Explainable Artificial Intelligence in Clinical Decision Support: A Comparative Evaluation of SHAP, LIME, Attention and Permutation Attribution for Cardiovascular Risk Prediction

Abstract

Explainable artificial intelligence is increasingly presented as the mechanism that will make machine learning acceptable in clinical decision support. Yet the leading attribution methods are rarely compared against one another on the same model and data, so it is unclear whether they agree, whether they are stable, and what they cost. Objective. To quantify the agreement, stability, fidelity and computational cost of four attribution methods — SHAP, LIME, learned feature attention and permutation feature importance — applied to cardiovascular risk prediction. Methods. Four classifiers (L2-regularised logistic regression, random forest, gradient-boosted trees and a purpose-built attention-gated network) were trained on the UCI Cleveland heart-disease cohort (n = 302 after de-duplication; 54.3% prevalence) under stratified ten-fold cross-validation. Attributions were generated on a held-out 30% partition and compared on four axes: rank agreement (Spearman ρ, Kendall τ), run-to-run stability (25 repeated explanations per case), fidelity (area under a per-instance deletion curve, AUDC) and wall-clock cost per explained patient. Results. Discrimination was statistically indistinguishable across model families: random forest 0.909 ± 0.059, logistic regression 0.899 ± 0.060, gradient boosting 0.885 ± 0.054 and the attention network 0.876 ± 0.059 AUC. The four attribution methods disagreed substantially, with pairwise Spearman ρ ranging from 0.582 (SHAP vs. permutation) to 0.846 (SHAP vs. LIME); serum cholesterol was ranked third most important by SHAP and last of thirteen by permutation importance. LIME attributions varied across identical repeated runs with a mean coefficient of variation of 0.279 and a top-five Jaccard similarity of 0.799, whereas SHAP and attention were exactly reproducible. SHAP was the most faithful (AUDC 0.662) and LIME the costliest at 13.48 ms per patient, some 6,700 times slower than reading attention weights. Conclusions. Explanation method is not an implementation detail: it changes which risk factor a clinician is told mattered. Attention weights were fast but collapsed onto a single feature and were the least faithful of the post-hoc methods; LIME was both the slowest and the least reproducible. Since a regularised logistic regression matched every black box here, the interpretability of the model itself remains the strongest available guarantee.