Massachusetts General Hospital team develops the FAHR-Face model, predicting survival risk of cancer patients from facial photos alone

AI "physiognomy" calculates a death list, and half of it comes true? Tech Fox November 01, 2026 09:00:39

"I see your Yintang is dark, I fear you may suffer a bloody calamity."

In the past, when Fox Sister heard this, her first reaction was: where did this master come from?

And it doesn't talk to you about wealth or marriage; it goes straight to maximum difficulty: using only a face, predict the mortality risk of cancer patients.

On October 19, the top medical journal The Lancet Digital Health published a study, and after reading it, Fox Sister felt a chill down her spine.

The team at Mass General Brigham in the United States developed an AI foundation model called FAHR-Face, and on that basis trained a tool that directly predicts the survival risk of cancer patients.

How outrageous were the results?

Among the 10% of patients judged by the AI as "highest risk," half (52.5%) did not survive 90 days; among the 10% judged "lowest risk," only 2.1% died within 90 days.

Can a face really reveal a person's book of life and death? Is this a medical miracle, or cyber mysticism?

Why must "face reading" be used to determine life and death?

This study was jointly completed by institutions including Harvard Medical School and Mass General Brigham.

The team lineup is quite cross-disciplinary: on one side is radiation oncology physician Raymond H. Mak, and on the other is Hugo J. W. L. Aerts, head of the medical AI program.

When these two met, they set their sights on an extremely critical problem: how do doctors determine whether a cancer patient can withstand treatment?

You should know that treating cancer cannot rely solely on tumor type and stage. The patient's physical foundation and degree of frailty directly determine whether radiotherapy, surgery, or systemic therapy can be used.

But the problem is that this "physical state" often can only be estimated through a doctor's subjective experience, and is hard to quantify precisely.

So the team set its sights on a zero-cost data source that everyone has but is often overlooked: the face.

They built the foundation model FAHR-Face, whose core logic is: first let the AI learn to "read faces," then teach it to do different "specialty tasks."

To lay this foundation, the team used a vision Transformer architecture and adopted a hardcore training method called "masked autoencoder."

During training, the system randomly masks about 75% of the area in a photo, forcing the AI to "fill in" and reconstruct the entire face based on the remaining information.

Once this "super base" was trained, they fine-tuned two specific tools:

FAHR-FaceAge: estimates age from the face, studying the relationship between facial aging and survival outcomes.

FAHR-FaceSurvival: directly looks at cancer patients' faces to predict survival risk.

They developed FAHR-FaceAge and trained it on more than 740,000 photos with real age labels. In public tests, the age it calculated differed from actual age by only 5.1 years on average.

But merely calculating age is nothing remarkable.

What the team really wanted to understand was: do facial aging features also hide codes related to health, or even life and death?

The team released its second big move: FAHR-FaceSurvival. They directly used 34,363 photos of cancer patients, combined with the patients' real survival outcomes, for training.

So where on the face is this pattern hidden?

The results found that the two models focus on completely different places:

FAHR-FaceAge (calculating age): focuses on the forehead, temples, and nasolabial folds. This fits human intuition very well, as these areas are naturally where wrinkles form and age shows.

FAHR-FaceSurvival (reading life and death): its gaze actually moved downward, locking firmly onto the under-eye area and the midface around the nose bridge.

This is quite interesting. Why does reading life and death require staring at the under-eye and midface areas?

Although heatmaps cannot directly tell us "why," combined with medical common sense, this makes sense.

In medicine, malignant tumors are extremely taxing on the body, and patients can easily show extremely subtle facial fat loss, sunken eye sockets, and muscle atrophy.

This is actually a manifestation of declining bodily reserve function and overall frailty.

But in daily consultations, this "slight facial structural collapse" is hard for doctors to quantify at a glance with the naked eye.

The research team also conducted independent validation. After removing all confounding factors that doctors already know, such as age, sex, and cancer stage, the results remained astonishing.

Among patients the AI predicted to be at highest risk, the mortality risk was still 3.23 times that of the lowest-risk quarter!

It is evident that an ordinary facial photo contains far more information than what we can see with the naked eye.

How far is it from truly entering hospitals?

Writing this, Fox Sister is also fired up, wanting to know when hospitals will actually be able to use it?

Obviously, the research team also does not want this technology to remain only in papers.

At present, they have launched the FaceAge Research Study public project (the website below), where ordinary people can also upload photos to obtain their own research results and incidentally help the team collect more diverse data.

The existing model may be very accurate in certain specific populations, but change the skin color or age group, and the effect may not be the same.

Only by feeding it sufficiently mixed and broad data can it possibly become a truly general-purpose tool.

From a clinical perspective, if this really works, the imaginative possibilities are enormous.

For example, whether elderly cancer patients can withstand intensive treatment has always been a headache for doctors in assessment.

If a facial photo can provide additional risk clues, it can be said that to some extent this spares elderly patients from "suffering," and this is also the original intention of this research.

Medical AI expert Professor Hugo Aerts sees further: in the future, this technology may not only be used for cancer.

After all, a photo does not require expensive equipment; a mobile phone is enough to take one.

In the future, in primary care clinics, home elderly care, and even telemedicine, taking a face photo regularly may help medical staff spot abnormal signals worth vigilance.

For chemotherapy patients, continuous facial records can also allow doctors to dynamically observe changes in their physical condition.

However, from "can predict" to "truly useful," there are still several mountains in between.

In addition, besides giving a score, can it provide extra value beyond the doctor's existing diagnosis? These all require larger-scale, cross-hospital real-world testing.

If it merely reminds doctors to pay more attention and intervene earlier, then it is definitely a good tool.

But if an incompletely mature risk score instead becomes a reason to "sentence patients to death" and discourage treatment, then it completely deviates from the original intention of technical assistance.

This is undoubtedly a beginning full of imagination. But the more powerful the technology, the clearer its boundaries must be.

It can become a sharp tool to help doctors see more information, but it can never replace doctors and become the final arbiter of fate.

After all, the underlying color of medicine is, in the end, human warmth.