🚨 AI FaceAge 🚨

Practical Application of the tool

AI Generated FaceAge to Predict Cancer Prognosis

Introduction & Objective

This study recently published in the Lancet introduces FaceAge, a deep learning-based AI tool designed to estimate a person's age from facial images and assess its potential as a prognostic biomarker in cancer patients. Unlike chronological age, FaceAge may capture biological aging markers that correlate with survival outcomes. The research validates FaceAge across multiple clinical oncology cohorts and compares its predictive power against physician assessments and traditional prognostic models.

Methods

Datasets & Ethical Considerations

The model was trained on 56,304 images from the IMDb-Wiki database, focusing on individuals aged ≥60 (representative of typical oncology populations). Validation was performed on:

1. UTKFace dataset (healthy individuals)

2. Four clinical cohorts:

   • MAASTRO cohort (non-metastatic cancer patients, Netherlands)

   • Harvard Thoracic cohort (patients with lung cancer, USA)

   • Harvard Palliative cohort (metastatic cancer patients, USA)

   • Harvard non-cancer cohort (benign/early-stage tumors)

Ethical approval was obtained, with retrospective data use permitted under IRB waivers. No patient data was used for training.

FaceAge Pipeline

1. Face Detection: A convolutional neural network (CNN) located faces with 95% accuracy.

2. Feature Extraction & Age Prediction: An Inception-ResNet v1 CNN generated facial feature vectors and predicted age via regression. Performance was strong for ages ≥60 (mean absolute error: 4.09 years).

Statistical & Survival Analysis

• Kaplan-Meier curves and Cox regression assessed survival stratification.

• FaceAge was compared to chronological age and integrated into the TEACHH model (a palliative care prognostic tool).

• Metrics included log-likelihood ratio, concordance index, and AUC-ROC.

Clinical Utility Testing : A Hybrid Human + AI approach analysis

• Physician Survey: Ten clinicians/researchers predicted 6-month survival for 100 palliative patients using:

  1. Face images alone (AUC: 0.61)

  2. Images + clinical data (AUC: 0.74)

  3. Images + clinical data + FaceAge Risk Model (AUC: 0.80, nearing FaceAge’s standalone AUC of 0.81).

• FaceAge improved prognostic accuracy beyond human judgment.

Genomic Analysis

• Whole-exome sequencing of 146 lung cancer patients linked FaceAge to senescence-associated genes.

• CDK6 (a cell-cycle regulator) was significantly associated with FaceAge but not chronological age.

Results

FaceAge was evaluated in several different ways and adjusted for age, gender and lifestyle factors such as smoking and BMI. Below is the pointwise breakdown of how it performed.

Prognostic Performance

1. MAASTRO Cohort (Non-Metastatic Cancer, n=4,906)

   • FaceAge stratified mortality risk (p=0.0013), remaining significant after adjusting for age, sex, and tumor type.

   • Strongest predictive power for breast, genitourinary, and gastrointestinal cancers.

2. Harvard Thoracic Cohort (Lung Cancer, n=573)

   • FaceAge predicted survival independently of stage, ECOG status, smoking history, and treatment intent (HR 1.15 per decade, p=0.011).

   • Chronological age was not predictive (p=0.16).

3. Harvard Palliative Cohort (Metastatic Cancer, n=717)

   • FaceAge significantly predicted survival (HR 1.12 per decade, p=0.021), unlike chronological age.

   • Substituting FaceAge into the TEACHH model improved risk-group separation (higher HRs for high-risk patients).

FaceAge vs. Chronological Age

• Cancer patients looked 4.79 years older than their chronological age (p<0.001).

• Non-cancer cohorts (e.g., benign tumors) showed smaller gaps (1.95–3.86 years).

Lifestyle & Clinical Factors

• Smokers appeared 33.24 months older than non-smokers (p<0.001).

• BMI had a weak correlation (r=−0.0999, p<0.0001).

• ECOG performance status did not significantly affect FaceAge predictions.

Genomic Insights

• FaceAge correlated with CDK6, a gene linked to cellular senescence.

• No genes were significantly associated with chronological age.

Real-World Implications

1. Novel and Clinically Relevant Approach: The study introduces FaceAge, an innovative AI tool that leverages facial imaging to predict biological age and cancer prognosis, offering a non-invasive biomarker that correlates with survival.

2. Robust Validation Across Multiple Cohorts: The model was trained on a large, diverse dataset (IMDb-Wiki) and validated in four independent clinical cohorts, including both localized and metastatic cancers with consistent performance.

3. Human-AI Comparative Analysis: The physician survey demonstrated that FaceAge enhanced clinicians' prognostic accuracy, particularly when combined with clinical data (AUC improved from 0.74 to 0.80) with potential to become a real-world practice tool.

4. Genetic Explanation of Biological Age Marker: The association between FaceAge and CDK6 (a senescence-related gene) provides mechanistic insight, supporting the idea that facial aging reflects underlying molecular changes.

To Keep in Mind for Future Use

1. Limited Representation of Ethnicities

   • The MAASTRO cohort lacked ethnicity data, raising concerns about bias in FaceAge’s performance across diverse populations.

   • Most training data came from IMDb-Wiki, which may overrepresent certain demographics (e.g., Western celebrities).

2. Modest Effect Sizes in Some Analyses

   • While statistically significant, some associations (e.g., BMI correlation: r = -0.0999) were weak, limiting clinical utility for certain factors.

   • The HRs for FaceAge (e.g., 1.12–1.15 per decade) could suggest meaningful but not dramatic prognostic improvements.

3. Potential Confounding by Disease Burden

   • The study did not fully disentangle whether FaceAge reflects general aging, cancer-specific effects, or treatment-related changes.

   • For example, cachexia (wasting syndrome) in advanced cancer may artificially inflate perceived age.

4. Lack of Integration with Other Biomarkers

   • FaceAge was not compared to other aging biomarkers (e.g., epigenetic clocks, inflammatory markers), leaving its relative predictive value unclear.

In summary, this is first of kind large scale deep-learning based FaceAge biomarker study that showed its potential in conjunction with traditional markers in Chronic Disease Prognosis (Cancer). This means it could be applied to several other chronic disease conditions such as kidney disease, transplantations, heart diseases etc.