Development and validation of machine learning-based prediction model for outcome of cardiac arrest in intensive care units

Figures

• 

  Fig. 1. Flow chart of patient selection.

• 

  Fig. 2. Comparison of the model performance based on different algorithms. RF, Random Forest; SVM, Support Vector Machine; MLP, Multi-Layer Perceptron; LR, Logistic Regression; GBDT, Gradient Boosting Decision Tree; XGBoost, eXtremely Gradient Boosting; KNN, K-Nearest Neighbor; AUC, areas under receiver operating characteristic curves.

• 

  Fig. 3. Comparison of the model performance in internal validation. (A) ROC of all of the models. (B) ROC full model vs. compact model. (C) ROC compact model vs. SOFA.

• 

  Fig. 4. Clinical practicability of the model. (A) Calibration plot of the CatBoost model and LR. (B) Decision curve analysis of the CatBoost model and LR.

• 

  Fig. 5. SHAP summary plot of the model. The plot was based on 72-hour variables. The colors of dots ranging from blue to red represent the attribution value, where blue corresponds to a lower feature value and red to a higher value. According to the overall contribution value, the importance of the variables was ranked as WBC, heart rate, bicarbonate, MBP, age, chlorine, DBP, platelets, sodium, heart failure, and creatinine.

• 

  Fig. 6. SHAP force plots of the model. The plots were based on 72-h variables. Each variable is represented by an arrow, with red indicating a positive effect on death and blue a negative effect. The size of the arrow indicates the importance of the variable in predicting mortality, where a larger arrow indicates a greater impact. The combined effects of all of the variables contribute to the predicted output value of the model. (A) The patient died at discharge with an output predictive value of 1.54. (B) The patient survived at discharge with an output predictive value of 0.42.

• 

  Fig. 7. Model performance in the external validation.