Explainable Machine Learning Models Using Routine Clinical Data for Early Prediction of Necrotizing Fasciitis: A Single-Center Retrospective Study.

BACKGROUND

Necrotizing fasciitis (NF), a fulminant soft tissue infection manifesting acute onset and rapid progression with substantial mortality risks, poses diagnostic challenges due to its predominant reliance on clinical evaluation during initial presentation, frequently leading to diagnostic inaccuracies. This study constructs an interpretable machine learning model leveraging clinical data to optimize early prediction of NF.

METHODS

This single-center retrospective study analyzed 288 soft tissue infection patients, partitioned into training (n = 201) and validation (n = 87) cohorts using a 7:3 allocation ratio. This study employed LASSO regression to identify key predictors, subsequently implementing five machine learning algorithms-support vector machines (SVM), decision trees (DT), Light Gradient Boosting Machine(LightGBM), eXtreme Gradient Boosting(XGBoost)and random forest(RF). Model performance was assessed through stratified 10-fold cross-validation.

RESULTS

Among 288 enrolled patients, 81 cases (28.1%) demonstrated NF confirmation through pathological/surgical evaluation. The XGBoost model demonstrated superior discriminative ability (AUC = 0.809), surpassing the conventional nomogram's predictive accuracy (AUC = 0.724).

CONCLUSIONS

The XGBoost model facilitates accurate early identification of NF leveraging routine clinical indices. The model's interpretability facilitates clinical decision support systems by providing evidence-based diagnostic workflow optimization.