acceptodds
Under review as a conference paper at ICLR 2027

FaceAge in Patients with Oligometastatic Cancer: External Validation and Racial Differences in Facial Age Estimates

Abstract

Purpose: FaceAge is a deep learning biomarker that estimates biological age from facial photographs and has been associated with survival in patients with metastatic cancer. We hypothesized that FaceAge would predict survival in patients with oligometastases and that the newer foundation model FaceAge 2.0 would have smaller racial differences compared to FaceAge 1.0. Methods & Materials: Between 2014 and 2023, 177 consecutive patients with 1 to 5 distant metastases were treated with comprehensive involved-site radiation at a community hospital. Frontal facial photographs obtained during radiation simulation were processed with FaceAge 1.0 and FaceAge 2.0/FAHR-Face. Photographs from 167 patients met quality criteria. Associations with overall survival (OS) were evaluated using Kaplan-Meier estimates, Cox regression and the concordance index. We compared FaceAge estimates across racial groups after adjusting for chronological age. Results: At a median follow-up of 44.0 months, the median OS was 42.8 months. After adjusting for chronological age, FaceAge 1.0 scores were underestimated in Black patients, with a mean difference of 10.6 years (p<0.001) compared to White patients. When using FaceAge 2.0, the difference decreased to 3.0 years (p=0.16). On univariable analysis, FaceAge 1.0 (HR 1.03 per year, p=0.004) and FaceAge 2.0 (HR 1.03 per year, p=0.002) both predicted OS, consistent with previous FaceAge studies. Chronological age was also strongly prognostic (HR 1.04 per year, p<0.001) and remained significant on multivariable analysis. Adding FaceAge to chronological age and established prognostic factors only modestly improved discrimination (C-index 0.74 without FaceAge vs. 0.75 with FaceAge 1.0 or 2.0). Conclusions: FaceAge was externally validated as a univariable prognostic biomarker in oligometastatic cancer. FaceAge 2.0 substantially reduced racial disparities in facial age estimates seen with FaceAge 1.0. Further studies are warranted to determine whether facial aging models perform consistently across racial groups before routine clinical use.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
algorithmic fairness, facial age estimation, computer vision, foundation models, medical artificial intelligence, oncology

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 2 reviews, above 0% of the others.
Reviewer_mj66OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies and validates FaceAge in patients with oligometastatic cancer for prognosis. The paper collects/utilizes a private dataset containing 177 patients with 1 to 5 distant metastases from 2014 to 2023. Unlike previous studies, the authors found that FaceAge did not outperform chronological age in predicting overall survival.

Critical strengths
  • The study addresses a real-world clinical problem of predicting survival in a specific type of cancer.
  • The authors compare FaceAge as a biomarker of survival with other biomarkers comprehensively and investigate sub-population performance (e.g. race).
Critical weaknesses
  • The study involves only 177 patients from only from 1 site, as far as I understand, and has a highly imbalanced patient background (145 white, 20 black, 12 other race). This dataset size is very small compared to previous similar studies, which makes the paper's claim less convincing. For reference: [1] (6,196 patients), [2] (24,556 patients), [3] (2,276 patients), [4] (670 patients).
  • While the authors provide p-values for statistical significance, there are inconsistencies in their interpretations. In Line 165, the authors say "mean FaceAge 2.0 estimate for Black patients was 1.4 years older than calendar age (p=0.62)", which I believe the interpretation is inaccurate, as the null hypothesis cannot be rejected because of a high p-value. In other places, such as in Line 198, the authors acknowledge a statistically insignificant association between FaceAge and post-radiation systemic therapy intensity based on p-values.
  • The authors include their main figure and tables after the References, which should be considered an Appendix. This makes reading uncomfortable, especially when looking at the main findings provided in the figure and tables.

References:

  1. "FaceAge, a deep learning system to estimate biological age from face photographs to improve prognostication: a model development and validation study", The Lancet Digital Health, 2025.
  2. "FaceAge as a biomarker for prognosis and treatment stratification in a large-scale oncology cohort", J Natl Cancer Inst., 2026.
  3. "Face aging rate quantifies change in biological age to predict cancer outcomes", Nature Communications, 2026.
  4. "Multimodal Assessment of Biological Age Following Radiation Therapy Among Patients With Early-Stage NSCLC", JAMA Netw Open, 2026.
Questions for authors
  • Lines 119-120. Do the authors perform sub-population survival analysis for patients with recorded dates of death and dates of last follow-up?
  • Is it possible to estimate biological age more reliably than by looking at patients’ photographs? If yes, do the authors try to verify FaceAge estimation with a more reliable method?
Reviewer_2uPBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

I strongly recommend rejection. The entire manuscript was written by AI agents, without meaningful human revision or reorganization. Submitting such an unedited manuscript demonstrates a serious lack of care and wastes the ICLR community’s limited reviewing time and resources. Authors must take responsibility for the quality and integrity of everything submitted under their names.

I request that the Area Chair examine the evidence and assess the submission’s compliance with ICLR’s AI-use policy and Code of Ethics. I will not provide further substantive review comments.

Critical strengths

See paper summary.

Critical weaknesses

See paper summary.

Questions for authors

See paper summary.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.