1049 - Deep Learning Based Facial Age Phenotyping in a Large-Scale, Real-World Cancer Patient Population

G. Lee¹, F. Haugg², D. Bontempi^2,3, J. He¹, O. Zalay^2,4, D. S. Bitterman^1,2, P. J. Catalano⁵, V. Prudente^2,6, S. Pai^2,6, C. V. Guthier^1,2, B. H. Kann^1,2, H. Aerts^1,2, and R. H. Mak^1,2; ¹Department of Radiation Oncology, Dana-Farber Cancer Institute/Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, ²Artificial Intelligence in Medicine (AIM) Program, Mass General Brigham, Harvard Medical School, Boston, MA, ³Department of Radiation Oncology (MAASTRO), Maastricht University, Maastricht, Netherlands, ⁴Division of Radiation Oncology, Queen’s University, Kingston, ON, Canada, ⁵Department of Biostatistics, Harvard T.H. Chan School Of Public Health, Harvard University, Boston, MA, ⁶Radiology and Nuclear Medicine, CARIM & GROW, Maastricht University, Maastricht, Netherlands

Purpose/Objective(s): Humans age at different rates and a person’s facial characteristics may yield insight into their biological age and physiological health. Our prior study demonstrated that FaceAge, a deep learning system that estimates biological age based on facial photographs, could be a potential prognostic biomarker among cancer patients. In this study, we evaluate the prognostic value of the extreme difference between FaceAge and chronological age (FaceAge–Age) across a large, real-world clinical dataset in predicting long-term survival and early mortality. Materials/

Methods: 24,556 patients with 28 different cancer types and age =60 who underwent radiation therapy (RT) between 2008-2023 at six clinic locations, with facial photographs obtained for identification prior to an RT course, were retrospectively analyzed. FaceAge estimates and chronological age were compared across different diagnoses/clinical contexts, and survival analyses (Cox and logistic regression) were conducted for overall survival (OS) and risk of early mortality.
Results: Median follow-up was 22 months for all and 38 months for surviving patients. Extreme FaceAge vs Age discordance (=5 years) was seen in 48%, with 13% who appeared much older (FaceAge–Age =10 years) and 13% who appeared much younger (FaceAge–Age =-5 years) than chronological age. Patients with younger age, female sex, diagnoses with worse prognosis (e.g. lung vs breast cancer), and palliative intent treatment had a higher likelihood of FaceAge–Age =10 years. Median OS was 89 months; mortality within 30 and 60 days was observed in 3% and 7%, respectively. Across a variety of clinical sub-groups, including metastatic disease and multiple primary cancer types (e.g. breast, prostate, lung, head/neck, and rectal cancer), FaceAge–Age was significantly associated with OS. Median OS was 51 vs 84 months among patients with FaceAge–Age =10 vs =-5 years (p < 0.01). On multivariate analysis (MVA), FaceAge–Age =10 years was significantly associated with worse OS (HR 1.26, p<0.01) while FaceAge–Age =-5 years predicted better OS (HR 0.90, p<0.01). FaceAge–Age =10 years was associated with greater odds of 30 day (OR 1.38, p<0.01) and 60 day (OR 1.33, p<0.01) mortality. Other significant variables adjusted for in the MVA included age, sex, race, cancer type, treated year, course intent, and RT technique.
Conclusion: Across a large, real-world dataset of cancer patients, we demonstrate that patients with more advanced, aggressive cancers tend to have significantly older FaceAge compared to age, and extreme discordance between FaceAge and chronological age is a novel, independent predictor of survival and early mortality. Our findings support future studies optimizing deep learning-based facial health assessments for potential integration into clinical prognostication models and personalized treatment decision-making.