Designing an explainable algorithm based on XGBoost and genetic algorithm for predicting hospitalization needs of...

Timely identification of COVID-19 outpatients who are at risk of hospitalization is critical for preventing clinical deterioration and optimizing healthcare resources. Although machine-learning models have demonstrated high predictive accuracy, their limited interpretability often hinders clinical adoption. This study aims to develop a hybrid explainable framework that combines the predictive strength of XGBoost with clinically interpretable rule-based explanations to support decision-making in real clinical settings. A retrospective dataset of 1278 COVID-19 patients was analyzed after applying strict inclusion and exclusion criteria. Twenty-seven clinical, laboratory, and demographic variables were preprocessed using outlier detection, multiple imputation by chained equations, and stratified train-test splitting validated through a Kolmogorov-Smirnov test. XGBoost was trained and benchmarked against logistic regression, random forest, LightGBM, and a neural network. For interpretability, candidate rules were extracted from a constrained Random Forest and optimized via a genetic algorithm (GA) using accuracy-support multi-objective fitness. Clinical validation of rules was performed by ten physicians using the Content Validity Index (CVI; threshold ≥ 0.85). XGBoost achieved superior predictive performance with an AUC of 0.85, sensitivity of 73.5%, specificity of 88.7%, AUPRC of 0.72, and a Brier Score of 0.085. Baseline models demonstrated lower discrimination and calibration. Fairness evaluation indicated stable model behavior across demographic and comorbidity subgroups. Sensitivity analysis identified SpO₂, CRP, age, D-dimer, ferritin, and lymphocyte percentage as the most influential predictors. From 400 initial rules, 80 were selected and refined, and 40 clinically valid rules were finalized through expert review. The explainability framework expands upon the classical Decision Tree Surrogate method, producing global IF-THEN rules that outperform local attribution tools such as SHAP and LIME in practical interpretability. The proposed hybrid system successfully integrates high-accuracy machine-learning predictions with clinically validated, interpretable rules, offering a transparent decision-support tool for hospitalization risk assessment in COVID-19 outpatients. Its modular design enables rapid adaptation to future infectious disease outbreaks, supporting broader clinical deployment and improved triage decision-making.