Development and external validation of a machine learning model for predicting 28-day mortality in patients with acute myocardial infarction complicated by malignant arrhythmia: a study using the MIMIC database and a Chinese cohort

Figures

• 

  Figure 1. Feature selection using LASSO regression and the boruta algorithm (A) Ten-fold cross-validation curve for tuning the penalty parameter (λ) in the LASSO model. The x-axis represents log (λ), and the y-axis shows the binomial deviance (error). The solid red line indicates the cross-validated error curve, with the grey error bars representing the 95% confidence intervals. The two vertical dashed lines indicate the λmin and λ1se values. At λ1se = 0.03409493, the model is optimally regularized, identifying a total of 15 clinically relevant variables with non-zero coefficients (Supplementary Table S4). (B) Coefficient shrinkage paths for the LASSO model under varying penalty strengths. The x-axis is log(λ), and the y-axis shows the coefficient profiles. Each curve represents a distinct clinical feature. As the penalty increases (moving from left to right), the coefficients shrink toward zero, retaining only the most influential predictors. (C) Results of the Boruta feature-importance analysis. Boxplots compare the importance distribution (y-axis) of each clinical variable with its corresponding shadow features (blue boxes) generated via random forest iterations. The green boxes…

• 

  Figure 2. Performance of the machine learning-based modelling frameworks in the validation set. (A) Receiver operating characteristic (ROC) curves showing the discriminative ability of the seven machine learning models. (B) Calibration curves comparing the model-predicted probabilities (x-axis) to the true observed outcomes (y-axis). Brier scores with 95% CIs are provided in the legend; lower scores indicate better calibration. (C) Decision curve analysis (DCA) illustrating the net clinical benefit (y-axis) of the machine learning models across different threshold probabilities (x-axis).

• 

  Figure 3. External validation and performance of the machine learning framework. (A) ROC curve. (B) Calibration curve. (C) DCA curve. (D) Confusion matrix.

• 

  Figure 4. SHAP interpretation of the optimal SVM model on the validation set. (A) SHAP Importance Plot showing showing the mean absolute SHAP value for each predictor, ranked by importance. (B) SHAP summary (beeswarm) plot. Each dot represents a single patient's SHAP value for a particular feature. The x-axis shows the SHAP value, indicating how much a feature's presence or absence changes the model's output prediction for that instance. Red dots indicate high feature values, while blue dots indicate low feature values. (C) SHAP dependence (scatter) plots illustrating the relationship between a feature's value and its corresponding SHAP value.

• 

  Figure 5. SHAP analysis results. (A, B) SHAP waterfall plots were used to explain personalized feature attributes for one patient. (C, D) SHAP force plots were used to explain personalized feature attributes for another patient.