Applying Deep Learning Based Facial Age Phenotyping to Improve Life Expectancy Prediction in Metastatic Cancer Patients Receiving Palliative Radiotherapy

Purpose/Objective(s)

Prognostic tools such as the TEACHH model (risk scoring based on cancer type, ECOG PS, age, prior palliative chemotherapy, hospitalization, and hepatic metastases) aim to predict life expectancy (LE) in metastatic cancer patients receiving palliative radiotherapy (RT). In our prior study, a deep learning model predicting biological age from facial photographs (FaceAge) was developed and showed prognostic potential in cancer patients. Here, we evaluated the prognostic significance of extreme discordance between FaceAge vs chronological age (FaceAge–Age) among patients receiving palliative RT and applied FaceAge to the TEACHH model.

Materials/Methods

A retrospective study of 690 patients with metastatic cancer treated by palliative RT between 2012 and 2018 at six clinic locations was conducted. FaceAge estimates were derived based on patients’ facial photographs taken before RT. Cox and logistic regression analyses were used to evaluate predictors of overall survival (OS) and early mortality (<3 months), respectively. FaceAge was substituted for chronological age in the TEACHH model and model fitness was compared via likelihood ratio test (LRT).

Results

Median OS was 9 months and 41% died within 3 months. Fifty-five percent had ≥ 5 years of absolute difference in FaceAge vs chronological age. Twenty-one percent had a much older FaceAge with FaceAge–Age of ≥ 10 years. In multivariate analyses, FaceAge–Age ≥ 10 years was significantly associated with worse OS (HR = 1.38, P = 0.01) and increased risk of early mortality within 3 months (OR = 1.68, P = 0.02), even after adjusting for other significant predictors (primary cancer type, ECOG PS, chemotherapy, and hospitalization). For all patients, substituting FaceAge for chronological age in the TEACHH model improved LE group stratification (LRT = 6.1, P < 0.01). Among patients with ≥ 5 years of FaceAge vs age discrepancy, the TEACHH model failed to significantly stratify into 3 expected LE groups, but substituting FaceAge for age allowed for significant stratification (Table; LRT = 7.0, P < 0.01).

Conclusion

Extreme discordance in facial aging may be a valuable prognostic marker for metastatic cancer patients receiving palliative RT. Moreover, substituting FaceAge for chronological age improved the performance of an existing LE prediction model, especially in patients with extreme discordance, by more accurately capturing biological age at end-of-life. Such AI biomarker may enhance LE predictions and aid in end-of-life treatment decision making.