Table 5 Expert agreement rates with SHAP attributions.
From: Explainable federated transformer framework for joint leukemia classification and stage prediction
Explanation component | Correct/clinically relevant | Partially correct | Incorrect/misleading |
|---|---|---|---|
Visual SHAP (image patches) | 92% (23/25 cases) | 6% (1/25 cases) | 2% (1/25 cases) |
Textual SHAP (clinical tokens) | 88% (22/25 cases) | 8% (2/25 cases) | 4% (1/25 cases) |
Multimodal consistency (image–text alignment) | 84% (21/25 cases) | 12% (3/25 cases) | 4% (1/25 cases) |
Helpfulness in explaining staging decisions | 80% (20/25 cases) | 16% (4/25 cases) | 4% (1/25 cases) |
Overall agreement (all modalities) | 88% (22/25 cases) | 8% (2/25 cases) | 4% (1/25 cases) |
