AI Model Predicts 90-Day Cancer Mortality Risk from Facial Photos

A recent study from Mass General Brigham has unveiled an AI model named FAHR-FaceSurvival that can assess cancer mortality risk from a single photograph of a patient's face. In this study, 122 patients were identified as high-risk, with over half (52.5%) succumbing within 90 days of the assessment. In contrast, only about 2% of the 975 patients categorized in the lowest-risk group faced similar outcomes. This groundbreaking research, published in The Lancet Digital Health, marks a significant advancement in the ability to predict cancer survival directly from facial images.

The FAHR-FaceSurvival model is built upon a more extensive foundation model called FAHR-Face, which has been trained using over 40 million facial images. Additionally, a related tool, FAHR-FaceAge, estimates a patient’s biological age based on facial features. Among patients undergoing radiotherapy, those whose biological age was assessed to be more than 20 years older than their chronological age exhibited a 21% mortality rate, compared to approximately 6% for those appearing at least a decade younger. The survival model primarily analyzes the bridge of the nose and the area beneath the eyes, while the age estimation focuses on nasolabial folds and forehead features.

Hugo Aerts, PhD, who leads the Artificial Intelligence in Medicine program at Mass General Brigham, described this innovation as a significant leap in facial health recognition AI. He noted that this is the first instance where an AI model can predict cancer survival directly from a simple photograph. Raymond Mak, a radiation oncologist involved in the research, emphasized that these tools are intended to complement existing prognostic evaluations rather than replace them.

While the findings are promising, external experts have urged caution regarding the model's application. Wilbur Lam from Emory University highlighted the compelling nature of the study, which is based on a substantial dataset, but stressed the necessity for prospective clinical validation across various hospitals and patient demographics. The authors of the study acknowledged limitations, including the predominance of white participants from the Boston area in the training cohort and the minimal impact of rhinoplasty on age estimations.

This research comes at a time of rapid advancement in healthcare AI, with 65 stories related to the field reported in the last 90 days alone. To further enhance the model's training data, a public research portal has been established, allowing volunteers to submit their own facial photographs for analysis. This initiative aims to broaden the dataset and improve the model's accuracy in diverse populations.