1154 - Comparing Face Photograph-Based and Blood-Based Biological Age Biomarkers for Survival Prediction in Patients Receiving Radiation Therapy
Presenter(s)
F. Haugg1,2, A. Warrington1, G. Lee1, L. L. Thompson1,2, D. S. Bitterman1,2, P. J. Catalano3,4, C. V. Guthier1,2, B. H. Kann1,2, V. N. Gladyshev5, H. Aerts1,2, and R. H. Mak1,2; 1Department of Radiation Oncology, Mass General Brigham/Dana-Farber Cancer Institute, Harvard Medical School, Boston, MA, 2Artificial Intelligence in Medicine (AIM) Program, Mass General Brigham, Harvard Medical School, Boston, MA, 3Department of Biostatistics, Harvard T.H. Chan School Of Public Health, Boston, MA, 4Department of Data Science, Dana-Farber Cancer Institute, Boston, MA, 5Harvard Medical School and Brigham and Women’s Hospital, Boston, MA, USA, Boston, MA
Purpose/Objective(s): Chronological age is routinely used to guide radiation therapy decisions but poorly captures physiologic reserve. Biological age biomarkers may better stratify risk, yet established measures like PhenoAge (a blood-based age estimate derived from chronological age plus albumin, creatinine, glucose, CRP, lymphocyte %, MCV, RDW, alkaline phosphatase, and WBC) require laboratory tests. We hypothesized that face photograph-based biological age biomarkers can approach blood-based PhenoAge for survival prediction and that training on biological age estimates improves prognostic performance over chronological age.
Materials/Methods: Three face photograph-based models with identical architecture were trained on separate cohorts with different training objectives: FaceAge (chronological age; 749,935 public images of presumed healthy adults), FacePhenoAge (PhenoAge; radiation therapy patients; n=4,088), and FaceSurvival (time-to-death; radiation therapy patients; n=36,939). These were compared with blood-based PhenoAge in an independent retrospective validation cohort of 1,167 radiation therapy patients with complete laboratory data from one academic medical center and four affiliated community sites (2008-2023). Pearson correlations assessed alignment with chronological age. Discrimination was assessed by Harrell C-index and time-dependent AUC at 3, 6, and 12 months. Top-decile 12-month event rates quantified risk enrichment. Multivariable Cox models adjusted for age, sex, race, and cancer site.
Results:
Among 1,167 patients (median age 64.9 years; 402 deaths; median follow-up 4.7 years), median biomarker values were FacePhenoAge 64.8 years [IQR 55.6-72.6], PhenoAge 69.0 years [58.2-79.1], and FaceSurvival 0.096 [0.041-0.209]. Correlations with age: FacePhenoAge r=0.83, FaceAge r=0.82, PhenoAge r=0.69, FaceSurvival r=0.21. All biomarkers were independently prognostic after covariate adjustment (all P<.001). FaceSurvival and PhenoAge achieved the strongest overall discrimination (C-index 0.694 and 0.692), followed by FacePhenoAge (0.613), FaceAge (0.576), and chronological age (0.551). FacePhenoAge exceeded FaceAge at every evaluated time point (?AUC 0.041-0.047; P<.001). Top-decile 12-month event rates were 47.0% (FaceSurvival), 39.3% (PhenoAge), 33.3% (FacePhenoAge), 25.6% (FaceAge), and 21.4% (age).
Conclusion: Face photograph-based biological age biomarkers provided independent prognostic value for overall survival in cancer patients receiving radiation therapy. Training AI models with biological age estimate biomarkers achieved better prognostic performance than chronological age. Photograph-based biological age estimation may offer a low-burden, scalable tool that allows more frequent and less invasive risk stratification in radiation oncology workflows.