CardioForest: Ensemble Methods in Cardiology

• CardioForest is a collection of tree ensemble methods that address diverse cardiology tasks using variations of Random Forest, Extra-Trees, and stacking strategies.
• It leverages heterogeneous cardiac data—from clinical metrics and ECG features to imaging descriptors—to deliver high-performance, clinically deployable models.
• The framework emphasizes interpretability and operational constraints, integrating techniques like SHAP and impurity-based ranking to support decision-making in clinical settings.

Searching arXiv for the cited CardioForest-related papers and term usage. Search query: CardioForest Search query: (Talukder et al., 2023) Search query: (Chakma et al., 30 Sep 2025) Search query: (Costa et al., 2024) Search query: (Tiwari et al., 2023) Search query: (Rodrigues et al., 28 May 2026) CardioForest denotes a set of cardiology-oriented machine-learning systems in which tree ensembles occupy the central modeling role. In the cited literature, the name is used for several distinct pipelines rather than for one fixed architecture: an Extra-Tree feature-selection plus Random Forest classifier for heart-failure survival prediction, an optimized Random Forest for automatic Wide QRS Complex Tachycardia diagnosis from ECG, a tree-based tertiary-care triage framework combining risk factors with standard ECG and vectocardiographic features, a stacked ensemble for cardiovascular disease prediction, and a Random Forest pipeline for cardiac fat segmentation on CT (Talukder et al., 2023, Chakma et al., 30 Sep 2025, Costa et al., 2024, Tiwari et al., 2023, Rodrigues et al., 28 May 2026). The common denominator is the use of ensemble decision trees to model heterogeneous cardiovascular data while preserving practical interpretability, modest preprocessing assumptions, and clinically meaningful feature ranking.

1. Scope, nomenclature, and recurring design motifs

The cited literature uses “CardioForest” for several forest-based or tree-ensemble systems rather than for one canonical model. Some instances are conventional Random Forests with task-specific optimization, some are two-stage pipelines in which a tree ensemble performs feature selection before downstream classification, some are stacked ensembles that include forest models as base learners, and one is a voxelwise image-segmentation system implemented with Weka’s RandomForest.

Variant	Clinical task	Core formulation
Heart-failure CardioForest	death_event prediction	Extra-Tree feature selection + tuned Random Forest
ECG CardioForest	WCT detection	optimized Random Forest + SHAP
Triage CardioForest framework	tertiary cardiology allocation	XGBoost decision-tree ensemble with risk, ECG, and GEH features
Cardiovascular-risk CardioForest	heart disease prediction	stacked RF/ETC/XGBoost/GBM ensemble
CT CardioForest	epicardial/mediastinal fat segmentation	intersubject registration + Random Forest voxel classifier

Across these variants, several motifs recur. First, the inputs are predominantly tabular or structured measurements: clinical covariates in heart failure, machine-generated ECG timings and axes in WCT detection, risk-factor and GEH-derived variables in tertiary triage, mixed categorical-numeric cardiovascular features in population-level heart disease prediction, and intensity-spatial-texture descriptors in CT segmentation. Second, impurity-based splitting and ensemble aggregation are central. Third, the systems are framed as clinically deployable decision support rather than as end-to-end representation-learning systems.

A further recurring property is task-specific interpretability. In heart-failure prediction, interpretability is delivered through Extra-Tree impurity-based importance ranking. In WCT diagnosis, SHAP is used to connect predictions to ECG features such as QRS duration. In tertiary triage, XGBoost gain scores identify the influence of age, QTc, and GEH variables. In CT segmentation, interpretability is more anatomical than explanation-centric, with centralization of the heart and explicit voxelwise features tied to HU range and spatial location.

2. Heart-failure survival prediction variant

In "Unleashing the Power of Extra-Tree Feature Selection and Random Forest Classifier for Improved Survival Prediction in Heart Failure Patients" (Talukder et al., 2023), CardioForest can be understood as a two-stage pipeline for heart-failure risk stratification. The study uses the UCI/UCL Heart Failure Clinical Records dataset with $N=299$ patients, all with left ventricular systolic dysfunction and prior heart failure (NYHA class III/IV), age $\ge 40$. The target is binary: $death\_event=1$ for deceased during follow-up and $death\_event=0$ for alive, with the paper reporting 32.1% death_event versus 67.9% alive. The feature set comprises age, anaemia, creatinine phosphokinase, diabetes, ejection fraction, high blood pressure, platelets, serum creatinine, serum sodium, sex, smoking, time, and the target itself. Crucially, the paper uses “time” as a predictor rather than as the survival outcome variable.

