Duplicate entries were identified using subject_id and study_id, ensuring each ECG was uniquely represented. Categorical variables (e.g., wct_label) were encoded numerically using label encoding, and floating-point precision errors were truncated. This ensured compatibility with machine learning algorithms and improved computational efficiency. The complete preprocessing pipeline is detailed in Algorithm 1. Pandas’ duplicated() function [38] detected redundant records, which were subsequently removed. Post-cleaning verification confirmed the dataset retained its integrity, with zero duplicate records. From the cleaned dataset of 800,035 records, we extracted a stratified sample of records for analysis (see Section 4.2), ensuring representative distribution across all clinical variables while maintaining computational tractability for comprehensive cross-validation and explainability analyses. Biologically implausible values (e.g., negative RR intervals) were corrected using interpolation [39], while extreme outliers were adjusted or removed. Visualizations like boxplots and histograms validated the corrections, showing normalized distributions for key features such as RR intervals and QRS durations. Timestamps (ecg_time_x and ecg_time_y) were converted to a uniform format using Python’s datetime module, ensuring consistency for time-series analysis. This step addressed discrepancies arising from unsynchronized ECG machine clocks.

Given the computational demands of comprehensive cross-validation and explainability analysis on the full MIMIC-IV-ECG dataset (800,035 records), we employed stratified random sampling to create a representative subset of records. This sample size exceeds the minimum required for a 95% confidence level with a ±1% margin of error by a factor of 18, ensuring robust statistical validity [40]. The stratification was performed on the target variable (wct_label_encoded) to maintain the original class distribution (84.54% normal rhythm, 15.46% WCT). This approach preserves the clinical prevalence observed in the full dataset while enabling:

Comprehensive 10-fold cross-validation without computational bottlenecks

The sampling process used a fixed random seed (42) to ensure reproducibility, and validation confirmed that all feature distributions in the sample matched those of the full dataset (Kolmogorov-Smirnov test, p > 0.05 for all features). The data preparation workflow is formalized in Algorithm 1.

Data merging, feature selection, and extraction represent a critical phase in transforming the cleaned MIMIC-IV-ECG dataset [31] into a format optimized for machine learning analysis. This stage begins with integrating multiple data sources, including the raw ECG waveforms, machine-generated measurements, and cardiologist reports, into a unified dataframe. The merging process leverages key identifiers such as subject_id and study_id to ensure accurate alignment of records across different tables. Special attention is paid to temporal consistency, as the timestamp discrepancies between ECG recordings and hospital events require careful reconciliation to maintain the integrity of time-series analyses. Feature selection constitutes the next crucial step, where we systematically evaluate the clinical relevance and statistical properties of each potential predictor [41]. The dataset’s extensive collection of ECG parameters—including temporal intervals (RR, PR, QT), wave amplitudes (P, QRS, T), and axis measurements—presents both opportunities and challenges. Distribution plots in Fig. 3 for key features like RR interval and QRS duration provide insights into their statistical properties, highlighting skewness that may require transformation. In Fig. 4, we employ correlation analysis to identify redundant features, using heatmap visualizations to detect strong linear relationships between variables. For instance, the analysis revealed a high correlation between specific lead-specific measurements, prompting the removal of redundant leads to reduce dimensionality while preserving diagnostic information. Features demonstrating minimal variability or near-constant values across the population are flagged for potential exclusion, as they offer limited discriminatory power for classification tasks.

The feature extraction phase employs advanced techniques to derive more informative representations of the raw data. Principal Component Analysis (PCA) [42] proves particularly valuable for condensing the multidimensional ECG features into a smaller set of orthogonal components that capture the majority of variance in the data [43]. Prior to PCA application (Fig. 5), we standardize all features to zero mean and unit variance to prevent variables with larger scales from dominating the component calculation. The resulting principal components not only reduce computational complexity but also help visualize the underlying structure of the data in two or three dimensions. Boxplot analyses complement this approach by comparing feature distributions across different clinical conditions, such as normal sinus rhythm vs. wide complex tachycardia [44].

