Mass General Brigham team publishes FAHR-Face study in The Lancet Digital Health, using facial photos to predict mortality risk in cancer patients

36Kr Auto Digital Kr Future Consumption Intelligent Emergence Future City Launch Power on 36Kr Global 36Kr Research Institute Tide 36Kr Enterprise Service Review 36Kr Finance Workplace bonus 36 Carbon Post-Wave Research Institute Dark Surge Waves Hard Kr Kr Rui Research Institute Media Brands

Enterprise Accounts Enterprise Service Review 36Kr Research Institute 36Kr Innovation Consulting Enterprise Services

Core Services City Window Government Services

Venture Capital Release LP Source Plan VClub VClub Investment Institution Database Investment Institution Job Referrals Investor Certification Investor Services

Event Search Project Recommendations I Want to Join City Cooperation Judging Life and Death by Facial Features—AI Actually Did It? Letter AI · October 21, 2026 13:11 Of the 122 extremely high-risk patients rated by the model, 64 died within 90 days.

On October 20, the international medical journal The Lancet Digital Health published a study: AI can identify differences in mortality risk among cancer patients using only an ordinary facial photo.

Mass General Brigham in the United States developed an AI model called FAHR-Face, tasking it with finding clues related to patient survival in facial photos.

The researchers recruited 3,820 cancer patients who had not participated in model training for testing. Among them, 122 were rated by the AI as extremely high risk, and within 90 days after the photos were taken, 64 of these 122 people died.

The researchers then used data from nearly 4,900 Dutch cancer patients for external validation and observed similar results.

Throughout the prediction process, the AI does not need to review medical records, analyze CT images, or reference blood test results. It only needs to look at a face.

01

To achieve this, the research team first trained a visual foundation model, FAHR-Face, using more than 40 million photos of ordinary people's faces.

The training method is somewhat like a jigsaw puzzle: the researchers randomly masked about 75% of each photo, then asked the AI to try to reconstruct the complete photo from the remaining parts. Through repeated training, the AI gradually learned to recognize the structure of faces and the connections between different facial features.

With this foundation, the researchers began teaching it to predict mortality risk. They collected 34,363 facial photos of cancer patients, along with corresponding survival records, including whether the patient had died, when they died, or how long they had survived as of the last follow-up.

Next, the researchers used a ranking learning method to have the AI compare the survival times of different patients. For example, after two patients had their photos taken, one died within three months while the other survived longer. The model needed to try to assign a higher mortality risk score to the former.

Ultimately, the researchers converted the model's output scores to a range between 0 and 1. The higher the score, the higher the AI considered the patient's mortality risk.

To verify whether the model's scores were accurate, the researchers recruited another 3,820 cancer patients who had not participated in training.

As a result, in addition to the 122 extremely high-risk patients mentioned at the beginning, the AI also gave risk scores between 0 and 0.1 to another 975 patients. Among these 975 people, only 20 died within 90 days after the photos were taken, a proportion of just 2.1%.

By contrast, the 90-day mortality rate among the extremely high-risk patients reached 52.5%, a difference of more than 25 times between the two ends. The high and low risk scores given by the model did indeed correspond to vastly different mortality rates.

Even after incorporating factors such as age, sex, cancer type, and treatment status into the statistical analysis, the mortality risk of the highest-risk group was still 3.23 times that of the lowest-risk group.

However, there was still a question at this point: this model performed well in American hospitals, but could similar results be obtained in another hospital with a different group of patients?

The researchers then recruited 4,892 patients from the Maastro Radiotherapy Center in the Netherlands and tested them using the same model.

From the test data at the American hospital to an independent patient cohort at another hospital in the Netherlands, this model was able to identify clear differences in mortality risk among cancer patients.

In fact, FAHR-Face is not limited to the mortality risk prediction function mentioned earlier. The research team also developed another model called FAHR-FaceAge, specifically designed to estimate a person's facial age.

The idea behind this model is relatively easy to understand: at the same age of 60, some people look only slightly over 50, while others appear much older. The researchers believe that how much older a person looks than their actual age may also reflect their health status.

