Interpretable machine learning model predicts the early gastrointestinal dysfunction risk in severely burned patients: a multicenter-based study.

BACKGROUND

Gastrointestinal (GI) dysfunction is a life-threatening complication following severe burn injury, significantly increasing risks of multi-organ failure and mortality. This study aimed to develop and validate the first machine learning (ML)-based clinical prediction model for GI dysfunction after severe burns by leveraging explainable artificial intelligence (AI) techniques to support early clinical intervention.

METHODS

In this retrospective multicenter study, 570 patients with severe burns were enrolled: 469 from Hospital A [randomly split into training (n = 328) and internal validation (n = 141) sets] and 101 from Hospital B (external validation set). Predictors of GI dysfunction were identified using least absolute shrinkage and selection operator (LASSO) regression and Boruta algorithm. Eight ML algorithms were developed and evaluated using a 7:3 training-validation split and external validation. Model performance was assessed by area under the curve (AUC), accuracy, specificity, and decision curve analysis (DCA). Model interpretability was provided using SHapley Additive exPlanations (SHAP).

RESULTS

Among 570 patients, the incidence of GI dysfunction was 35.61% (203/570). The XGBoost algorithm showed superior discrimination, with an AUC of 0.910 (95% CI: 0.878-0.941) in the training set, 0.851 (0.790-0.913) in the internal validation set, and 0.908 (0.837-0.979) in the external validation set. SHAP analysis identified five key predictors by importance: SOFA score, TBSA, inhalation injury, blood culture result, and hematuria.

CONCLUSION

We developed and validated the first interpretable ML-based model for predicting GI dysfunction after severe burn injury, with XGBoost achieving high performance. This model could help identify high-risk patients for personalized pre-emptive management.