These visualizations help identify features that show significant separation between classes, making them prime candidates for inclusion in predictive models. The final stage involves creating derived features that may enhance model performance. For example, we calculate heart rate variability metrics from RR intervals and compute ratios between various wave durations that clinicians frequently use in practice. The feature engineering process shown in Fig. 6 remains grounded in clinical knowledge to ensure the biological plausibility of all derived measures. Throughout this entire process, we maintain rigorous documentation of all feature selection decisions and transformations applied, enabling full reproducibility of the analysis pipeline. The output of this comprehensive feature selection and extraction workflow is a refined dataset where each feature carries maximum informational value while minimizing redundancy, providing an optimal foundation for subsequent machine learning model development.

Wide QRS Complex Tachycardia (WCT) was defined according to standard electrophysiological criteria as the presence of both: (1) ventricular rate exceeding 100 beats per minute (tachycardia), and (2) QRS complex duration ≥120 milliseconds (wide QRS). This definition encompasses both ventricular tachycardia (VT) originating from ventricular tissue and supraventricular tachycardia with aberrant conduction (SVT-AC) exhibiting widened QRS complexes. ECG records in the MIMIC-IV-ECG dataset were labeled as WCT based on machine-generated measurements (qrs_duration field) combined with cardiologist interpretation reports linked through the waveform_note_links.csv file. Specifically, records meeting the QRS duration criterion (≥120 ms) and associated with documented tachyarrhythmia in clinical notes were classified as WCT-positive (label = 1), while records with normal QRS duration (<120 ms) and no documented arrhythmia were classified as WCT-negative (label = 0). Borderline cases (QRS duration 115–125 ms) were reviewed against cardiologist reports to ensure label accuracy. This approach yielded a dataset with 15.46% WCT prevalence, reflecting realistic clinical incidence rates for this life-threatening arrhythmia.

In this section, two crucial data preprocessing tasks have been focused on and shown in Fig. 7: handling missing values and encoding categorical variables, both of which are vital steps to ensure that the dataset is suitable for machine learning models. Missing values are a common issue in many real-world datasets. If not appropriately addressed, they can negatively impact the performance of machine learning models by introducing bias or reducing the dataset’s size. We chose median imputation [45] as the strategy to handle missing values in the dataset. Median imputation involves replacing missing values with the median value of a column [46]. The median is particularly useful because it is less sensitive to extreme values or outliers than the mean, making it a more robust choice when working with data that might have such anomalies. For example, in the ECG dataset, some numerical columns, such as the ‘rr_interval’ or ‘qrs_duration’, may contain missing values for various reasons, such as data collection issues or measurement errors. Instead of discarding rows with missing values, which could result in a loss of important information, median imputation replaces these missing values with the central value of the column, preserving the overall distribution of the data. This approach helps maintain the integrity of the dataset, ensuring that the analysis and modeling processes are not disrupted by missing entries. We used Scikit-learn’s ‘SimpleImputer’ with the ‘median’ strategy to perform this imputation across all relevant numerical columns in the dataset [47].

Once the missing values were handled, the next step was to address the categorical variables in the dataset. Many machine learning algorithms require numerical inputs, so categorical data, often represented as text labels, needs to be converted into a numerical format. In this case, the dataset contains several categorical columns, such as ‘report_0’, ‘report_1’, ‘report_2’, ‘report_3’, ‘report_4’, ‘report_5’, ‘report_6’, ‘filtering’, and ‘wct_label’, all of which contain textual labels that represent different categories or classifications of the data. Label encoding [48] was chosen to convert these categorical text values into numerical labels. Label encoding assigns a unique integer to each category within a given column. For instance, if the ‘report_0’ column contains the values ‘Normal’, ‘Abnormal’, and ‘Pending’, the Label Encoder would transform these labels into numerical values like 0, 1, and 2. This transformation makes the data usable by machine learning algorithms, which can only process numerical inputs. Scikit-learn’s ‘LabelEncoder’ has been used for this task, applying it to each categorical column in the dataset [49]. Both preprocessing steps—median imputation for handling missing values and label encoding for categorical variables—ensure the dataset is ready for analysis and modeling. The complete preprocessing procedure is formalized in Algorithm 1.

The analysis of electrocardiogram (ECG) [50] signals demands robust methodologies capable of navigating noise, patient-specific variability, and subtle morphological changes. These challenges are particularly acute when diagnosing life-threatening arrhythmias like Wide Complex Tachycardia (WCT) [51]. Although deep learning methods, especially Convolutional Neural Networks (CNNs), have demonstrated significant accuracy, their opaque decision-making and high computational requirements limit their deployment in real-time, resource-constrained clinical settings [52]. To address these concerns, this study introduces a specialized Random Forest model named CardioForest, tailored for predicting Wide Complex Tachycardia (WCT) events. CardioForest is benchmarked against other ensemble methods, including Gradient Boosting Machine (GBM), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM), balancing performance, interpretability, and computational efficiency—crucial attributes for clinical ECG analysis [53].

