Ensemble SHAP Aggregation and Attribution Variability in Clinical Machine Learning: A COVID-19 Mortality Study

Objective: To combine performance-weighted ensemble SHAP aggregation with resampling-based assessment of feature-importance variability and patient-level attribution alignment for interpreting COVID-19 mortality predictions. Methods: We analyzed a prospective cohort of 1,857 patients hospitalized with COVID-19 at two hospitals in Peru. Ten predictive algorithms were evaluated using five-fold cross-validation, and their mean AUROC values determined their attribution-aggregation weights. Within each model and training trial, signed SHAP values were normalized by the mean total absolute attribution across evaluation patients before performance-weighted aggregation. Feature importance was summarized across 30 resampled training trials using means and normal-approximation 95% confidence intervals. Patient-level attribution alignment was assessed using cosine similarity between signed feature-attribution vectors for corresponding patients, with whole-trial patient-correspondence randomization and Benjamini--Hochberg correction. Subgroup observations were reweighted toward the evaluated population's feature distributions for secondary comparisons of existing absolute SHAP values. Results: Among the 1,857 included patients, 982 (52.9%) died during hospitalization. Random forest achieved the highest mean AUROC (0.925 +/- 0.010), followed by AdaBoost (0.919 +/- 0.017) and logistic regression (0.918 +/- 0.007), with variability reported as the SEM. Attribution profiles were more similar across repeated training trials of the same algorithm (mean Pearson r=0.783; SEM, 0.004) than across algorithms within the same trial (mean Pearson r=0.369; SEM, 0.006). The largest normalized attribution magnitudes were observed for dexamethasone use at home without oxygen support (6.186%; 95% CI, 5.795--6.577), PaO2/FiO2 ratio (3.014%; 95% CI, 2.787--3.242), shortness of breath (2.774%; 95% CI, 2.592--2.956), and FiO2 (2.560%; 95% CI, 2.284--2.836). Features differed in patient-level attribution alignment, indicating that importance magnitude and robustness provided complementary information. Among 1,763 eligible subgroup--feature comparisons, 64 had nominal one-sided p < 0.05, although none remained below 0.05 after Benjamini--Hochberg adjustment. Conclusions: Algorithms with similar predictive performance produced substantially different feature-attribution profiles. Normalized performance-weighted SHAP aggregation provided a representative relative-importance summary across algorithms, while resampling-based alignment assessment qualified the consistency of individual feature contributions. The resulting attribution patterns describe fitted-model behavior and should not be interpreted as causal clinical effects.