Fairness of machine learning readmission predictions following open ventral hernia repair.

introductionFew models have predicted readmission following open ventral hernia repair (VHR), and none have assessed fairness. Fairness evaluation assesses whether predictive performance is similar across demographic groups, ensuring that biases are not propagated. Therefore, we generated an interpretable machine learning model to predict readmission following open VHR while assessing fairness.

methodsNSQIP (2018-2021) was queried for open VHR. We developed an XGBoost model to predict unplanned readmissions within 30 days of surgery with fivefold cross-validation. Performance and fairness were assessed by demographic groups: gender (female vs. male), ethnicity (Hispanic vs. non-Hispanic), and race (non-White vs. White). We identified influential features within demographic groups using SHapley Additive exPlanations (SHAP).

results59,482 patients were included with a readmission rate of 5.5%. The model had an AUC of 0.72 and a Brier score of 0.16. Fairness metrics revealed minimal performance differences between demographic groups. SHAP revealed that influential factors were similar across demographic groups and included days from operation to discharge, morbidity probability, and operative time.

conclusionUsing interpretable machine learning, we identified unique predictors for unplanned readmission following open VHR. Fairness metrics revealed minimal differences in performance between demographic groups. SHAP showed similar influential factors across demographic groups. Future surgical machine learning models should similarly assess models using fairness metrics and interpretation of predictions.