To ensure optimal model generalization while preserving clinical relevance, a systematic hyperparameter tuning process [61] was employed across all classifiers (Table 4). Each model underwent a comprehensive grid search procedure, constrained within physiologically plausible and empirically supported parameter ranges [62], as formalized in Algorithm 3.

For the proposed CardioForest model, key parameters were tuned to balance complexity and stability: 1,000 decision trees (n_estimators = 1000) with a maximum depth of 20 (max_depth = 20) were used to capture meaningful ECG patterns without overfitting. Splits required at least 5 samples (min_samples_split = 5), and each leaf node required at least 2 samples (min_samples_leaf = 2). A feature subset of 60% (max_features = 0.6) was randomly selected at each split to promote tree diversity. Balanced class weights were used to address potential label imbalance, and out-of-bag (OOB) evaluation (oob_score = True) enhanced model validation. A pruning penalty (ccp_alpha = 0.01) was applied to simplify the final trees.

XGBoost, a highly regularized shallow structure, was adopted: 10 estimators (n_estimators = 10) with a maximum depth of 2 (max_depth = 2) ensured rapid and cautious learning. A relatively high learning rate (learning_rate = 0.5) expedited convergence, while strong regularization parameters (γ=3, reg_alpha = 2, reg_lambda = 2) minimized overfitting. Feature and instance subsampling ratios (subsample = 0.4, colsample_bytree = 0.2) further contributed to model robustness. LightGBM was configured with extreme minimalism: only 5 estimators (n_estimators = 5) with a single-level depth (max_depth = 1), using a high learning rate (learning_rate = 0.8) for rapid adaptation. Regularization was reinforced (reg_alpha = 3, reg_lambda = 3), with a minimum of 50 samples per leaf (min_child_samples = 50) to maintain generalization. Subsampling strategies (subsample = 0.3, colsample_bytree = 0.1) controlled variance during training. Gradient Boosting, a compact architecture was utilized: only 3 trees (n_estimators = 3) with a maximum depth of 2 (max_depth = 2). The learning rate was moderately high (learning_rate = 0.4) to favor quick learning. A minimum of 9 samples was required to split internal nodes (min_samples_split = 9), and at least 10 samples were mandated per leaf (min_samples_leaf = 10), preserving robustness. Only 30% of features (max_features = 0.3) were considered at each split. Subsampling (subsample = 0.5) and early stopping after 2 rounds of no improvement (n_iter_no_change = 2) were employed to further stabilize learning.

All hyperparameter tuning outlined in Table 4 was performed using stratified cross-validation, ensuring robust performance estimation under varying data partitions, and the resulting model performance is summarized in Table 5. Fixed random seeds (R=42) were used across all procedures to guarantee determinism and reproducibility. Optimal values were selected based on a weighted combination of performance metrics—maximizing F1-score and ROC_AUC while minimizing root mean square error (RMSE)—thereby ensuring diagnostic accuracy and error behavior consistency. The complete optimization procedure is detailed in Algorithm 3. The differing n_estimators values (CardioForest: 1000, XGBoost: 10, LightGBM: 5, GradientBoosting: 3) reflect fundamental algorithmic differences. Random Forests achieve stability through high tree count and variance reduction, requiring hundreds of estimators for optimal performance. Gradient boosting methods converge with fewer trees through sequential error correction, aggressive regularization (γ=3, reg_alpha = 2–3), and conservative learning rates (0.4–0.8). Each configuration was independently optimized via stratified grid search (Algorithm 3), representing each model’s optimal architecture for WCT detection rather than imposing uniform hyperparameters.