The preprocessing stage applies StandardScaler with $z=(x-\mu)/\sigma$. Binary indicators were already encoded as $0/1$, no missing-value imputation is described because the dataset has no missing values, and no explicit outlier treatment is reported. The paper standardized the full dataset before splitting, which introduces a data-leakage risk; the same source notes that best practice would be to fit the scaler on training data only. Class imbalance is moderate and is handled through $class\_weight='balanced'$; no SMOTE or oversampling is reported in this study.

Feature selection is performed with ExtraTreesClassifier using impurity-based importance. The mathematical basis is explicit. For a node $t$, the impurities are given as

$G(t)=1-\sum_c p(c\mid t)^2$

and

$H(t)=-\sum_c p(c\mid t)\log p(c\mid t).$

For a split on feature $\ge 40$0,

$\ge 40$1

and the mean decrease in impurity is

$\ge 40$2

The ET ranking identifies time, ejection_fraction, serum_creatinine, and age as the most predictive variables, followed by serum_sodium, creatinine_phosphokinase, platelets, diabetes, smoking, sex, anaemia, and high_blood_pressure.

These selected variables are then used for grid search over Random Forest hyperparameters. The explored grid includes $\ge 40$3, $\ge 40$4, $\ge 40$5, $\ge 40$6, and $\ge 40$7. The paper reports an 80:20 train/test split with an internal validation loop but does not report $\ge 40$8-fold cross-validation or out-of-bag scoring.

The final tuned model achieves 98.33% accuracy, 100% precision, 94.12% recall, 100% specificity, 96.97% F1-score, 97.06% balanced accuracy, 97.06% ROC-AUC, Cohen’s Kappa of 95.82, and MCC of 95.91 (Talukder et al., 2023). The confusion matrix figure is described as showing only one misclassification, and the reported values are consistent with $\ge 40$9, $death\_event=1$0, $death\_event=1$1, $death\_event=1$2 on an approximately 60-patient test set. The paper also reports Mean Squared Error $death\_event=1$3 and Gini coefficient $death\_event=1$4, while noting that these quantities are not standardly presented in this form for binary classification.

Methodologically, this CardioForest instantiation is a classification system rather than a survival-analysis model. That distinction is central. It yields deployable probabilities of $death\_event=1$5, but it does not estimate hazard functions or time-dependent risk, and censoring is not modeled explicitly. This suggests that its reported performance depends materially on treating follow-up time as a covariate rather than modeling time-to-event structure directly.

3. Explainable ECG diagnosis of Wide QRS Complex Tachycardia

In "CardioForest: An Explainable Ensemble Learning Model for Automatic Wide QRS Complex Tachycardia Diagnosis from ECG" (Chakma et al., 30 Sep 2025), CardioForest is an optimized Random Forest for automatic WCT detection from ECG. The clinical target is Wide QRS Complex Tachycardia, defined in the study as a rapid rhythm with widened QRS complexes, typically reflecting ventricular tachycardia or supraventricular tachycardia with aberrant conduction. The paper states that “if the QRS duration exceeds 120 milliseconds, the rhythm may be suggestive of WCT,” and this threshold informs interpretation.

The data source is the MIMIC-IV-ECG Diagnostic Electrocardiogram Matched Subset, described as approximately 800,035 ten-second 12-lead ECG recordings at 500 Hz from 161,352 unique subjects. Records are stored in WFDB format, with machine-generated measurements available in machine_measurements.csv and linked notes available through waveform_note_links.csv. Duplicate removal is performed by subject_id and study_id, yielding 0 duplicates in the final dataset. Missing numerical values are handled by median imputation, negative RR intervals are corrected via interpolation, categorical fields are label-encoded, and numerical features are standardized by $death\_event=1$6-score before PCA and modeling. Correlation heatmaps are used to remove redundant variables, while PCA is used for exploration rather than final inference.

