Could your face reveal health clues doctors cannot easily see? A new AI model trained on millions of facial photographs links apparent age and subtle facial features to survival outcomes in cancer patients, opening the door to a potential new way of assessing prognosis.
Study: Foundation artificial intelligence models for health recognition by use of face photographs (FAHR-Face): a model development and validation study. Image credit: wedmoments.stock/Shutterstock.com
A study published in The Lancet Digital Health describes FAHR-Face, an artificial intelligence (AI) foundation model designed to extract prognostic information regarding cancer survival from facial photographs. The findings suggest that facial photographs may provide noninvasive prognostic information, although they require validation before clinical translation.
Can facial photographs reveal more than biological age?
Facial appearance reflects a combination of genetic, environmental, lifestyle, and illness-related factors. Therefore, it may mirror health and aging. Facial photographs can be collected noninvasively and cheaply. However, visual assessments of facial features are subjective, and facial complexity makes reliable assessment difficult. AI models could potentially identify facial features associated with physiological aging or illness and convert them into objective measures.
Previous facial-analysis models have estimated chronological or biological age, but their performance has been limited by training data and generalizability. Predicting survival from facial appearance presents a separate challenge: a person’s apparent age and their risk of death may overlap, but they are not necessarily equivalent.
The authors previously developed and tested FaceAge, a deep-learning model that estimates biological age from facial photographs. They found that a higher facial age was associated with poorer survival in cancer patients. However, they developed it in a mostly White cohort of patients aged 60 years or older.
In the current study, the researchers developed Foundation Artificial Intelligence for Health Recognition-Face (FAHR-Face), a deep-learning model trained on more than 40 million consumer-grade digital facial photographs.
Development of FAHR-Face
FAHR-Face was pretrained using a self-supervised learning approach. Rather than requiring a health label for every image, this approach enables the model to learn visual representations from large amounts of image data.
The researchers used masked autoencoders, an approach that learns from images by randomly masking and reconstructing patches within images. This self-supervised process helps the model learn broader patterns and relationships between facial features. In facial images, multiple structures are highly correlated, providing an ideal setting for masked autoencoders to capture higher-level correlates of health.
The model was then fine-tuned to develop two distinct models for two tasks: estimating biological age (FAHR-FaceAge) and directly predicting survival risk in patients with cancer (FAHR-FaceSurvival).
FAHR-FaceAge
This tool estimates age based on facial appearance. The foundation model was fine-tuned using 743,152 publicly available facial images, with a two-stage approach designed to balance representation across age groups.
The authors compared the model with benchmark models and evaluated its ability to maintain accuracy across image variations. The model was tested in two independent clinical cohorts. The first included 38,183 patients treated at Harvard-affiliated radiotherapy centers in the United States, a small proportion of whom had benign conditions. The second comprised 4,892 patients with cancer who received radiotherapy in the Netherlands.
FAHR-FaceSurvival
This model predicts mortality risk directly from facial photographs rather than from the estimated age alone. It was fine-tuned using 34,363 photographs of patients with cancer, a subset of one of the clinical datasets that were used to test FAHR-FaceAge. The remaining 3,820 patients were used to evaluate the final model, along with another independent patient dataset used to test FAHR-FaceAge.
The researchers examined the performance of the models across demographic and cancer subgroups. They also tested the robustness of age estimation to factors such as makeup, lighting, facial pose, and cosmetic surgery. Saliency mapping was used to map the facial regions that contributed to the age and survival predictions.
Facial age differences reveal links to mortality and lifestyle
FAHR-FaceAge achieved a mean absolute error of 5.1 years when estimating age in public image datasets. This was better than the benchmark models evaluated by the researchers. Its performance did not vary substantially across the human lifespan. The difference between estimated facial age and chronological age has been shown to be associated with survival outcomes. Survival analyses used multivariable Cox models adjusted for clinical prognostic factors.
