Can an ordinary photograph of a face reveal a person’s health and even help predict their life expectancy? Scientists have developed FAHR-Face, an artificial intelligence system that can estimate a person’s biological age and identify visual features associated with mortality risk. To train the model, researchers used more than 40 million facial photographs, and tested its predictive capabilities using data from patients with cancer. The results were promising.
Does this mean that selfies will soon make it possible to assess someone’s health and predict their life expectancy? It is too early to say: the models were studied primarily in the context of cancer, and their results cannot accurately predict how long a specific person will live. Let’s look at how FAHR-Face works, what the researchers found, and what this discovery could mean for the future of medicine.
The FAHR-Face base model was first trained on more than 40 million photographs from the WebFace42M dataset. According to the authors’ documentation in the AIM-Harvard repository, these are images of faces. This is not a database of 40 million cancer records with known treatment outcomes.
Tool | What it was trained to do | How the result is used |
|---|---|---|
FAHR-FaceAge | Predict chronological, or calendar, age from a photograph | Researchers study whether the difference between the estimated and actual age is associated with survival |
FAHR-FaceSurvival | Estimate the risk of death from patients’ photographs | Researchers assess how well the measure separates patients by prognosis |
The phrase “biological age from the face” needs a caveat here. The algorithm does not measure the aging of the heart, kidneys, or immune system. Researchers consider a photographic estimate as a possible biomarker—a proxy that may be associated with health.
The FAHR-FaceAge algorithm estimated a person’s age with an average deviation of just 5.1 years from their actual age, according to the authors’ repository.
The researchers tested the AI on two groups of patients who had received radiation therapy. The first included more than 38,000 people from the Harvard clinical database; the second consisted of nearly 4,900 patients from the independent Maastro database. However, publicly available study materials contain a small discrepancy in the size of the first group that needs clarification.
The results showed that people whose AI-estimated age was five or more years higher than their actual age had an increased risk of death. The difference was 21% in relative risk, even after accounting for other clinical factors. However, this does not mean that each such patient’s probability of death increases by 21 percentage points.
Another model, FAHR-FaceSurvival, also demonstrated the ability to predict risk. In the test group, its C-index was 0.72. This means that in about 72% of comparable patient pairs, the algorithm correctly identified who would experience an adverse event sooner. However, the model’s performance varied across patient groups, with scores ranging from 0.53 to 0.88.
AI can identify facial features in photographs that are associated with disease prognosis, but it cannot yet reliably predict the life expectancy of a specific person.
According to Mass General Brigham, combining the two measures improved the model’s fit to survival data in an external validation group. Improving a statistical model does not, by itself, prove that it will benefit treatment.
The authors conducted a retrospective analysis, comparing photographs with clinical data that had already been collected. They did not test an approach in which doctors use an AI recommendation, change treatment, and then check whether patients fare better. The study examined overall survival, which includes deaths from any cause, so it would be a mistake to automatically attribute every outcome to cancer.
The same question arises here as when discussing a whole-body MRI scan in the absence of symptoms: what will actually change for a person after they receive the result? A technology’s ability to produce a measure does not answer this question.
Raymond Mak, part of the development team, describes the system’s purpose as supporting personalized treatment decisions. Mass General Brigham also reports that some of the authors have filed a provisional patent application for algorithms that analyze health from facial images. This interest should be taken into account when evaluating future claims about the technology.
You can already upload a photograph and get a research estimate from FaceAge: this option is available through the open FaceAge Research Study. The authors intend the published models for academic research and say they need further validation before clinical use.
The study described does not establish that it can. Its results concern overall survival; tumor recurrence is a separate outcome that requires separate validation.
The reliability of this approach has not been confirmed in the study presented. A difference between two age estimates does not prove that treatment has worked or that someone’s health has worsened.
You should not submit someone else’s photograph to a research service without their informed consent. First, discuss the purpose of uploading it, how it will be stored, and how it might be used later.
A photographic age estimate alone is not a reason to undergo this kind of scan. The decision depends on symptoms, medical history, and a doctor’s recommendations; the limitations of whole-body MRI for preventive screening are discussed separately.
No. A research estimate from FaceAge does not replace your prescribed follow-up care. A reassuring number based on a photograph does not confirm that you are disease-free or cancel the need for examinations.
A research measure based on a photograph should not be used as the sole basis for denying treatment. Even an unfavorable prognosis calls for a discussion of available options, goals of care, and the patient’s own preferences.
How was this article?
Share