The feature set includes RR interval, $death\_event=1$7, $death\_event=1$8, $death\_event=1$9, $death\_event=0$0, $death\_event=0$1, $death\_event=0$2, $death\_event=0$3, $death\_event=0$4, $death\_event=0$5, and derived heart-rate and heart-rate-variability variables. The key computations are explicit:

$death\_event=0$6

if RR is in seconds, or

$death\_event=0$7

if RR is in milliseconds, and

$death\_event=0$8

The model architecture is a Random Forest with $death\_event=0$9, $z=(x-\mu)/\sigma$0, $z=(x-\mu)/\sigma$1, $z=(x-\mu)/\sigma$2, $z=(x-\mu)/\sigma$3, $z=(x-\mu)/\sigma$4, $z=(x-\mu)/\sigma$5, $z=(x-\mu)/\sigma$6, $z=(x-\mu)/\sigma$7, $z=(x-\mu)/\sigma$8, and $z=(x-\mu)/\sigma$9. Splits minimize Gini impurity,

$0/1$0

and ensemble prediction is by majority vote,

$0/1$1

Hyperparameters are selected by systematic grid search with stratified 10-fold cross-validation, optimizing a weighted combination of F1 and ROC-AUC while minimizing RMSE.

CardioForest outperforms the reported XGBoost, LightGBM, and GradientBoosting baselines on the headline test metrics. The paper reports accuracy $0/1$2, balanced accuracy $0/1$3, precision $0/1$4, recall $0/1$5, F1 $0/1$6, ROC-AUC $0/1$7, RMSE $0/1$8, and MAE $0/1$9 (Chakma et al., 30 Sep 2025). Comparative deltas are also reported: versus XGBoost, $class\_weight='balanced'$0 percentage points in accuracy and $class\_weight='balanced'$1 in F1; versus LightGBM, $class\_weight='balanced'$2 percentage points in accuracy and $class\_weight='balanced'$3 in F1; versus GradientBoosting, $class\_weight='balanced'$4 percentage points in accuracy and $class\_weight='balanced'$5 in F1, though MAE is slightly higher than for GradientBoosting.

A distinctive property of this variant is explicit explainability. SHAP analysis is used for global and local explanation, and the paper states that QRS duration emerges as the top feature, with RR interval, QRS axis, and T axis also contributing. The directionality is clinically aligned: increased $class\_weight='balanced'$6 increases WCT probability, shorter RR increases WCT likelihood, and certain axis deviations may correlate with ventricular origin. This is the most explicit use of Explainable AI among the cited CardioForest systems.

The paper also records several limitations. Labels are derived from machine-generated interpretations and categorical encodings, but the mapping rules are not fully specified. Stratified 10-fold cross-validation is used, yet cross-patient partitioning is not reported, so multiple ECGs from one subject may appear in different folds. External validation is not performed. These constraints are important because the model’s strong discrimination is demonstrated within a large single-source dataset rather than across institutions or devices.

4. Tertiary cardiology triage with ECG, VCG, and GEH features

The study "Evaluating the Efficacy of Vectocardiographic and ECG Parameters for Efficient Tertiary Cardiology Care Allocation Using Decision Tree Analysis" is presented in the source material as a CardioForest-style triage framework (Costa et al., 2024). Its problem formulation differs from the heart-failure and WCT variants: the aim is efficient tertiary care allocation using risk factors, standard ECG features, and vectocardiographic measurements of global electric heterogeneity.

The cohort is a single-center, real-world prospective sample from Instituto Dante Pazzanese de Cardiologia in São Paulo. The initial cohort contains 303 first-visit subjects, with 293 outcomes confirmed at 6 months and 274 at 1 year; the final analytic cohort is 274. Baseline demographics include median age 59.0 years $class\_weight='balanced'$7, 52.2% female, median BMI $class\_weight='balanced'$8 kg/m$class\_weight='balanced'$9 $t$0, hypertension 69.0%, diabetes 27.7%, previous MI 28.1%, previous PCI 14.6%, previous cardiac surgery 11.3%, and previous stroke 6.6%. The composite positive outcome is mortality or a new nonfatal cardiovascular event including stroke, MI, PCI, and cardiac surgery.

