Cancer Prognosis From a Selfie: AI Reads Biological Age in the Face

An artificial intelligence tool that predicts a cancer patient's chances of survival from a single facial photo shows just how cheap prognosis in medicine could become. But the same technology also heralds an era in which a selfie could be turned into a health score. That's why the real question isn't the model's accuracy. The real question is what the model is looking at, and who gets to use the result.

What happened?

Researchers at Mass General Brigham, an academic health system in Massachusetts, have developed an AI system that estimates a person's biological age from a photo of their face. In patients diagnosed with cancer, the system links this estimate directly to survival outcomes. The idea is simple: a patient whose face looks older than their chronological age likely has a worse prognosis.

Why does it matter?

Doctors already do this by intuition. In oncology, a patient's overall condition is assessed using performance scores such as ECOG, which rely largely on observation. This assessment is one of the main factors determining how aggressive treatment will be. But it is a subjective measure. Two oncologists can give the same patient different scores. Estimating biological age from the face aims to turn that intuition into a reproducible number.

This approach isn't alone, either. In recent years, numerous models have been published that derive "biological age" from retinal photos, ECGs and chest X-rays. The common logic is to extract additional information from data that is already routinely collected. What sets the facial photo apart is cost. It requires no special device; a phone camera is enough. In resource-limited clinics, that could be a serious advantage.

The problem begins with the "black box" issue. What is the model actually looking at? If it is picking up clinical cues such as hollowing at the temples, skin turgor or weight loss, it is a useful tool that standardizes the physician's eye. But if it latches onto misleading signals such as lighting, camera quality, makeup or skin tone, the results will only look good on the dataset it was trained on. There are precedents for this in medical imaging: some models have been shown to learn traces of the hospital where an X-ray was taken rather than the disease itself.

Is it reliable across ethnic groups and in Turkey?

Facial analysis technologies do not have a good track record on bias. The Gender Shades study revealed that commercial facial recognition systems made markedly more errors on darker-skinned women. During the pandemic, it was also debated that pulse oximeters could overestimate oxygen levels in people with darker skin. A measurement error tied to skin tone in a prognostic tool translates directly into treatment decisions.

From Turkey's perspective, the picture looks like this: such models are typically developed using patient data from a single health system in the US. Patient profiles, aging patterns, sun exposure and smoking rates in Turkey are different. Assuming the tool will perform equally well on patients here without a local validation study would be a mistake.

Who does it affect, and how?

The most troubling scenario lies outside the clinic. Everyone has selfies, and they are easy to collect. If an insurer or employer could extract a "biological age" score from an application photo, a person would be profiled without their knowledge. Many countries have legal protections against discrimination based on genetic data. Health predictions derived from the face, however, remain in a gray area.

In Europe, the AI Act classifies biometric categorization and medical AI as high-risk areas. In Turkey, the KVKK (Personal Data Protection Law) protects health and biometric data as special categories of personal data. But how a "health inference" derived from an ordinary photo would fall under this scope has yet to be clarified.

What's next?

What to watch for is clear: independent validation of the model across different countries and ethnic groups, studies explaining its decision-making mechanism, and prospective clinical trials. Most importantly, showing whether the tool actually improves treatment decisions. A statistical association does not mean it benefits patients.

My assessment: estimating biological age from the face could be a reasonable aid that makes the physician's eye measurable. But letting it move beyond the clinic before its validation and limits of use are clear would cause far more harm than good.