Harvard University researchers developed an AI tool that decodes facial cues to predict biological age
Researchers at Mass General Brigham and Harvard developed "FaceAge," an AI tool that analyzes facial photos to predict biological age and cancer survival. Trained on thousands of images, it found cancer patients often appear biologically older, correlating with worse outcomes. This tool aims to help doctors personalize treatments by offering a more nuanced health assessment than chronological age alone.
Raymond Mak (left) and Hugo Aerts. (Image source: The Harvard Gazette)
AI Tool Gauges Health Through Facial Cues: In a striking convergence of artificial intelligence and medical diagnostics, researchers from Mass General Brigham (MGB) and Harvard Medical School (HMS) have pioneered a sophisticated tool, dubbed "FaceAge," capable of discerning an individual's biological age and forecasting cancer survival times merely by analyzing a facial photograph. This innovative approach promises to arm clinicians with unprecedented insights, potentially revolutionizing how patient treatments are personalized and prognoses are assessed.
The fundamental premise, articulated by Dr. Hugo Aerts, co-senior author of the study and Director of MGB’s Artificial Intelligence in Medicine program, challenges the conventional reliance on chronological age. "We all recognize the truism that aging is a deeply individual process," Dr. Aerts, also a professor of radiation oncology at HMS, explained. “A person's date of birth dictates their chronological age, but this often diverges significantly from their biological age – the true barometer of their physiological health and life expectancy. Factors such as lifestyle choices, genetic predispositions, and underlying health conditions all contribute to this biological clock. Our central hypothesis was that a person's facial appearance might serve as a tangible reflection of this internal, biological age.”
This intriguing idea spurred the development of FaceAge. Under the guidance of scientists at MGB’s Artificial Intelligence in Medicine Program, the deep-learning algorithm underwent rigorous training. Its digital mind was fed a massive dataset: over 58,000 facial photographs of healthy individuals with confirmed ages, and a further 6,000 images of cancer patients for whom both age and detailed clinical outcomes were known.
The results gleaned from this extensive training were compelling. The algorithm consistently estimated that, on average, cancer patients presented a FaceAge approximately five years older than their actual chronological age. More significantly, the research established a strong correlation: an older-appearing FaceAge was linked with poorer prognoses across various types of cancer.
Dr. Aerts highlighted that assessing a patient's health based on visual cues is an established, albeit informal, practice in medicine. "Physicians routinely perform what we might call an 'eyeball test' upon entering a patient's room," he noted. “This encompasses an array of observations: Is the patient using a wheelchair? How robust or frail do they appear? Are there overt signs of illness?”
However, the study, published in the esteemed journal The Lancet Digital Health in early May and supported by funding from the National Institutes of Health, also exposed the limitations of this human-led visual assessment. When ten clinicians and researchers were tasked with predicting short-term life expectancy using photographs of 100 terminal patients undergoing palliative radiation therapy, their accuracy was only marginally better than random chance. This held true even when they were privy to crucial information like the patient's chronological age and cancer status. Notably, their predictive accuracy improved when FaceAge data was incorporated into their assessments.
Dr. Raymond Mak, a faculty member at the Artificial Intelligence in Medicine Program, HMS associate professor of radiology oncology, and co-senior author, emphasized the clinical utility of this enhanced understanding. Knowing a patient's biological age and their likely remaining lifespan allows oncologists to fine-tune treatment strategies with greater precision.
He recounted the case of an 86-year-old lung cancer patient who, despite his chronological age, appeared markedly younger. This apparent vitality influenced Dr. Mak's recommendation for a more aggressive treatment regimen. Today, at 90, the patient continues to thrive. Retrospective analysis using FaceAge on a photo taken at the time of treatment remarkably assessed his biological age as ten years younger than his chronological age.
Conversely, Dr. Mak pointed out, patients who appear frailer than their years might suggest may benefit from less intensive treatments, aligning medical intervention with their body's actual capacity to tolerate it. "We hypothesize that FaceAge could serve as an invaluable biomarker in cancer care," Dr. Mak stated, “quantifying a patient’s biological age and empowering doctors in these challenging decision-making processes.”
The efficacy of FaceAge has been demonstrated across multiple cancer types, and its creators are now actively exploring its potential to predict outcomes in other diseases. The algorithm's power stems from deep learning, a process where it autonomously learns and refines its predictive capabilities as it is exposed to vast quantities of data with known outcomes.
Interestingly, Dr. Aerts conceded that researchers do not yet fully comprehend the specific facial cues that FaceAge prioritizes in its analysis. "It's highly probable," he mused, “that the algorithm is detecting subtleties different from what a human physician might focus on, such as wrinkles, hair graying, or baldness. If this is indeed the case, its utility is amplified, as it introduces a novel, complementary perspective to the physician's existing analytical framework.”
Both Dr. Aerts and Dr. Mak stressed that FaceAge is envisioned not as a standalone diagnostic oracle but as a sophisticated tool within the physician's arsenal. Its role would be to augment, not replace, clinical judgment. Beyond aiding in initial treatment planning, it could also serve as a dynamic monitoring system, tracking changes in a patient's appearance over time and alerting medical staff if an individual appears to be deteriorating.
However, before FaceAge can be integrated into routine clinical practice, it requires further validation through extensive testing across diverse patient populations to ensure its accuracy and reliability are maintained across different ethnicities, genders, and age groups.
"The potential clinical impact is truly substantial," Dr. Aerts concluded. "We are on the cusp of having a non-invasive, easily deployable method to continuously monitor a patient’s health status – before, during, and after treatment. This could dramatically enhance our ability to predict the risk of complications following major surgeries or other intensive therapies, ultimately leading to more tailored and effective patient care." The algorithmic gaze of FaceAge may soon offer a clearer window into our inner health, changing the face of medicine as we know it.