Baseline ECG measurements include P-wave interval, PR segment interval, QRS interval, QT interval, QTc, and RR interval. GEH features are derived after ECG-to-VCG conversion using a Kors transformation, formalized as

$t$1

The paper defines peak and area QRST angles by

$t$2

defines the spatial ventricular gradient by

$t$3

and defines the vector magnitude QT integral as

$t$4

Feature sets are grouped as $t$5 for risk factors, $t$6 for standard ECG, $t$7 for VCG/GEH, and $t$8 for the combined model.

The learning algorithm is XGBoost rather than Random Forest. Training uses a 70/30 random split, internal cross-validation, AUCPR as the optimization metric, and bootstrap oversampling plus undersampling to address a roughly $t$9 negative-to-positive class ratio. For each feature set, 50 XGBoost instances are trained under selected settings, and the instance with the highest AUC is chosen as the representative “winner tree.” The paper emphasizes recall by selecting a fixed high-sensitivity operating point.

At the reported threshold, all models operate at 94.12% sensitivity, but specificity differs materially. The combined $G(t)=1-\sum_c p(c\mid t)^2$0 model yields F2 $G(t)=1-\sum_c p(c\mid t)^2$1, AUC $G(t)=1-\sum_c p(c\mid t)^2$2, sensitivity $G(t)=1-\sum_c p(c\mid t)^2$3, and specificity $G(t)=1-\sum_c p(c\mid t)^2$4, compared with F2 $G(t)=1-\sum_c p(c\mid t)^2$5, AUC $G(t)=1-\sum_c p(c\mid t)^2$6, specificity $G(t)=1-\sum_c p(c\mid t)^2$7 for $G(t)=1-\sum_c p(c\mid t)^2$8; F2 $G(t)=1-\sum_c p(c\mid t)^2$9, AUC $H(t)=-\sum_c p(c\mid t)\log p(c\mid t).$0, specificity $H(t)=-\sum_c p(c\mid t)\log p(c\mid t).$1 for $H(t)=-\sum_c p(c\mid t)\log p(c\mid t).$2; and F2 $H(t)=-\sum_c p(c\mid t)\log p(c\mid t).$3, AUC $H(t)=-\sum_c p(c\mid t)\log p(c\mid t).$4, specificity $H(t)=-\sum_c p(c\mid t)\log p(c\mid t).$5 for $H(t)=-\sum_c p(c\mid t)\log p(c\mid t).$6 (Costa et al., 2024). The source explicitly states that adding $H(t)=-\sum_c p(c\mid t)\log p(c\mid t).$7 to $H(t)=-\sum_c p(c\mid t)\log p(c\mid t).$8 improves AUC and specificity substantially at the same sensitivity.

The study also reports statistically significant group differences for several GEH parameters: peak QRST angle $H(t)=-\sum_c p(c\mid t)\log p(c\mid t).$9, area QRST angle $\ge 40$00, peak SVG elevation $\ge 40$01, and SVG integral $\ge 40$02. In the combined model, the top gain-based features include age at 9.4%, SVG magnitude at 8.6%, QTc at 7.5%, area SVG elevation at 7.3%, area SVG azimuth at 6.9%, PR interval at 5.8%, and previous PCI at 5.1%.

Within the CardioForest family, this triage framework is notable for showing that tree ensembles can incorporate spatial-repolarization variables beyond conventional ECG intervals. It is also the clearest example of capacity-aware thresholding: the model is evaluated not merely for discrimination, but for specificity under a fixed high-sensitivity operating point intended to avoid overburdening tertiary services. Its limitations are equally explicit: single-center design, modest $\ge 40$03, no external validation, no calibration analysis, and outcome ascertainment by phone follow-up without reported independent adjudication.

5. Stacked-ensemble cardiovascular disease prediction

In "Ensemble Framework for Cardiovascular Disease Prediction" (Tiwari et al., 2023), CardioForest is a forest-centric stacked ensemble for heart-disease prediction using the “Heart Disease Dataset (Comprehensive)” from IEEE Data Port. This combined dataset merges the Hungarian, Cleveland, Long Beach VA, Switzerland, and Statlog cohorts, yielding 1190 instances with 11 features plus a binary target. The class distribution is slightly imbalanced toward disease: 53% positive and 47% negative.

