Machine learning classifiers using echocardiographic parameters to predict shock occurrence in severe dengue: A secondary analysis of a prospective cohort

Abstract

Background: Shock is the most critical complication of dengue, yet early risk stratification remains challenging. This study performed a secondary analysis of a prospective cohort to evaluate machine learning (ML) models for predicting shock and to identify the most influential clinical and haemodynamic predictors.

Results: Eighty-six patients were included, of whom 67 (77.9%) developed shock. In basic features only Model, performance was limited, indicated by AUC ≤0.68. Addition of demographic and haemodynamic features to the model showed the best balance of discrimination and calibration, with Random Forest achieving an AUC of 0.80 and accuracy of 0.80. Combining all features yielded slightly higher accuracy (up to 0.82 with Random Forest) but no clear calibration advantage over Model 2. Feature importance analyses ranked haemodynamic parameters, such as stroke volume, stroke volume index, cardiac index, and aortic VTI, as the most influential predictors.

Methods: Data from patients admitted with dengue shock syndrome at the Hospital for Tropical Diseases, Ho Chi Minh City, were reanalyzed. Demographic, vital sign, haemodynamic, and echocardiographic variables were included. Shock occurrence served as the binary outcome. Seven ML classifiers (Decision Tree, Random Forest, XGBoost, LightGBM, AdaBoost, Gradient Boosting, Naive Bayes) were trained using stratified 10-fold cross-validation. Performance was assessed by accuracy, AUC, sensitivity, specificity, F1 score, and RMSE. Feature importance was analyzed to explore mechanistic relevance.

Conclusion: ML classifiers showed moderate ability to predict shock in dengue, with Random Forest, LightGBM, and XGBoost performing best. A parsimonious model using demographic and haemodynamic features achieved calibration similar to more complex models and was supported by feature importance findings.