Revolutionising Cancer Prognosis with Facial Recognition Technology
A groundbreaking study has emerged from researchers at Mass General Brigham (MGB), revealing that a simple selfie may provide insights into the survival probabilities of cancer patients. The team has developed an innovative artificial intelligence (AI) model capable of calculating biological age and predicting mortality risk in individuals with malignant neoplasms through facial image analysis. This technology could serve as a crucial tool for tailoring personalised treatment plans across various conditions.
Experts suggest that facial images hold the potential to act as a non-invasive and accessible biomarker for assessing overall health and the progression of ageing. Nonetheless, the challenge remains in quantifying and qualifying the information derived from facial appearance, primarily due to the difficulties in establishing objective interpretations of facial features.
Previous research indicates that advancements in AI could help address this issue. In recent years, experimental models have been devised to predict, among other factors, mortality risk from various causes and the progression of certain diseases by analysing photographs of millions of patients.
However, many of these algorithms face inherent limitations stemming from a lack of representative data across diverse populations and biases present in the datasets used for training. Consequently, this can lead to inaccurate results, particularly when applied to groups absent during the model’s development.
FAHR-Face: A Comprehensive AI Solution
The specialists at MGB aimed to overcome these inconsistencies with FAHR-Face, a sophisticated AI model trained on over 40 million facial photographs using self-supervised learning techniques. This method enabled the system to learn to recognise visual features and patterns independently of medical diagnoses associated with those traits.
Following this foundational phase, the model developed the capability to correlate elements within images and specialise in specific tasks, such as identifying signals linked to diseases. Smaller sets of photographs containing clinical information were employed in this advanced training phase.
Hugo Aerts, director of the AI in Medicine (AIM) programme at MGB and co-author of the study, explains, “This fundamental model approach provides a scalable means to develop multiple facial health biomarkers, rather than training each model from scratch.”
The findings, published in The Lancet Digital Health, detail how the researchers utilised FAHR-Face to create two tools designed to offer complementary prognostic information regarding mortality risk.
Estimating Biological Age through Facial Analysis
The first tool, known as FAHR-FaceAge, aims to estimate an individual’s biological age from their facial photograph. To create this tool, the algorithm underwent a two-stage training process utilising ten publicly accessible datasets comprising 743,152 images of individuals with known chronological ages.
To assess its efficacy, researchers conducted a control test where the system was tasked with calculating the ages of various individuals using photographs not included during the training phase. This exercise was later repeated with images of over 4,890 patients who had undergone radiotherapy.
The results indicated that the algorithm’s estimations deviated from chronological age by an average of 5.1 years in the general test and nearly seven years in the group of radiotherapy patients. The researchers interpret this discrepancy as a potential indicator of accelerated biological ageing.
Linking Biological Age to Cancer Survival Rates
In a subsequent analysis, the specialists investigated the relationship between this biological age indicator and the survival rates of 38,183 cancer patients who had received radiotherapy. The analysis unveiled that those with an estimated biological age at least five years older than their chronological age faced a 21% higher risk of mortality compared to other participants. This suggests that, within the studied cohort, individuals whose facial features reflected a significantly older age were also more likely to succumb to their illness.
Thus, the researchers concluded that FAHR-FaceAge could effectively estimate the biological age of patients and that this estimation correlates with their survival probabilities following a cancer diagnosis.
Predicting Mortality Risk with Facial Recognition
In contrast, the second algorithm, FAHR-FaceSurvival, was developed to predict mortality risk directly from a facial photograph, bypassing the need for biological age estimation. This system was trained on over 34,360 photographs of cancer patients, accompanied by survival data, enabling the model to discern facial features that may correlate with life expectancy post-diagnosis.
To evaluate its performance, scientists categorised patients into four groups according to the risk level determined by the AI. The group identified as having the highest risk exhibited a mortality rate over three times greater than that of those with the lowest risk. This disparity persisted even after accounting for variables such as age, sex, cancer type, treatment intent, and radiation technique.
Aerts noted, “FAHR-FaceSurvival advances the field of facial health recognition AI by learning survival risk directly from facial photographs of cancer patients. To our knowledge, this is the first demonstration that an AI model can predict cancer survival directly from a simple image.”
Complementary Tools for Clinical Decision-Making
While both tools utilise the same AI model and analyse facial features to derive their predictions, the authors emphasise that the biological age estimates and direct survival predictions do not rely on identical physiological characteristics. FAHR-FaceAge primarily focused on areas such as nasolabial folds and temples, which are associated with visible signs of ageing, while FAHR-FaceSurvival concentrated on regions below the eyes and around the nose.
These distinctions support the researchers’ hypothesis that ageing indicators and survival-related features could complement each other to yield more accurate prognoses.
The scientists clarify that both systems are designed as assistive tools to enhance clinical evaluations and support patient-tailored decisions, rather than replace the judgement of healthcare professionals.
Future Implications of AI in Patient Care
Raymond Mak, a member of the AIM and co-author of the study, remarked, “Our goal is to make health assessments based on facial photographs more quantitative, consistent, and scalable, so they can support personalised treatment decisions rather than replace medical judgement.”
The authors acknowledge that the proposed tools have limitations. Although the models can estimate biological age and predict survival from photographs taken under various lighting conditions, facial expressions, or angles—even when individuals are wearing makeup—errors are still possible. Additionally, the algorithms do not identify the biological mechanisms that could explain their predictions.
Despite these limitations, Mak highlights their potential utility in meeting the current demands of medicine. “As the population ages, physicians require better methods to estimate physical reserves and frailty before recommending intensive treatments such as radiotherapy, surgery, or systemic therapy,” he concludes.