The feature set includes age, resting blood pressure, serum cholesterol, maximum heart rate achieved, ST depression, sex, chest pain type, fasting blood sugar, resting ECG result, exercise-induced angina, and slope of peak exercise ST segment. Pearson correlations with the target identify ST slope $\ge 40$04, exercise-induced angina $\ge 40$05, chest pain type $\ge 40$06, ST depression $\ge 40$07, sex $\ge 40$08, and age $\ge 40$09 as positive correlates, while cholesterol $\ge 40$10 and maximum heart rate achieved $\ge 40$11 are negatively correlated. Preprocessing consists of exploratory analysis, correlation analysis, outlier removal, and an 80/20 train-test split; normalization is not explicitly reported, and no class-rebalancing method is described.

The architecture is stacking. Ten candidate classifiers are first evaluated, then the four strongest by 10-fold cross-validation accuracy are selected as base learners: Random Forest, ExtraTrees Classifier, XGBoost, and Gradient Boosting Classifier. A meta-classifier $\ge 40$12 combines their outputs, but the exact algorithm is not specified. The paper gives the generic stacking formalism

$\ge 40$13

and illustrates logistic stacking as a possible but not explicitly claimed choice.

The algorithmic detail is uneven across base learners. For Random Forest, entropy is reported as the selected split criterion. For ExtraTrees, $\ge 40$14 are tried, with best performance at 500 trees. For XGBoost, $\ge 40$15 are tried, again with best performance at 500 trees. The meta-learner hyperparameters and the precise out-of-fold stacking protocol are not reported.

On the test set, the stacked CardioForest reaches accuracy $\ge 40$16, precision $\ge 40$17, sensitivity $\ge 40$18, specificity $\ge 40$19, F1-score $\ge 40$20, ROC $\ge 40$21, MCC $\ge 40$22, AUC-ROC $\ge 40$23, and average precision $\ge 40$24 (Tiwari et al., 2023). The paper compares this performance with several baselines, including ExtraTrees at 90.93% accuracy, XGBoost at 91.91%, and Random Forest at 90.21%. It also reports literature comparisons, placing CardioForest’s 92.34% accuracy above several cited earlier works.

This variant broadens the meaning of CardioForest beyond a single Random Forest or ET+RF pipeline. Here the defining feature is the forest-centric composition of the ensemble rather than exclusive reliance on one forest model. The limitations follow directly from that design and reporting style: the exact meta-classifier is unspecified, confidence intervals are absent, external validation is not reported, and the merged multi-source dataset may introduce site and coding heterogeneity.

6. Automatic segmentation of epicardial and mediastinal fat on CT

In "Towards the automated segmentation of epicardial and mediastinal fats: A multi-manufacturer approach using intersubject registration and random forest" (Rodrigues et al., 28 May 2026), CardioForest is a Random Forest pipeline for fully automatic segmentation rather than a classifier for diagnosis or prognosis. The target structures are epicardial fat, located within the visceral pericardium and contiguous with the myocardium and coronary arteries, and mediastinal fat, external to the fibrous pericardium in the mediastinal space. The paper notes that adipose tissue on non-contrast coronary calcium scoring CT typically lies around $\ge 40$25 to $\ge 40$26 HU; operationally, the range of interest is set to $\ge 40$27 HU.

The workflow combines intersubject registration, handcrafted feature extraction, and supervised voxel classification across Siemens and Philips scans. The localization stage constructs a 2D retrosternal atlas from 10 manually pre-aligned patients, binarized by thresholding at $\ge 40$28 HU and averaged within the fat range. On a new case, the atlas is slid over one slice and scored using weighted mutual information. This alone yields a 70% correct localization rate. A heuristic confirmation step using two fat-tracking points then raises localization success to 100% across 82 assessed instances. After successful localization, all slices are translated to a common reference, and DICOM-informed rescaling normalizes physical dimensions.

Feature extraction is explicitly spatial and textural. Per-voxel inputs include HU gray level, Cartesian coordinates $\ge 40$29, and $\ge 40$30 and $\ge 40$31 relative to the image’s center of gravity. Local features are computed in a $\ge 40$32 neighborhood and include arithmetic mean, Haralick-style grey-level co-occurrence moments, geometrical moments of grey values, run-length features such as run percentage and grey-level non-uniformity, and a coefficient of smooth variation implemented by convolution with kernel weights derived from the sup metric and a 1D Gaussian. No dedicated denoising is applied.

