For millennia, healers have read the face as a map of the body's inner state — and now, a team of researchers has trained an artificial intelligence to do the same. Published in the Lancet Digital Health, the tool analyzes a single photograph to assign cancer patients a mortality risk score, drawing on patterns the algorithm identified independently, patterns that echo observations Hippocrates recorded more than two thousand years ago. The work sits at the intersection of ancient intuition and modern computation, raising a question as old as medicine itself: what does a face truly tell us abou
AI tool predicts cancer survival odds from facial photo in breakthrough study
One more piece of information for doctors to consider
Mark
So the AI looks at your face and tells doctors how long you might live. That's quite a claim. How does it actually work?
Mimi
It was trained on forty million face photos to learn age estimation, then retrained on thirty thousand cancer patients' images and medical records. The algorithm learned to spot patterns in facial features that correlate with survival outcomes.
Luke
But we should be clear—correlation isn't causation. The study shows the tool sorts patients into risk groups that have different mortality rates. That's measurable. But we don't actually know why those facial features matter biologically.
Mark
The researchers said it focuses on the nose bridge and under the eyes. Why those spots?
Mimi
They found that Hippocrates described those same features—sunken eyes, hollowed temples, a sharp nose—as signs of impending death. So the AI independently rediscovered something from ancient medicine.
Luke
That's interesting historically, but it doesn't explain the mechanism. And here's the bigger issue: the study didn't compare the AI's predictions to what experienced doctors already judge just by looking at a patient.
Mark
So we don't know if this actually tells doctors something new?
Mimi
Exactly. That's what Miroslaw Bober, the outside expert, flagged. He said the tool needs to be tested against physician judgment and across different hospitals before anyone should use it clinically.
Mark
What does Mak say about that?
Mimi
He frames it as one more piece of information for doctors to consider, not a replacement for their judgment. He's also careful to say it shouldn't be used to deny patients care.
Luke
Which is reassuring, but it doesn't answer whether the tool actually improves outcomes or just makes doctors more confident in decisions they'd make anyway.
Der Puls
- An AI trained on forty million faces and thirty thousand cancer patient records can now estimate, from a single photograph, whether a patient is likely to die within ninety days.
- The algorithm zeroes in on the bridge of the nose and the area beneath the eyes — the same features Hippocrates flagged as signs of approaching death, a convergence that unsettled even the researchers themselves.
- Outside experts warn that predictive accuracy is not the same as clinical utility, and the study never measured what doctors already perceive, leaving the tool's added value unproven.
- The risk of harm is real: a mortality score attached to a face could quietly shape treatment decisions, potentially nudging physicians toward undertreating patients who still have time.
- Researchers frame the tool as a supplementary data point, not a verdict — but the line between a second opinion and a self-fulfilling prophecy in end-of-life care is dangerously thin.
For millennia, healers have read the face as a map of the body's inner state — and now, a team of researchers has trained an artificial intelligence to do the same. Published in the Lancet Digital Health, the tool analyzes a single photograph to assign cancer patients a mortality risk score, drawing on patterns the algorithm identified independently, patterns that echo observations Hippocrates recorded more than two thousand years ago. The work sits at the intersection of ancient intuition and modern computation, raising a question as old as medicine itself: what does a face truly tell us about the life behind it, and who — or what — should be trusted to read it?
A research team has developed an AI system capable of estimating a cancer patient's survival odds from a single facial photograph, publishing their findings in the Lancet Digital Health. The algorithm assigns risk scores, with the highest scores flagging patients most likely to die within ninety days. The scientists describe it as a first of its kind.
The model was built in two stages: first trained on more than forty million photographs to learn age estimation, then retrained on images and medical records from over thirty thousand cancer patients. This layered approach allowed it to move beyond age and begin detecting facial patterns linked to mortality. It performed consistently across ages, races, genders, and cancer types — a breadth the team considers one of its strengths.
