Figure 1.

Workflow of research to compare mortality predictive models of sepsis patients based on machine learning AUC: area under curve; ROC: Receiver operacting characteristic; ICU: intensive care unt; LoS: length of stay"

Figure 2.

The patient selection processes for data extraction from the MIMIC-Ⅳ database."

Table 1.

Features of raw data with missing ratio>5% obtained from MIMIC-IV (n=19,903)"

Figure 3.

Performance of models at different number of features. Three lines represent AUC scores in different models with top n features selected by MIC in (A), with different number of bins for continuous features discretization in (B). The color shadow area represents the fluctuation of AUC in different folds. The red dashed line represents the feature numbers we chose. LR: Logistic Regression; RF: Random Forest; BN: Bayesian Network."

Table 2.

MIC values of the features in feature selection."

Table 3.

Comparisons of features between survived and dead patients (n=12,664)"

Table 4.

Performance comparison of five machine learning models and SAPS Ⅱ score predictor"

Figure 4.

The ROC of the five the machine learning models and SAPS II score prediction. AUC: area under the curve; ROC: receiver operating charactoristic."

Figure 5.

Feature explanation of lightGBM model using the SHapley Addictive exPlanations (SHAP) value Each point in the figure is the SHAP value of a sample, and the color of the point represents the value of the feature. Red represents high values while blue represents low values. SHAP value greater than 0 indicates that the feature is a risk factor for death, while less than 0 indicates a protective factor."