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

To simulate shortcut learning mechanisms in AI-based breast cancer genomic subtyping and to develop an interpretable, reproducible framework capable of auditing feature over-reliance using a low-code environment.

A retrospective secondary analysis of The Cancer Genome Atlas-Breast Invasive Carcinoma dataset was conducted. A multinomial logistic regression model was trained on 691 complete cases using only clinical predictors (estrogen receptor [ER], progesterone receptor [PR], HER2 status, tumor stage, and age) to predict PAM50 molecular subtypes, a genomic classification system originally developed by Parker et al, and widely used in breast cancer subtyping. A pseudo-Shapley additive explanations (SHAP) attribution method was implemented in Stata to estimate the marginal influence of each variable. Model performance was evaluated using McFadden's pseudo-R ² and validated by comparing pseudo-SHAP scores against canonical SHAP values derived in Python.

To enhance reproducibility, we now detail of the pseudo-SHAP methodology implemented in Stata. This approach estimates first-order marginal effects of each clinical predictor by simulating variable perturbations within a fitted multinomial logistic regression model. Specifically:1. Predicted probabilities for each PAM50 subtype are generated using the baseline model.2. Each predictor variable is perturbed individually (eg, PR status toggled from negative to positive), holding other features constant.3. New probabilities are calculated postperturbation, and the change in probability (ΔP) for each class is recorded.4. These ΔP values are averaged across all samples to provide directional importance scores, representing monotonic effects similar to mean SHAP values.This simulation assumes linear additive contributions and does not capture interaction or nonlinear effects.

The clinical-only model achieved a pseudo-R ² of 0.396, indicating moderate explanatory power. Progesterone receptor and HER2 status were statistically significant predictors (P <.01) for Luminal B and HER2-enriched subtypes, respectively. The pseudo-SHAP framework identified disproportionate influence of these features (ΔP up to +0.29 for HER2-enriched). Correlation between pseudo-SHAP and SHAP values was strong (Spearman r = 0.91, P<.001), confirming method validity.

To assess how shortcut learning manifests in prediction outcomes, we examined the directional shifts in subtype assignments associated with PR and HER2 status perturbations. Perturbing PR status resulted in increased predictions of Luminal A and Luminal B, and decreased predictions of Basal-like subtypes. Similarly, toggling HER2 status elevated the probability of HER2-enriched classification even in the absence of supporting genomic evidence.For example, the average ΔP from HER2-negative to HER2-positive status led to a +0.29 increase in HER2-enriched classification probability. Such shifts demonstrate systematic over-reliance on biomarker correlates, consistent with shortcut learning..

Shortcut learning was evident as the model over-relying on surrogate clinical biomarkers. Such behavior may distort biological interpretation and compromise clinical decision support reliability.Importantly, while PR and HER2 are biologically linked to PAM50 subtypes, their disproportionate weight in a clinical-only model reveals reliance on clinically available proxies rather than genomic signals. The model effectively substitutes biomarker status for molecular data, which constitutes shortcut behavior in this analytic context.To illustrate operational relevance, pseudo-SHAP attribution shifts could be integrated into clinical quality assurance pipelines. For instance, if the ΔP associated with HER2 status exceeds 0.25 persistently across batches, this could trigger retraining alerts or attribution-based warnings in a clinical decision support system.Additionally, as PR and HER2 status may be differentially available across populations, their overuse may propagate disparities. Pseudo-SHAP auditing can flag these risks early, improving fairness and model generalizability.

The pseudo-SHAP method provides a scalable, interpretable auditing framework for detecting shortcut learning in resource-constrained environments. Its application could enhance transparency and equity in AI models for precision oncology.