Auditing the Reliability of Multi-Omics Fusion Models in Breast Cancer: An Attention--Explanation Concordance and Faithfulness Framework

Background: Attention weights in multi-omics fusion models are routinely presented as intrinsically interpretable, yet whether they agree with independent post-hoc explanations of the same trained model -- and whether that agreement is causally meaningful -- has not been systematically tested. Methods: We train a four-modality cross-attention fusion model (mRNA, miRNA, DNA methylation, CNV) for early-stage breast cancer detection (Stage 0/DCIS and Stage I, TCGA-BRCA), benchmarked against two single-omics baselines and an explicit CNV ablation. We define a cosine-similarity concordance metric 𝐶𝑖 between attention weights and SHAP explanations, test a priori whether 𝐶𝑖 predicts prediction confidence (Hypothesis 1) and causal feature-perturbation faithfulness (Hypothesis 2), and externally validate via random-effects meta-analysis across four independent public GEO cohorts. Results: Concordance reliably predicts prediction confidence (Spearman 𝜌 = 0.335--0.339, replicated, 𝑝 ≈ 0.01), but is significantly associated with lower, not higher, feature-perturbation faithfulness (𝜌 = −0.333 to −0.388), the opposite of what a simple concordance-validatesexplanation account predicts. External validation across four independent GEO cohorts yields a pooled AUC of 0.917 (95% CI [0.859, 0.975]), evidence against a project-wide technical batch effect. CNV inclusion did not improve discrimination and reduced the usable cohort by 59%. Conclusion: Attention--explanation agreement is not sufficient evidence of causal explanation reliability; the two properties require independent evaluation.