A retrospective study of mortality for perioperative cardiac arrests toward a personalized treatment

Figures

• 

  Figure 1. ROC curves for the six ML models on the test data. The AUC value of each model is represented by “(AUC = mean ± standard deviation)”, which was estimated from 1000 bootstrap resamples of predictions on the test data. Each ROC curve is visualized by corresponding plot with shaded bands.

• 

  Figure 2. Probability curves for each ML model. Survivors indicated in green, and non-survivors in red. p < 0.005 for ensemble versus other models.

• 

  Figure 3. SHAP importance plots of the mortality and risk factors for the ensemble ML model (VotingClassifier). The features are ranked by importance. Each row represents the impact of a feature on the outcome of mortality, with higher SHAP values indicating higher likelihood of a positive outcome. For a binary feature, like gender, “male” → “1” is shown in red while “female” → “0” is shown in blue. For the detailed mapping of categorical features, please refer to the code online (such as “ < 12 ys” → “0”, “12 ~ 40 ys” → “1”, “40 ~ 65 ys” → “2”, “ > 65 ys” → “3”).

• 

  Figure 4. LIME explainer for four typical scenarios. (a) True positive, patient died, i.e., a correctly classified non-survivor, (b) True negative, patient survived, i.e., a correctly classified survivor, (c) False positive, patient survived, i.e., an incorrectly classified survivor (predicted to die), and (d) False negative, patient died, i.e., an incorrectly classified non-survivor (predicted to survive). Features with a green bar favored survival, and those with a red bar were predictive of mortality. The x-axis shows how much each feature added to or subtracted from the final probability value for the patient. Each weight can be interpreted in the context of the original probability; if a feature is absent for a patient, it can be numerically added to or subtracted directly from the initial probability.