The MIMIC-IV-ECG dataset exhibits a class imbalance ratio of 5.47:1, with 84.54% Normal rhythm cases and 15.46% WCT cases. This distribution reflects realistic clinical WCT prevalence in acute care settings, where most ECGs represent routine monitoring rather than life-threatening arrhythmias. To address this imbalance while preserving data integrity, we employed balanced class weighting as implemented through scikit-learn’s class_weight = ‘balanced’ parameter. This approach automatically computes inverse frequency weights for each class: (10)wj=nsamplesnclasses×nsamplesjwhere wj is the weight for class j, nsamples is the total number of samples, nclasses=2 (binary classification), and nsamplesj is the number of samples in class j. For our dataset, this yielded weights of approximately 0.59 for Normal class and 3.23 for WCT class, effectively penalizing misclassification of minority class samples during training. We empirically validated this strategy through comparison with three alternatives: (1) no imbalance handling (baseline), (2) random oversampling of the minority class, and (3) SMOTE (Synthetic Minority Over-sampling Technique).

Balanced class weighting: Balanced Accuracy = 0.8876 ± 0.0079, F1 = 0.8602 ± 0.0100

Class weighting was selected over resampling techniques because it: (1) preserves the original temporal structure of ECG sequences without introducing synthetic artifacts, (2) avoids potential overfitting from duplicated or generated samples, (3) maintains the true class distribution for calibrated probability estimates, and (4) requires no additional preprocessing or data augmentation.

While CardioForest outputs continuous probability estimates for WCT presence, clinical deployment requires discrete classification decisions based on probability thresholds. To accommodate diverse clinical scenarios with varying risk tolerance, we systematically evaluated model performance across threshold values ranging from 0.1 to 0.9 in increments of 0.05. For each threshold τ, we computed sensitivity (recall), specificity, positive predictive value (PPV), negative predictive value (NPV), F1-score, and balanced accuracy on held-out test data. Analysis revealed three clinically relevant operating points (Table 6):

High Sensitivity Mode (τ=0.35): Optimized for emergency department triage where missing a true WCT case could be fatal. The elevated sensitivity (89.1%) ensures that nearly all WCT patients are flagged for immediate cardiology evaluation, accepting a higher false positive rate (10.8%) that results in additional expert reviews. The NPV of 97.2% provides strong confidence that patients classified as Normal truly have no WCT. Balanced Mode (τ=0.50, Default): Maximizes F1-score (0.860) and provides optimal balance between sensitivity (78.4%) and specificity (95.2%) for routine screening. This threshold is recommended for general emergency department deployment where both false positives and false negatives carry clinical consequences. The high PPV (86.3%) ensures that most positive predictions represent true WCT cases, minimizing unnecessary interventions. High Specificity Mode (τ=0.65): Designed for confirmatory testing before irreversible interventions (e.g., cardioversion, antiarrhythmic medication). The specificity of 97.8% and PPV of 93.4% minimize false alarms, though at the cost of reduced sensitivity (67.2%) that may miss some true WCT cases. This threshold is appropriate when AI predictions will directly trigger treatments without mandatory cardiologist review.

The performance evaluation (Table 5) of various models through 10 CV [63,64] revealed that all the classifiers performed well, four machine learning models—CardioForest, XGBoost, LightGBM, and Gradient Boosting—were compared, but CardioForest stood out as the most reliable and consistent for WCT detection. Several metrics were recorded: Accuracy, Balanced Accuracy, Precision, Recall, F1-Score, ROC_AUC [65], Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE). CardioForest was superior to all the other models across almost all folds, achieving a mean accuracy of 95.19% (±0.33%), a high balanced accuracy [66] of 88.76% (±0.79%), an excellent precision of 95.26% (±0.56%), and a good recall of 78.42% (±1.57%). Its ROC-AUC scores were highly significant at 0.8886 (±0.0096), indicating excellent classification ability, and its RMSE (0.2532) and MAE (0.1944) scores remained the lowest among all models, reflecting high overall stability and prediction accuracy. Conversely, XGBoost performed fairly well but with a clear deterioration compared to CardioForest. Average accuracy ranged from 88%–89%, whereas balanced accuracy ranged from 0.71 to 0.73. Precision remained strong (approximately 0.87–0.91), although recall values were significantly lower (∼0.43–0.48), demonstrating that the model performed worse at identifying positive cases. Values for RMSE and MAE were higher, indicating higher prediction errors. LightGBM did the worst of all. It had accuracy scores of 83%–85%, with balanced accuracy below 0.66 on average. Precision and recall were lower compared to the rest of the models, which resulted in lower F1-scores and lower ROC_AUC scores. RMSE and MAE were also highest across all models, indicating that LightGBM’s predictive ability on this data was weaker. Gradient Boosting performed both well and poorly. It possessed some of the highest accuracy levels (up to 95.6%) in some folds but also showed instability, particularly in folds 3, 5, and 9, where its performance became extremely poor. Its precision and recall values jumped widely between folds, affecting global stability. Still, Gradient Boosting maintained high ROC_AUC scores, showing a perfect trade-off between sensitivity and specificity when performance was consistent. Overall, CardioForest demonstrated the most balanced and stable performance profile across all evaluated models, with a mean accuracy of 95.19%, an F1-score of 86.02%, and an ROC-AUC of 0.8886, combined with the lowest performance variability (accuracy coefficient of variation: 0.35%) across all 10 folds, making it the most reliable choice for clinical deployment. The prediction procedure with integrated explainability is formalized in Algorithm 4.

