Brecahad Mimic Iv Fusion Eval

brecahad-mimic-iv-fusion-eval

Multimodal Fusion of Histopathology Images and Electronic Health Records for Early Breast Cancer Diagnosis — Khandelwal et al. (2026) (arXiv:2604.17122, 2026)

What this evaluates

Evaluates a multimodal framework's ability to fuse patch-level histopathology features with structured EHR data for early breast cancer diagnosis, specifically probing performance on class-imbalanced minority classes like mitosis.

Datasets

• BreCaHAD — total ?; splits: test (-1)
• MIMIC-IV — total ?; splits: test (-1)

Metrics

• Macro-average AUC (primary) — range: [0, 1]
  • Unweighted mean of per-class ROC-AUC scores computed via one-vs-rest binary classification.
• Mitosis AUC — range: [0, 1]
  • ROC-AUC computed specifically for the mitosis class against all other classes combined.
• Accuracy — range: [0, 1]
  • Proportion of correctly classified instances out of the total.
• Macro F1 — range: [0, 1]
  • Unweighted mean of per-class F1-scores.

Input / output format

Input: Histopathology image patches for CNN/ResNet branches; structured tabular EHR features (age, comorbidity count, biomarkers) for MLP/XGBoost branches; concatenated latent vectors for the fusion model.

Output: Predicted class label (non-tumour, mitosis, tumour) and/or class probability scores.

Scoring recipe

def evaluate(y_true, y_pred, y_prob, classes=['non-tumour', 'mitosis', 'tumour']):
    acc = np.mean(y_true == y_pred)
    macro_f1 = f1_score(y_true, y_pred, average='macro')
    macro_auc = np.mean([roc_auc_score(y_true == c, y_prob[:, i]) for i, c in enumerate(classes)])
    mitosis_auc = roc_auc_score(y_true == 'mitosis', y_prob[:, 1])
    return {'accuracy': acc, 'macro_f1': macro_f1, 'macro_auc': macro_auc, 'mitosis_auc': mitosis_auc}

Common pitfalls

• Class imbalance causes overall accuracy to remain near-perfect (>0.99) while minority-class (mitosis) performance degrades significantly, obscuring clinical utility.
• Tabular baselines (MLP, XGBoost) operate on patient-level EHR data and cannot compute patch-level 'Mitosis AUC', making direct metric comparison with image models invalid without explicit notation.
• Macro-averaging masks per-class disparities; relying solely on macro-AUC without inspecting one-vs-rest ROC curves can hide poor discrimination on the mitosis class.

Evidence (verbatim from paper)

The fusion model achieves the highest macro-average AUC overall and the best mitosis-specific AUC, confirming the added value of multimodal integration.

Citation

@misc{khandelwal2026multimodal,
  title={Multimodal Fusion of Histopathology Images and Electronic Health Records for Early Breast Cancer Diagnosis},
  author={Khandelwal et al. (2026)},
  year={2026},
  note={arXiv:2604.17122}
}

• arXiv: 2604.17122