They trained the model using hundreds of thousands of facial photos labeled with age, then calculated the gap between the AI-estimated facial age and actual age. The results showed that among cancer patients, those who looked older than their actual age often also faced a higher mortality risk.

For example, after accounting for factors such as age, sex, and cancer type, patients whose AI-estimated facial age was at least 5 years older than their actual age had a mortality risk about 21% higher than other patients.

But the research team did not stop there.

The results showed that the two focused on noticeably different areas.

The researchers also used statistical methods to compare the two models. Among the 4,892 patients in the Netherlands, the correlation coefficient between the gap between facial age and actual age and the mortality risk score given by the AI was only 0.16, meaning the association between the two was very weak.

Moreover, when the researchers used the two indicators together to predict patients' survival risk, both retained independent predictive effects, and the performance of the combined statistical model also improved.

But then the question arises: what exactly is this information?

In the paper, the research team proposed that features below the eyes and near the nose may be related to the facial changes in seriously ill patients that doctors have long observed. For example, as a patient's body gradually weakens, facial fat, muscles, and skin condition may all change, and these changes may leave traces in photographs.

Miroslaw Bober, a professor of machine learning at the University of Surrey in the UK, also pointed out this issue in an expert commentary released by the UK Science Media Centre. He believes that the study does demonstrate a statistical link between patients' facial features and survival risk, but being able to predict an outcome does not mean the cause has been figured out.

Speaking of this, it brings to mind the traditional Chinese medicine practice of "observation, listening, questioning, and pulse-taking." Among these, "observation" means looking at a person's complexion, demeanor, and physique to look for clues related to health status. In modern medicine, doctors likewise pay attention to whether patients show outward changes such as weight loss or a haggard appearance.

Interestingly, the paper for this study even specifically cites medical literature on "Hippocratic facies"—"Hippocratic facies" refers to the special facial appearance that may appear in some critically ill patients, especially those near death: sunken eye sockets, hollow temples, a sharpened nose, and a pale or ashen complexion. More than two thousand years ago, Hippocrates recorded these features and used them as clues for judging whether a patient's condition was critical.

But whether it has discovered new clues that are difficult for the human eye to detect, or has simply turned bodily changes that doctors could already observe into numbers, there is still no answer.

03

How much medical information is still hidden in a face?

Interestingly, some of the photos used by the researchers were originally just ordinary photos taken by hospitals in routine practice to verify patient identity.

And this was not their first time studying facial photos.

In May 2025, the team published the initial FaceAge study in The Lancet Digital Health. At that time, they also tried to have AI assist doctors in predicting the survival period of patients with advanced cancer.

The researchers recruited 100 patients receiving palliative care and asked doctors and researchers to judge whether they would still be alive six months later. After adding the information provided by FaceAge, the predictions improved somewhat: one indicator used to measure predictive ability, AUC, rose from 0.74 to 0.80.

By the time of this FAHR-Face study, the researchers had already begun trying to predict mortality risk directly from photos.

The most appealing aspect of this line of research is that it is simply too convenient.

CT and MRI require specialized equipment, and pathological examinations require collecting tissue samples, while facial photos require only an ordinary camera. For doctors, this means that in the future there may be an additional low-cost, easy-to-operate auxiliary assessment method that does not increase the physical burden on patients.

The study leader, radiation oncologist Raymond Mak, mentioned that when doctors decide whether to arrange surgery, radiotherapy, or other intensive treatment for patients, they need to assess the patients' physical condition and whether they can withstand the treatment. He hopes that in the future AI can make such judgments more objective and consistent.

However, there are still many issues to resolve before this goal can truly be achieved.

First, although ordinary photos are easy to obtain, they are also easily affected by shooting conditions. Miroslaw Bober, a professor at the University of Surrey in the UK, pointed out that simply having a person smile, or changing the lighting when taking a photo, can cause the AI-estimated facial age to change by about two years on average, and sometimes the difference can be even greater.

Although this does not mean the mortality risk model will be affected in the same way, if facial photos are to become a reliable medical indicator, it must first be clarified how much interference these shooting conditions actually cause.