The classifier is Weka’s RandomForest with parameters $\ge 40$33. Training data comprise approximately 3 GB of feature vectors derived from about 1000 manually segmented CT slices across 20 patients, with two experts producing ground-truth masks. Background voxels outside the fat HU range are excluded from feature extraction and training. Validation uses both a random 66%/34% split and 10-fold cross-validation.

The reported metrics are high. For the 66% split, epicardial fat reaches 98.3% accuracy and 98.1% TPR, while mediastinal fat reaches 98.0% accuracy and 92.9% TPR. Under 10-fold cross-validation, epicardial fat reaches 98.5% accuracy and 98.3% TPR, and mediastinal fat 98.4% accuracy and 94.2% TPR. Overall mean accuracy across labels is 98.4%, mean TPR is 96.2%, and average Dice similarity is 96.8%; a specific epicardial Dice of 97.9% is also reported in comparison with earlier epicardial-only methods (Rodrigues et al., 28 May 2026). Dice is defined as

$\ge 40$34

and volume is computed by voxel counting times voxel dimensions:

$\ge 40$35

A defining characteristic of this imaging CardioForest is that it does not explicitly segment the pericardial membrane. Instead, separation between epicardial and mediastinal fat is learned from context, texture, and spatial priors after centralization of the heart. The source treats this as a practical strength, but it also identifies failure modes near thin pericardial regions and boundaries with lung or chest-wall fat. Processing time is another major limitation: on a dual-core CPU with 4 GB RAM, end-to-end processing per patient takes several hours, dominated by feature extraction.

7. Cross-cutting methodological themes, misconceptions, and limitations

A common misconception would be to treat CardioForest as synonymous with a particular algorithm such as Random Survival Forests or a single standardized software package. The cited literature does not support that interpretation. Instead, the name is attached to ET+RF classification for heart failure, optimized Random Forest ECG diagnosis, XGBoost-based triage, a stacked ensemble over multiple learners, and Random Forest CT segmentation (Talukder et al., 2023, Chakma et al., 30 Sep 2025, Costa et al., 2024, Tiwari et al., 2023, Rodrigues et al., 28 May 2026).

Despite that architectural diversity, several methodological themes recur. One is the preference for tree ensembles on heterogeneous cardiology data: mixed continuous and binary clinical variables in heart failure, machine-generated ECG measurements in WCT detection, combined risk-factor and GEH features in tertiary triage, heterogeneous structured features in heart-disease prediction, and high-dimensional handcrafted descriptors in CT segmentation. Another is the use of feature importance or explainability to maintain clinical alignment, whether via impurity-based ET ranking, XGBoost gain, or SHAP. A third is an emphasis on deployability under practical constraints such as high-sensitivity triage thresholds, near-real-time ECG inference, or automated volumetry from standard CAC CT.

The limitations are similarly recurrent. External validation is absent in the heart-failure, WCT, tertiary-triage, and stacked cardiovascular-risk variants, and the CT segmentation paper reports multi-manufacturer data but not per-vendor stratified metrics. Data leakage concerns are explicit in the heart-failure paper because standardization was applied before train-test splitting, and they are also plausible in the WCT paper because multiple ECGs from the same patient may appear in different folds. Calibration is not assessed in the heart-failure and tertiary-triage studies, while the stacked cardiovascular-risk paper omits detailed meta-learner specification. The heart-failure CardioForest additionally uses “time” as a predictor in a binary classification task, so censoring is not modeled explicitly.

Taken together, these reports position CardioForest as a flexible design label for tree-based cardiology systems rather than as a monolithic method. The strongest evidence across the cited studies is task-specific: high-accuracy heart-failure death_event classification, explainable WCT detection from ECG, improved specificity at fixed sensitivity in tertiary triage when GEH features are added, competitive stacked-ensemble heart-disease prediction, and accurate fully automatic segmentation of epicardial and mediastinal fat. A plausible implication is that the enduring identity of CardioForest lies less in one fixed model class than in a recurring methodological choice: forest-centric ensemble learning applied to clinically structured cardiovascular data, with interpretation mechanisms and operational constraints built into the pipeline.