A simple photo of a face can reveal more than appearance. This study shows how tracking subtle changes in facial aging over time could help predict survival and reshape cancer care.
Study: Facial aging rate quantifies change in biological age to predict cancer outcomes. Image credit: hedgehog94/Shutterstock.com
A study published in Nature communications examines the predictive ability of photo-based facial aging rate (FAR) for overall survival in cancer patients.
Artificial intelligence-derived facial age as a measurable biological marker
Rates of biological aging vary significantly between individuals and can influence cancer outcome regardless of chronological age. However, their clinical use remains limited due to the lack of practical, non-invasive biomarkers that can be easily implemented in routine care.
FaceAge is an AI-based tool that calculates biological age from facial features such as skin texture, volume loss and structural changes. Previous studies have shown that cancer patients predicted to be older than their chronological age have worse survival outcomes, supporting its potential as a prognostic biomarker.
Using Face Age to measure the rate of aging
The authors previously developed a model called Foundation Artificial Intelligence Models for Health Recognition (FAHR-FaceAge), which was trained to recognize signs of poor health in more than 40 million face images. When used with Face Age, they found that patients whose predicted age was five or more years older than their chronological age had a 21% higher risk of mortality.
Based on this, the researchers examined serial photographs to understand signs associated with disease progression or response to treatment. Such longitudinal measures are already widely used in clinical practice. For example, changes in prostate-specific antigen (PSA) levels over time help assess prostate cancer risk, while blood pressure variation provides information on cardiovascular risk.
LONG and overall survival in cancer
Researchers conducted a retrospective study of 2,276 cancer patients receiving radiation therapy. Most participants were White, with a median age of 63.4 years, and 62.9% had metastatic cancer in the first course of radiotherapy, increasing to 78.7% in the second.
The researchers used two photographs of each patient, taken as part of routine clinical practice for identification purposes at the start of each course of radiotherapy. These were used to predict biological age using the Face Age AI algorithm.
FAR was calculated as the change in age of the face divided by the time between photographs and provided a measure of the rate of aging. This was analyzed for associations with overall survival.
Intervals between photographs were categorized into short (10–365 days), intermediate (366–730 days), and long (731–1,460 days). The FAR range was very large in the short-term group, due to the small denominator. Thus, only a FAR >20 was reported to be significant in this group, while in the intermediate and long-term groups, the threshold was set at FAR >10 and >1, respectively.
High FAR is associated with lower overall survival
For many patients, facial age predicted greater age than chronological age from the second photograph. A high FAR was associated with poorer overall survival in all groups, after adjusting for time between images, sex, race, and cancer diagnosis in the second course of radiation therapy.
In the short-term group, mortality risk was 25% higher with high FAR. In the intermediate and long-term groups, a high FAR was associated with a 37 % and 65 % higher risk of mortality.
The researchers repeated the analysis with only metastatic cancer patients. The same associations were found, but with a sharper separation in survival outcomes between groups.
FAR is a stronger predictor of long-term survival
They also examined the combined effects of initial deviation of predicted facial age from chronological age (FADRT1) and FAR. This showed that when both high FADRT1 and FAR were high, patients always had the highest risk of mortality.
With increasing intervals between shots, especially in the long-term group, the differences in FAR values become smaller. Even so, FAR becomes the dominant predictor of survival outcomes, although both measures still play a substantial role in increased mortality risk.
This suggests that “FAR consistently outperforms FADRT1 as a prognostic marker at all time intervals, with the strongest performance at longer intervals.”
Possible mechanisms underlying FAR-based prediction
The authors emphasize the non-linear nature of biological aging, with accelerated molecular senescence, such as DNA damage and cellular senescence, often occurring at specific inflection points. In cancer patients, such dynamic parameters reflect not only the disease process but also the effect of cancer treatment.
By quantitatively measuring facial aging, FAR could reflect changes in health during treatment. Advantages of using FAR include accessibility, convenience, and cost-effectiveness, allowing repeated measurements to assess health changes during treatment.
If validated, it could be incorporated into current prognostic parameters to identify high-risk patients across multiple cancer categories and to guide decision-making regarding monitoring intensity, supportive care, and therapeutic approaches, particularly in advanced disease where less intensive or palliative strategies may be appropriate.
Study restrictions
The ethnic/racial and age composition of the sample limits the generalizability of the findings. Furthermore, the lack of data on disease progression and treatment meant that the higher FAR could not be interpreted as causal. Unmeasured factors such as cancer cachexia or treatment-related toxicities could have influenced the observed associations between FAR and survival.
Because photographs were taken at specific radiotherapy time points rather than regular intervals, their use could have introduced indication bias, as different interval groups may reflect different clinical scenarios, limiting generalizability. Pending validation of this work, ethical and privacy issues, as well as the potential for bias in such facial recognition systems, need to be addressed before its clinical translation.
Future studies should correlate disease type, stage and treatment in different populations, using easily accessible algorithms with strong data protection barriers. The current findings need to be validated in prospective studies and in combination with other markers of aging. If so, FAR could be a tool to help deliver personalized cancer care.
