According to Bloomberg, researchers have developed an artificial intelligence tool capable of predicting the survival probability of cancer patients using only a single facial photograph. This remarkable tool can supplement the clinical judgment and experience of physicians. It promises to provide valuable guidance for formulating treatment plans, proving especially beneficial in scenarios where medical data is profoundly scarce.
Training the Predictive Algorithm
This novel model underwent rigorous training to estimate human age and predict the survival outcomes of oncology patients. The research team trained the model using a massive dataset comprising over 40 million facial images, alongside photographs of more than 30,000 cancer patients, many of whom have since passed away. According to the findings published in The Lancet Digital Health, the tool generates a specialized risk score for each patient. Consequently, individuals with the highest scores face a significantly greater likelihood of mortality within 90 days.
Scientists state that such groundbreaking research is the first of its kind. Raymond Mak, a radiation oncologist at the Mass General Brigham Cancer Center and co-author of the paper, noted that while doctors learn to assess a patient’s physical condition through appearance during training, they usually struggle to accurately estimate the remaining life expectancy of terminally ill patients.
“We believe that through artificial intelligence, capturing a glimpse of the body’s physiological reserve and mortality risk is possible from just a single facial photograph,” Mak explained during an interview.
Accuracy and Clinical Implementation
This AI tool estimates age with a margin of error of approximately five years and reliably categorizes patients into high-risk and low-risk groups. This reveals a stark contrast in mortality risk between the two cohorts. Furthermore, the model remains effective across diverse demographics, encompassing variations in age, race, gender, and differing types of cancer.
Increasingly, AI tools are undergoing trials in the realm of cancer detection and treatment. Some assist in diagnosis by analyzing medical imagery, while others analyze genetics and lifestyle risk factors to forecast an individual’s susceptibility to disease. While previous reports have shown how AI eroded doctors’ ability to spot cancer within months in a recent study, this new facial analysis model seeks to complement human clinical intuition rather than replace it.
Although the experimental results appear highly promising, Miroslaw Bober, a professor of video processing and machine learning at the University of Surrey, argues that the tool is not yet ready for clinical deployment. Bober was not involved in this specific study.
“Predicting a result does not equate to explaining its underlying causes,” Bober asserted. “This study has not demonstrated that these predictions offer additional, clinically valuable information beyond a physician’s holistic assessment of a patient’s overall health. Furthermore, the research itself did not measure this type of physician assessment.” He added that the tool requires further rigorous testing across various hospitals and diverse patient populations, heavily relying on the subjective judgments of medical professionals as a benchmark.
Historical Context and Future Use
Mak, who developed the tool alongside his team, expressed his profound surprise at the model’s exceptional predictive accuracy regarding life expectancy. The team discovered that the algorithm primarily focuses on the bridge of the nose and the region beneath the eyes during its evaluation. Upon reviewing historical medical literature, Mak found that Hippocrates first described these exact features: sunken eyes, collapsed temples, and a sharpened nasal tip serve as early external signs that life is drawing to a close.
Mak emphasized that the tool will never be used to deny or limit a patient’s access to vital medical treatment. “We recognize that when facing terminally ill patients, doctors sometimes resort to overtreatment,” he reflected. In his perspective, this innovation simply provides “an additional layer of reference information for physicians.”