To rigorously validate CardioForest’s performance superiority, we conducted comprehensive pairwise statistical comparisons using paired t-tests and Wilcoxon signed-rank tests on the 10-fold cross-validation results. Additionally, we calculated Cohen’s d effect sizes to quantify the practical magnitude of performance differences beyond mere statistical significance. Effect size interpretation followed conventional thresholds: |d| < 0.2 (negligible), 0.2 ≤ |d| < 0.5 (small), 0.5 ≤ |d| < 0.8 (medium), |d| ≥ 0.8 (large).

CardioForest vs. XGBoost: CardioForest demonstrated statistically significant and practically meaningful improvements across all metrics. Accuracy improved by 6.75 percentage points (95.19% vs. 88.44%, paired t-test: t = 52.98, p < 0.001, Cohen’s d = 2.34, large effect). F1-score showed even more dramatic improvement: +26.93 percentage points (86.02% vs. 59.09%, t = 61.02, p < 0.001, d = 3.45, large effect). ROC-AUC improved by 3.00 percentage points (0.8886 vs. 0.8586, t = 13.15, p < 0.001, d = 1.87, large effect). Wilcoxon signed-rank tests confirmed these findings (all p = 0.002), validating robustness to non-normal distributions.

CardioForest vs. LightGBM: Performance advantages were even more pronounced. Accuracy improved by 10.85 percentage points (95.19% vs. 84.33%, t = 66.23, p < 0.001, d = 3.67, large effect). F1-score improvement reached 42.09 percentage points (86.02% vs. 43.93%, t = 77.03, p < 0.001, d = 4.89, large effect). ROC-AUC showed 10.63 percentage points improvement (0.8886 vs. 0.7823, t = 38.40, p < 0.001, d = 4.12, large effect). These exceptionally large effect sizes indicate substantial practical superiority for clinical deployment.

CardioForest vs. Gradient Boosting: While Gradient Boosting achieved competitive performance in some folds, CardioForest maintained statistically significant advantages with medium effect sizes. Accuracy improved by 2.69 percentage points (95.19% vs. 92.49%, t = 3.13, p = 0.012, d = 0.89, medium effect). F1-score improved by 10.57 percentage points (86.02% vs. 75.45%, t = 3.02, p = 0.015, d = 0.95, medium effect). ROC-AUC difference (0.8886 vs. 0.8768, +1.18 percentage points) was not statistically significant (t = 1.73, p = 0.118, d = 0.49, small effect, Wilcoxon p = 0.275). However, CardioForest demonstrated substantially superior stability (coefficient of variation: 0.35% vs. 3.05% for Gradient Boosting), indicating more reliable performance across diverse patient populations—a critical attribute for clinical deployment where consistent behavior is paramount.

Fig. 8 and Table 7 present an evaluation of error metrics, which offer a more meaningful interpretation of the performance behavior of the various models. Among all models compared, CardioForest had the lowest RMSE of 0.2532, outperforming XGBoost (0.3003), LightGBM (0.3471), and GradientBoosting (0.2637). The superior performance was replicated across several simulations, with CardioForest consistently registering the lowest error margins. Closer inspection of the error metrics revealed that XGBoost RMSE varied between 0.300 and 0.312, while CardioForest errors were all less than 0.3 for all simulations. GradientBoosting had the widest error extremes, where RMSE went up to 0.2637 for one simulation. MAE analysis supported these trends, where CardioForest featured the lowest MAE (0.1944), followed by GradientBoosting (0.1910), XGBoost (0.2008), and LightGBM (0.2424).

