers need broader datasets, evidence that the model works across diverse groups, and careful tests of its practical and ethical effects. For now, the results are a promising research lead—not a diagnosis or a substitute for a full medical assessment.