@Article{jimh.2026.075201,
AUTHOR = {Vaskar Chakma, Xiaolin Ju, Heling Cao, Xue Feng, Xiaodong Ji, Haiyan Pan, Gao Zhan},
TITLE = {CardioForest: An Explainable Ensemble Learning Model for Automatic Wide QRS Complex Tachycardia Diagnosis from ECG},
JOURNAL = {Journal of Intelligent Medicine and Healthcare},
VOLUME = {4},
YEAR = {2026},
NUMBER = {1},
PAGES = {37--86},
URL = {http://www.techscience.com/JIMH/v4n1/65616},
ISSN = {2837-634X},
ABSTRACT = {Wide QRS Complex Tachycardia (WCT) is a life-threatening cardiac arrhythmia requiring rapid and accurate diagnosis. Traditional manual ECG interpretation is time-consuming and subject to inter-observer variability, while existing AI models often lack the clinical interpretability necessary for trusted deployment in emergency settings. We developed CardioForest, an optimized Random Forest ensemble model, for automated WCT detection from 12-lead ECG signals. The model was trained, tested, and validated using 10-fold cross-validation on 800,000 ten-second-long 12-lead Electrocardiogram (ECG) recordings from the MIMIC-IV dataset (15.46% WCT prevalence), with comparative evaluation against XGBoost, LightGBM, and Gradient Boosting models. Performance was assessed using accuracy, balanced accuracy, precision, recall, F1-score, ROC-AUC, RMSE, and MAE. SHAP (SHapley Additive exPlanations) analysis provided feature-level interpretability to ensure clinical validity. CardioForest achieved superior and consistent performance across all metrics: test accuracy 95.19% (<mml:math id="mml-ieqn-1"><mml:mo>±</mml:mo></mml:math>0.33%), balanced accuracy 88.76% (<mml:math id="mml-ieqn-2"><mml:mo>±</mml:mo></mml:math>0.79%), precision 95.26%, recall 78.42%, F1-score 86.02%, and ROC-AUC 0.8886, with the lowest error rates (RMSE: 0.2532, MAE: 0.1944). Statistical significance testing confirmed CardioForest’s advantages over XGBoost (accuracy <mml:math id="mml-ieqn-3"><mml:mo>+</mml:mo><mml:mn>6.75</mml:mn><mml:mi mathvariant="normal">%</mml:mi></mml:math>, <i>p</i> &lt; 0.001), LightGBM (accuracy <mml:math id="mml-ieqn-4"><mml:mo>+</mml:mo><mml:mn>10.85</mml:mn><mml:mi mathvariant="normal">%</mml:mi></mml:math>, <i>p</i> &lt; 0.001), and Gradient Boosting (accuracy <mml:math id="mml-ieqn-5"><mml:mo>+</mml:mo><mml:mn>2.69</mml:mn><mml:mi mathvariant="normal">%</mml:mi></mml:math>, <i>p</i> = 0.012). Notably, CardioForest demonstrated exceptional stability (coefficient of variation: 0.35%) compared to competing models. SHAP analysis revealed that QRS duration-the primary clinical diagnostic criterion—dominated model predictions (mean SHAP value: 0.45), with additional contributions from QRS morphology and axis measurements, perfectly aligning with established cardiological knowledge. CardioForest represents a clinically validated, interpretable AI solution for WCT detection that balances diagnostic accuracy with transparent decision-making. With inference times under 10 milliseconds and comprehensive explainability through SHAP visualizations, the model is deployment-ready for real-time emergency department screening. By providing cardiologists with both accurate predictions and clinically interpretable feature attributions, CardioForest addresses the critical gap between AI performance and clinical trust, offering a practical tool for timely, evidence-based cardiac diagnosis in high-stakes scenarios.},
DOI = {10.32604/jimh.2026.075201}
}