Lead author Raymond Mak, a radiation oncologist at Mass General Brigham, noted that while doctors routinely form impressions from a patient's appearance, they are generally poor at predicting survival timelines. When his team examined which features the algorithm weighted most, they found it focused on the nose bridge and the area beneath the eyes. Mak then discovered that Hippocrates had described those same features — sunken eyes, hollowed temples, a sharpened nose — as signs of impending death, suggesting the AI had independently recovered one of medicine's oldest observations.
Not everyone is convinced the tool is ready. Miroslaw Bober of the University of Surrey argued that prediction and understanding are not the same thing, and that the study's most significant gap is its failure to measure physician assessments alongside the AI's scores. Without that comparison, it remains unknown whether the algorithm offers anything beyond what an experienced clinician already perceives. He called for testing across multiple hospitals and direct benchmarking against human judgment.
Mak acknowledged the tool's limits and positioned it modestly — as one more data point in a difficult conversation, not a replacement for clinical wisdom. He suggested it might help address the well-documented tendency to overtreat patients near the end of life. Whether it will genuinely change decisions, or simply give numerical form to what doctors already sense, is a question the research has opened but not yet answered.
A team of researchers has built an artificial intelligence system that can estimate how long a cancer patient might live by analyzing a single photograph of their face. The tool assigns risk scores to patients, with the highest scores indicating those most likely to die within ninety days. The work, published in the Lancet Digital Health on October 19, represents what the scientists describe as a first in the field.
The AI was trained on more than forty million facial photographs to learn how to estimate a person's age, then retrained on images and medical records from over thirty thousand cancer patients, many of whom had already passed away. This dual training allowed the algorithm to move beyond simple age prediction and begin recognizing patterns in facial features that correlate with survival outcomes. When tested, the tool estimated age with a margin of error of roughly five years and successfully sorted patients into higher- and lower-risk groups that showed measurably different mortality rates. Importantly, the model performed consistently across different ages, races, genders, and cancer types.
Raymond Mak, a radiation oncologist at Mass General Brigham Cancer Institute and one of the study's lead authors, explained that while doctors are trained to make judgments about patients based on appearance, they are generally poor at predicting how long someone will live. "We think that using AI, we can capture some of that physiological reserve, risk for mortality, just from a face photo," Mak said. When his team examined which facial features the algorithm relied on most heavily, they found it focused primarily on the bridge of the nose and the area directly beneath the eyes. Mak then discovered that Hippocrates, writing over two thousand years ago, had identified sunken eyes, hollowed temples, and a sharpened nose as early physical signs of approaching death—suggesting the AI had independently rediscovered observations made in ancient medicine.
Yet the research has drawn caution from outside experts. Miroslaw Bober, a professor of video processing and machine learning at the University of Surrey who was not involved in the study, argued that the tool is not yet ready for use in actual clinical settings. He pointed out that being able to predict an outcome is fundamentally different from understanding why that prediction works. More critically, Bober noted that the study did not measure doctors' own assessments of patients' general health and fitness, making it impossible to know whether the AI tool provides information that doctors do not already possess through their own clinical judgment. He called for the algorithm to be tested directly against physician assessments and validated across multiple hospitals and different patient populations.
Mak expressed surprise at how effectively the tool worked for predicting survival and framed its potential role narrowly. He acknowledged that doctors tend to overtreat patients at the end of life and suggested the AI could serve as an additional data point for physicians to consider—not as a replacement for their judgment or as a tool to restrict access to care. The question of whether this new capability will actually change how doctors make decisions, or whether it simply confirms what experienced clinicians already sense, remains open.
Bemerkenswerte Zitate
Doctors are trained to assess patients by how they look, but they are generally not very good at estimating end-of-life expectancy. Using AI, we can capture physiological reserve and risk for mortality from a face photo.— Raymond Mak, radiation oncologist at Mass General Brigham Cancer Institute
Predicting an outcome is not the same as explaining it. The tool has not shown it adds clinically useful information beyond doctors' assessments of a patient's general health and fitness.— Miroslaw Bober, professor of video processing and machine learning at University of Surrey
Möchten Sie die ganze Geschichte?
Das Original lesen bei The Straits Times ↗
Health & Medicine