Auditing shortcut learning and misclassification in artificial intelligence-based breast cancer genomic subtyping

Area of Science:

• Computational biology and bioinformatics
• Artificial intelligence in healthcare
• Genomic data analysis

Background:

• AI models are increasingly used for breast cancer genomic subtyping.
• Concerns exist regarding AI models exhibiting 'shortcut learning,' over-relying on specific features.
• Developing interpretable frameworks to audit AI behavior is crucial for clinical trust.

Purpose of the Study:

• To simulate shortcut learning mechanisms in AI-based breast cancer genomic subtyping.
• To develop an interpretable and reproducible framework for auditing feature over-reliance.
• To implement this framework in a low-code environment using readily available clinical predictors.

Main Methods:

• Retrospective analysis of The Cancer Genome Atlas-Breast Invasive Carcinoma dataset (n=691).
• Trained a multinomial logistic regression model using clinical predictors (ER, PR, HER2, stage, age) to predict PAM50 subtypes.
• Implemented a pseudo-Shapley additive explanations (SHAP) method in Stata to estimate variable influence and validated against canonical SHAP values.

Main Results:

• The clinical-only model achieved moderate explanatory power (pseudo-R²=0.396).
• Progesterone receptor (PR) and HER2 status were significant predictors for Luminal B and HER2-enriched subtypes, respectively.
• Pseudo-SHAP identified disproportionate influence of PR and HER2 status (ΔP up to +0.29), with strong correlation to canonical SHAP values (Spearman r=0.91).

Conclusions:

• Shortcut learning was evident, with the AI model over-relying on surrogate clinical biomarkers (PR, HER2) instead of genomic signals.
• The pseudo-SHAP method provides a scalable and interpretable auditing framework for detecting shortcut learning in AI models.
• This approach can enhance transparency, reproducibility, and equity in AI for precision oncology, especially in resource-constrained settings.