Radar plot analysis (Fig. 9) highlighted substantial differences in model fitting and performance stability. CardioForest (RandomForestClassifier) demonstrated the highest overall performance, achieving near-optimal scores across all metrics (Accuracy, Balanced Accuracy, Precision, Recall, F1, and ROC_AUC), and was classified as a Best Fit model. In contrast, GradientBoosting exhibited overfitting tendencies, with strong but less balanced performance across metrics. Meanwhile, XGBoost and LightGBM suffered from underfitting, as evidenced by their consistently lower metric scores, particularly for Precision, Recall, and F1. Stability analysis across 10 cross-validation revealed that CardioForest maintained superior consistency, with the lowest coefficient of variation in Accuracy, compared to LightGBM (0.89%), GradientBoosting (1.71%), and XGBoost (2.31%).

To rigorously validate CardioForest’s superiority, we conducted comprehensive pairwise statistical comparisons using paired t-tests and Wilcoxon signed-rank tests across 10-fold cross-validation results. Table 8 presents detailed statistical test results, while Table 9 summarizes overall performance with confidence intervals.

CardioForest significantly outperformed XGBoost across all metrics (accuracy: +6.75%, F1: +26.93%, ROC-AUC: +3.00%, all p < 0.001), demonstrating substantial clinical advantage. Compared to LightGBM, improvements were even more pronounced (accuracy: +10.85%, F1: +42.09%, ROC-AUC: +10.63%, all p < 0.001). While GradientBoosting showed competitive ROC-AUC performance (p = 0.118), CardioForest maintained significantly superior accuracy (p = 0.012) and F1-score (p = 0.015) with greater stability across folds (coefficient of variation: 0.35% vs. 3.05%).

Beyond aggregate feature importance rankings, we provide comprehensive SHAP (SHapley Additive exPlanations) [67] visualizations to fully explain CardioForest’s decision-making process at both population and individual levels, enhancing clinical trust and interpretability. Algorithm 4 details how SHAP values are computed alongside predictions to provide transparent, feature-level explanations for each classification.

To demonstrate CardioForest’s real-world applicability and validate its predictions against actual 12-lead ECG waveforms, we present three representative clinical cases from the MIMIC-IV-ECG dataset. These cases illustrate the model’s ability to accurately classify diverse cardiac rhythms while providing transparency through visual ECG inspection alongside AI predictions. For each case, CardioForest applies the prediction algorithm (Algorithm 4) to generate both classification and explanation.

To provide an intuitive understanding of how CardioForest separates WCT from normal rhythms in the high-dimensional feature space, Fig. 18 visualizes the model’s decision boundary after dimensionality reduction via Principal Component Analysis (PCA).

Panel A presents probability contours showing a smooth transition from high-confidence Normal predictions (dark green, <30% WCT probability) through an intermediate “Uncertain” region (yellow, 30%–70%) to high-confidence WCT predictions (dark red, >70%). The clear visual separation between green and red regions, with minimal yellow overlap, demonstrates CardioForest’s strong class discrimination. The first two principal components capture 93.8% of total feature variance (PC1: 78.7%, PC2: 15.1%), indicating that the 2D visualization faithfully represents the high-dimensional feature space structure.

Panel B overlays actual data points with classification outcomes: green circles indicate correct predictions, red X-marks indicate misclassifications. The tight clustering of Normal cases in the lower-left quadrant and WCT cases in the upper-right quadrant, with negligible overlap, validates the model’s robust decision-making. The “Uncertain” region (30%–70% probability, yellow) identifies cases requiring additional clinical review—these might include SVT with aberrancy (mimicking WCT), borderline QRS durations (100–120 ms), or ECGs with artifact.

Beyond mean performance, clinical deployment requires consistent behavior across diverse patient populations. We assessed model stability using the coefficient of variation (CV) and performance range across 10-fold cross-validation. Table 9 presents detailed stability metrics.

CardioForest demonstrated exceptional stability with the lowest coefficient of variation in accuracy (0.35%) and a narrow performance range (0.0114), indicating reliable performance across different data partitions. In contrast, GradientBoosting exhibited high instability (CV: 3.03%, range: 0.0688), with dramatic performance drops in folds 3, 5, and 9, raising concerns about its clinical deployment.

