Risk stratification of patients with syncope in the emergency department using ECG based artificial intelligence models

Figures

• 

  Fig. 1. CONSORT diagram of the derivation of the study population. A total of 65,389 patients were eligible for inclusion in this study. Of these a total of 25,654 were excluded due to pacemaker implantation prior to syncope (n = 2869), immigration (n = 6453), age < 18 years (n = 3840), cancer (n = 8305), chronic obstructive pulmonary disease (n = 2662), anemia (n = 1525). The final study population of 39,735 were divided into training (n = 25,430), validation (n = 6358) and holdout test (n = 7947) sets.

• 

  Fig. 2. Comparison of the receiver operating characteristic curves for predicting cardiovascular death within 100 days for the neural network, logistic regression, random forest, k-nearest neighbor, and XG Boost model. The legend reports the area under the receiver operating characteristic curve (AUC-ROC) for each model.

• 

  Fig. 3. Precision-Recall curves for predicting cardiovascular death within 100 days illustrating the tradeoff between precision and recall in the diagnostic performance of the neural network, logistic regression, random forest, k-nearest neighbor, and XG Boost model. The legend reports the area under the precision-recall curve (AUC-PR) for each model.

• 

  Fig. 4. Bar plots showing the mean absolute SHAP (SHapley Additive exPlanations) values for features contributing to the prediction of cardiovascular death across the neural network, logistic regression, random forest, and XG Boost model. Each bar represents the average magnitude of a feature’s impact on model predictions, indicating its overall importance. The figure highlights differences in which features influence cardiovascular death risk across model types.

• 

  Fig. 5. Kaplan-Meier survival curves showing the probability of cardiovascular death over time, stratified by predicted risk groups from the neural network, logistic regression, random forest, and XG Boost model. Patients were classified into high- and low-risk groups based on model-predicted probabilities. Differences between high- and low-risk groups were assessed using the log-rank test, and hazard ratios were estimated with Cox regression.

• 

  Fig. 6. Cumulative incidence curves of 1-year cardiovascular diagnoses stratified by high- and low-risk groups from the logistic regression model. Differences between high- and low-risk groups were assessed using the log-rank test,