In addition, the photos of cancer patients currently used in the study are still predominantly of white people. Whether the model can maintain the same predictive performance for other populations still requires more data validation.

More importantly, predicting mortality risk and actually helping doctors treat patients are still two different things. FaceAge in 2025 has already shown potential to assist doctors in making judgments, but this new model has not yet fully demonstrated how much additional help it can provide on top of doctors' existing clinical assessments.

It can divide patients into different groups according to mortality risk, but it has not yet proven whether these predictions can help doctors choose more appropriate treatment plans. Therefore, FAHR-Face cannot yet be used directly in clinical practice.

By the way, Mass General Brigham is not the only one trying to find disease clues in ordinary facial photos.

As early as 2020, researchers from institutions including Fuwai Hospital of the Chinese Academy of Medical Sciences and Tsinghua University published a study in the European Heart Journal, attempting to have AI identify coronary heart disease through facial photos.

In tests with 1,013 patients, the model achieved an AUC of 0.73, showing a certain ability to identify the disease.

In February this year, a research team from institutions including the University of Copenhagen in Denmark turned its attention to the emergency room. They collected mobile phone photos of 27,660 emergency patients and had AI predict these patients' mortality risk.

The results found that, based on facial photos alone, AI could identify some high-risk patients that traditional emergency triage systems easily overlook. Some people originally considered to have milder conditions were rated by the AI as having a higher mortality risk, and they were indeed more likely to die later.

However, this study is currently only a preprint and has not yet undergone formal peer review.

From cancer to coronary heart disease to emergency patients, multiple studies have already demonstrated that an ordinary facial photograph does indeed contain information related to a patient's health, and that this information may be richer than we previously imagined.

As for what role this information can play, researchers may still need to keep searching for the answer.

This article comes from the WeChat public account "Letter AI"; author: Yuan Xinyue; editor: Wang Jing; published by 36Kr with authorization.

The views in this article represent only the author personally; the 36Kr platform provides only information storage space services.+110

Good articles need your encouragement. Letter AI, guest author, 2, Favorites+10, Comments. Open WeChat "Scan" to open the webpage, then tap the share button in the upper right corner. Weibo, immersive reading, back to top, report, join the comments. Comments number in the millions, friendliness first. Log in to join the discussion. Submit comment 0/1000. You may also like these articles. Harvard's major study: AI can predict the mortality risk of cancer patients from facial photos. Managing corporate AI bills and raising $60 million: who is North.cloud eyeing? Will AI be humanity's final invention? Tencent leads, Baidu second; the AI office market is far from settled. Li Zexiang's team made an e-ink screen Agent that sticks to the back of a phone, backed by Vertex, 5Y Capital, Source Code, and others | Hard Kr exclusive. Anthropic's latest prospectus disclosure: revenue of 16.3 billion, book losses as high as 80.7 billion US dollars. Codex team in-depth retrospective: how was the top coding stream born? The AI employee bubble bursts; Meta urgently halts layoff plan; Zuckerberg: humans are still easier to use. Terence Tao and Deng Yu's latest conversation calls for "slow research" in mathematics. Letter AI guest author.

Focusing on frontier technology, see the future first. Published 505 articles.

Recent content: Reading fate from faces, AI actually did it? 1 hour ago. Yesterday product managers were still tearing into each other; today all have been forced by REA into comrades-in-arms. 23 hours ago. Letting "foreign Agents" browse the Chinese internet, I became their assistant. 2026-10-19. Read more content, poke here hard. Choose reliable AI, see real reviews. View the 36Kr AI Review official communication community. Join consultation project review and onboarding. Contact 36Kr project recommendations. Subscribe to the account and follow.

36Kr Europe Station 36Kr Europe Station 36Kr Europe Station AI Product Daily Online Rumor Information Reporting Portal Hot Recommendations

Push and interpret cutting-edge, substantive technology venture capital information.

Primary market financial information and system service provider.

Focus on outstanding global entrepreneurs, project financing rate close to 97%, leading the industry.