The median FAHR-FaceAge deviation (FAD), calculated by subtracting chronological age from estimated facial age, was 1.1 years. Patients were categorized into groups ranging from those appearing more than 10 years younger to those appearing more than 20 years older than their chronological age. Patients whose faces looked older than their chronological age had poorer survival.
FAD was also associated with early mortality, with 21.3% mortality in the highest FAD group (>20 years) at 90 days, compared to 6.4% in the lowest group (-10 years or less). It outperformed FaceAge in patients aged 60 years or more and in chronological age-based estimates across the first two years post-assessment (3, 6, 12, and 24 months).
The quartile with the highest risk had over threefold the mortality hazard of the lowest-risk quartile. After adjustment for clinical factors, patients with an FAD of at least 5 years had a 21% higher mortality hazard than those with a lower FAD. Conversely, an FAD of -5 years or less was associated with a 17% lower mortality hazard compared with higher values.
The researchers also found associations between lifestyle factors and estimated facial age. Smoking showed a particularly strong relationship with older apparent age, regardless of whether the cancer was associated with smoking. This association was dose-dependent, both when comparing current, former, and never smokers, and when comparing lifetime exposure to smoking (never, light, or moderate-to-heavy smokers).
Heavy alcohol and drug use were also associated with a higher FAD. However, this does not mean that the model can detect such behaviors at the individual level.
The results suggested that facial features may capture aspects of physiological aging not fully reflected by chronological age alone.
Facial AI scores reveal striking differences in cancer survival
Higher FAHR-FaceSurvival scores were associated with poorer survival among patients with cancer. After adjusting for clinical prognostic factors, every 0.1-point increase in the score was associated with a 17% higher mortality hazard.
The differences were particularly striking for early mortality. Within 90 days, 52.5% of patients in the highest-risk score group had died, compared with 2.1% in the lowest-risk group. At five years, estimated survival was 17.7% and 78.1%, respectively.
Facial age deviation (FAD) also provided prognostic information beyond chronological age, with greater deviations associated with higher mortality risk. Both models showed associations with survival across multiple patient groups, including in the independent Dutch cohort, although their performance varied between subgroups.
The two models appeared to capture different aspects of facial appearance. FAHR-FaceSurvival focused more on regions beneath the eyes and around the nasal bridge, while FAHR-FaceAge emphasized the nasolabial folds, forehead, and temples. These differences suggest that the models provide complementary prognostic information, although the underlying biological mechanisms remain unclear.
Why facial AI needs further testing before clinical use
The models still show several issues before routine clinical use. The survival models were developed and validated in cancer populations from developed countries, limiting generalizability to healthy people or patients with other diseases, or from other regions. The study was also retrospective, and its clinical cohorts consisted of patients receiving radiotherapy, which may limit applicability to broader cancer populations.
Residual confounding may have persisted: Facial appearance may reflect age, illness severity, nutritional status, frailty, and other factors, but the study does not establish which of them contribute to the model’s performance.
The accuracy of the models across racial and ethnic groups needs to be rigorously established in the interests of equity and generalizability, especially since FAD-based survival stratification was weaker in Black and Hispanic patients than in White and Asian patients.
Prospective studies must validate the clinical utility of these models, but they add useful information beyond current assessments, and their predictions are actionable, resulting in better patient care. Privacy and consent issues must also be addressed when using facial photographs, which are identifying biometric data.
Facial photographs emerge as potential cancer prognostic tools
FAHR-Face uses a foundation model trained on more than 40 million facial images to estimate biological age and predict mortality risk in patients with cancer. Its two derived models produced distinct but complementary prognostic information. Patients in the highest survival-risk quartile had more than three times the mortality hazard of those in the lowest quartile.
The findings highlight the potential of facial photographs as a source of health-related biomarkers. However, they do not yet establish a clinically validated method for directing treatment or predicting an individual’s lifespan. Independent, prospective validation across diverse populations and clinical settings will be important before these tools can be incorporated into routine care.
“Integration of this model into clinical workflows could enhance prognostic accuracy and clinical decision making, particularly in settings where traditional biomarkers or comprehensive assessments are scarce.”