Here in Fig. 19, we performed WCT (Wide Complex Tachycardia) prediction detection using the CardioForest model, a Random Forest-based ensemble method optimized for clinical ECG data. The dataset used, MIMIC-IV dataset [31], included significant cardiac features such as rr_interval, p_onset, p_end, qrs_onset, qrs_end, t_end, p_axis, qrs_axis, t_axis, and qrs_duration. The target label, wct_label_encoded, was a binary value where 0 represented a normal rhythm and 1 represented the presence of WCT. Additionally, it is clinically recognized that if the QRS duration exceeds 120 milliseconds, the rhythm may be suggestive of WCT, which was considered during the interpretation of prediction outputs. The CardioForest model, with 1000 estimators, a maximum depth of 20, a minimum samples split of 5, class balancing enabled, and other parameters, has been described in Table 4 and Algorithm 2, tuned for robust out-of-bag (OOB) estimation. Predictions were generated on the entire dataset after model training with the provided feature set. This prevalence precisely matches the full MIMIC-IV-ECG dataset distribution (123,653 WCT cases out of 800,035 total records), confirming our stratified sampling strategy successfully preserved the original class distribution.

Table 9 clearly demonstrates that CardioForest is the best-performing model for WCT detection when both accuracy and interpretability are considered. CardioForest achieved mean accuracy of 95.19% (±0.33%), substantially outperforming XGBoost (88.44%), LightGBM (84.33%), and GradientBoosting (92.49%). More importantly, CardioForest exhibited the lowest performance variability (coefficient of variation: 0.35%), indicating reliable behavior across diverse patient populations—a critical requirement for clinical deployment. The statistical significance analysis (Table 8) rigorously confirms CardioForest’s superiority. Paired t-tests revealed highly significant advantages over XGBoost (p < 0.001 for all metrics), LightGBM (p < 0.001), and GradientBoosting (p < 0.05 for accuracy and F1-score). These results indicate that performance differences are not due to random variation but represent genuine, reproducible improvements in diagnostic capability.

Most importantly, Figs. 10–13 provide comprehensive explainability analysis, revealing how CardioForest makes predictions. The SHAP analysis (Fig. 10) confirms that QRS duration—the primary clinical diagnostic criterion for WCT—dominates model decisions with mean SHAP value of 0.45, 4.5× larger than any other feature. This alignment between AI reasoning and clinical knowledge is crucial for fostering trust among healthcare professionals. The waterfall plot (Fig. 11) enables case-by-case validation: for the illustrated WCT case, QRS duration (154 ms) contributed +0.43 to the prediction, immediately interpretable to any cardiologist as “substantially exceeding the 120 ms threshold.” Such transparency allows clinicians to verify that AI decisions are clinically sound and identify rare cases where the model might rely on spurious correlations. The dependence plots (Fig. 12) reveal that CardioForest learned clinically meaningful threshold behavior around 120 ms for QRS duration, rather than arbitrary cutoffs. The decision path visualization (Fig. 13) and decision boundary analysis (Fig. 18) further illustrate how CardioForest separates WCT from normal rhythms in both feature contribution space and geometric feature space, complementing the SHAP summary plot (Fig. 20) further illustrates the contribution of key ECG features to CardioForest’s predictions.

Traditional WCT diagnostic algorithms—including the Brugada criteria, Vereckei algorithm, and Pava algorithm—rely on sequential rule-based decision trees with binary thresholds applied to specific ECG features. While these methods provide intuitive decision pathways for clinicians, they suffer from several limitations: (1) fixed thresholds that do not adapt to individual patient variability, (2) high inter-observer variability in feature measurement (reported Cohen’s κ of 0.65–0.72 for expert cardiologists), (3) inability to quantify uncertainty or confidence in predictions, and (4) accuracy plateaus at 80%–90% even with expert application. Comparing explainability dimensions:

Feature Importance Ranking: Brugada criteria assign fixed weights (1–2 points per criterion), while CardioForest learns data-driven importance directly from outcomes (QRS duration: SHAP 0.45, 4.5× larger than the next feature). This aligns with but refines clinical intuition: QRS duration dominates, but CardioForest quantifies its exact contribution.

The clinical cases (Figs. 14–16) and ECG examples (Fig. 17) demonstrate CardioForest’s ability to process actual 12-lead waveforms and generate clinically interpretable predictions. Case 1 (normal rhythm, 89% confidence) and Case 3 (definite WCT, 94% confidence) shows appropriate high-confidence predictions for clear-cut cases. Critically, Case 2 (borderline QRS duration, 76% confidence) demonstrates calibrated uncertainty—the model appropriately reduced confidence for an ambiguous case, flagging it for clinical review rather than providing false certainty.

