Additional file 1 of An interpretable prognostic model for early unfavorable outcomes after bronchiectasis surgery

SHAP summary plot of the top 10 features in the random forest comparator model. This figure presents the 10 most influential features in the random forest comparator model, ranked by mean absolute SHAP value. Each dot represents one patient.

The horizontal position indicates the direction and magnitude of that feature's contribution to the predicted risk of early unfavorable outcome; positive SHAP values indicate a higher predicted risk and negative SHAP values indicate a lower predicted risk. Dot color represents the original feature value, with red denoting higher values and blue denoting lower values. Features are ordered from top to bottom according to mean absolute SHAP value, with higher-ranked features contributing more to the overall model prediction.

SHAP, Shapley Additive exPlanations. Supplementary Figure S2. Forest plot of odds ratios from conventional multivariable logistic regression including the predictors retained in the Elastic Net model.

Dots represent odds ratios (ORs), and horizontal bars represent 95% confidence intervals (95% CIs) from a conventional multivariable logistic regression including the five predictors retained in the Elastic Net model: preoperative pCO2, left ventricular ejection fraction (LVEF), forced expiratory volume in 1 second as percentage of predicted (FEV1%pred), maximal voluntary ventilation as percentage of predicted (MVV%pred), and total Bronchiectasis Severity Index (BSI) score.

All continuous predictors were entered per 1-standard deviation increase. The vertical dashed line indicates the null value of OR = 1.0. OR > 1 indicates an increased risk of early unfavorable outcome, whereas OR < 1 indicates a decreased risk.

In this supplementary analysis, preoperative pCO2 and total BSI score were associated with higher risk, FEV1%pred showed a protective association, and LVEF and MVV%pred showed directions consistent with the primary model. Supplementary Table S1. Sensitivity analyses showing model performance after sequential removal of correlated predictors.

Sensitivity analysis for the inclusion of FEV1%pred and MVV%pred in the primary model. The primary model was the final Elastic Net model. Alternative models were reconstructed after excluding FEV1%pred or MVV%pred separately, with all other analytic settings unchanged.

Excluding FEV1%pred worsened model performance (AUC 0.821 to 0.794; Brier score 0.171 to 0.183), whereas excluding MVV%pred produced minimal change (AUC 0.822; Brier score 0.170). This suggests that FEV1%pred carried more key predictive information, while MVV%pred added relatively limited incremental value.