Accuracy is not enough: explainable boosting machine model and identification of candidate biomarkers for real-time sepsis risk assessment in the emergency department

Background Sepsis poses a significant threat in emergency settings, necessitating tools for early and interpretable risk assessment. This study aimed to develop a robust explainable boosting machine (EBM) model, one of the explainable artificial intelligence (XAI) technologies, to construct a predictive model that balances high accuracy and clinical interpretability for use in emergency departments (EDs) and to examine candidate biomarkers. Methods The study identified a significant class imbalance problem in the sepsis distribution among 560 sepsis and 1012 non-sepsis patients. To address the imbalance issue, SMOTE-NC was applied in the training data. The data was divided into two parts, 80% training and 20% testing. To ensure the reliability of the models and to report unbiased results, this process was repeated 100 times and the average performance was reported. To determine the best model for sepsis prediction, five different models (AdaBoost, Gradient Boosting, CatBoost, LightGBM, and EBM) were trained, and their performances were evaluated. In the last stage, we presented local and global explanations of EBM. Results The EBM model achieved the highest success by reaching 79.1% F1-score, 80.9% sensitivity, and 84.8% AUC after resampling. In the global explanations, the variables with the highest weights in the model’s decision process were identified as positive blood culture, oxygen saturation, and procalcitonin, respectively. Conclusion The EBM model accurately predicts sepsis risk based on clinically relevant biomarkers. The model’s high performance and inherent transparency can foster trust among clinicians and facilitate its integration into emergency department workflows for real-time decision support.