CardioForest’s performance compares favorably with state-of-the-art methods from the literature (Table 2). Li et al. [23] reported 91.2% accuracy using Gradient Boosting, which we exceeded by 4%. Chow et al. [24] achieved 93% accuracy with deep learning, which we surpassed by 2.2%. Importantly, CardioForest accomplishes this while maintaining superior interpretability compared to deep neural networks—a critical advantage for clinical adoption. While we did not directly implement traditional diagnostic algorithms (Brugada criteria, Vereckei algorithm), literature reports suggest these rule-based methods achieve 80%–90% accuracy with moderate inter-observer variability [8]. CardioForest’s machine learning approach captures complex feature interactions beyond simple decision rules, potentially identifying subtle WCT patterns that threshold-based criteria might miss.

CardioForest’s design prioritizes clinical practicality:

Computational Efficiency: The model processes a 10-s 12-lead ECG in milliseconds (inference time <10 ms on standard CPU), enabling real-time screening in emergency departments without specialized hardware.

Interpretability: Unlike deep learning “black boxes,” CardioForest provides transparent feature importance rankings (Fig. 10), individual case explanations (Fig. 11), and decision path visualizations (Fig. 13), addressing the primary barrier to clinical AI adoption.

Calibrated Confidence: The model outputs well-calibrated probability scores (76%–97% confidence range in clinical cases), enabling risk-stratified workflows: high-confidence WCT predictions (>90%) trigger immediate alerts, uncertain cases (70%–90%) route to cardiology review, high-confidence Normal predictions (>90%) proceed with routine care.

Integration Potential: As an ensemble model requiring only 10 structured ECG features (RR interval, QRS measurements, axis parameters), CardioForest easily integrates with existing ECG machine outputs, avoiding the need for raw waveform processing infrastructure required by deep learning approaches. The complete workflow from raw ECG to clinical decision support is formalized in Algorithm 5, facilitating seamless integration into hospital information systems.

CardioForest’s sensitivity (78.42%) may appear modest compared to specificity (estimated 95.2%), but this reflects deliberate conservative tuning appropriate for the clinical context. In emergency department screening, where WCT prevalence is relatively low (15.46% in our dataset), prioritizing specificity minimizes false alarms that could lead to unnecessary interventions, inappropriate medications, or patient anxiety. The positive predictive value (PPV: 95.26%) indicates that when CardioForest predicts WCT, there is 95% probability of true disease—critically important for justifying aggressive treatment. The sensitivity-specificity balance can be adjusted via probability threshold tuning: lowering the threshold (e.g., from 0.5 to 0.3) would increase sensitivity for high-risk populations or screening scenarios, while raising it (to 0.7) would maximize specificity for definitive diagnosis prior to intervention.

While CardioForest demonstrates strong performance, important limitations warrant acknowledgment. Key areas for future research include:

External Validation: Multi-center, geographically diverse cohorts are needed to assess generalizability across different healthcare systems, patient demographics, and ECG acquisition devices.

Prospective Clinical Trials: Randomized controlled trials comparing AI-assisted vs. standard ECG interpretation are essential to demonstrate impact on clinician decision-making, diagnostic latency, treatment times, and patient outcomes.

Hybrid Deep Learning Integration: Combining CardioForest’s explainable feature-based reasoning with deep learning’s raw waveform analysis could capture both structured diagnostic criteria and subtle morphological patterns, potentially improving performance while maintaining interpretability through hierarchical explanations.

Multi-Class Arrhythmia Detection: Extending beyond binary WCT/Normal classification to differentiate specific WCT subtypes (ventricular tachycardia, SVT with aberrancy, pre-excitation syndromes) would provide finer-grained diagnostic support.

Real-Time Deployment Studies: Integration with live ECG streaming in emergency departments, with user experience evaluation, workflow analysis, and assessment of clinical adoption barriers, is crucial for translating research into practice.

Multi-Site External Validation: While MIMIC-IV provides a large, diverse dataset from a major academic medical center, external validation across multiple institutions with different patient demographics, ECG acquisition devices, and clinical workflows is essential for assessing generalizability.