Explainable AI framework blends calibrated models to predict heart disease risk

Cardiovascular disease remains the most prevalent cause of illness and death worldwide, and machine-learning models built to flag it early have long been criticised as opaque, poorly calibrated, and fragile. A team reporting in Discover Artificial Intelligence has now assembled a framework designed to attack all three weaknesses at once. Their system, called FWC-TabXAI++, combines clinically constrained data cleaning, second-order feature-interaction expansion, mutual-information-based feature weighting, a heterogeneous ensemble of classifiers, validation-guided probability blending, and threshold optimisation into a single pipeline for heart disease prediction. On a cleaned clinical dataset of 836 records, the framework achieved an accuracy of 91.07%, a precision of 92.22%, a recall of 91.21%, a specificity of 90.91%, and an F1-score of 91.71%, alongside an ROC-AUC of 94.01% and a PR-AUC of 93.24%. Crucially, the model did not stop at a prediction: it attached SHAP values and LIME-based explanations that identified ST-segment slope and exercise-induced angina as dominant predictive contributors for individual patients.

The motivation stems from persistent gaps in the machine-learning literature on cardiovascular prediction. The authors note that many existing studies rely on small or single-centre datasets, lean on black-box deep learning architectures that resist interpretation, or chase accuracy while neglecting calibration and class imbalance. Prior efforts illustrate the landscape: a two-level stacking ensemble on the Cleveland data reached 92% accuracy, while a random forest classifier applied to a cross-sectional dataset of 391 cardiovascular patients and 260 controls achieved 98.04% accuracy. Other work has combined SMOTE oversampling with feature selection, achieving 97.57% accuracy with XGBoost, or deployed IoT-connected systems such as DEEP CARDIO, which reported 99.90% accuracy using a BiGRU model with attention. Yet, according to the researchers, few studies jointly incorporate second-order feature interactions, calibrated probability estimation, statistical validation, and both global and local explainability in one framework, which is the niche FWC-TabXAI++ was designed to fill.

The data behind the study come from the publicly available Kaggle Heart Disease repository, based on the UCI Heart Disease dataset. The raw data contained 918 observations and 12 variables spanning demographics, electrocardiographic measurements, exercise-related indicators, and biochemical values. Preprocessing was deliberately clinical in its constraints: continuous measurements such as serum cholesterol and resting blood pressure, which cannot physiologically take a value of zero in this diagnostic context, had zeros replaced with the non-zero mean of the same variable. Outlier filtering using the interquartile range method then removed 27 records from RestingBP, 40 from Cholesterol, 2 from MaxHR, and 13 from Oldpeak, while Age required no removal. After these steps, 836 samples were retained, comprising 454 heart-disease-positive and 382 non-heart-disease records, a moderately imbalanced but clinically feasible distribution. Categorical variables such as chest pain type, resting ECG status, exercise-induced angina, and ST-segment slope were one-hot encoded, and numerical variables were standardised using training-set statistics only, to prevent information leakage.

The methodological core of the framework is a staged pipeline rather than a single black-box classifier. The cleaned features are first expanded with second-order interaction terms, generating all pairwise products between variables so that, for example, the significance of maximum heart rate can vary with age and ST-segment depression can act synergistically with exercise-induced angina. Because this expansion inflates dimensionality, mutual information between each expanded feature and the diagnostic label is computed, and only the top 60 most informative variables are retained. The selected representation feeds a heterogeneous pool of candidate learners spanning linear, kernel-based, tree-based, boosting-based, and neural components, including logistic regression, support vector machines, k-nearest neighbours, random forests, and gradient boosting. Rather than crowning a single winner, the framework learns a convex combination of the candidates’ probability outputs, with blending weights optimised on the validation set through Dirichlet-distributed random simplex search. The decision threshold was likewise tuned on validation data, settling near 0.46 instead of the default 0.50.

The mathematical scaffolding behind these choices carries weight. The authors prove that a linear classifier operating on the interaction-expanded representation can represent any second-order decision function of the original variables, which justifies using simpler learners instead of deep architectures. A second proposition shows that the validation-optimised convex ensemble cannot underperform the best single candidate on the same validation criterion, provided the simplex search can reach the vertices. A third guarantees that blending valid probabilities yields a valid probability, a property essential for the downstream calibration analysis. The authors are careful to note that this optimisation property is empirical in nature and does not by itself establish external clinical generalisation; that question is addressed through independent testing and confidence-interval analysis instead.

On the independent test set of 168 samples, FWC-TabXAI++ delivered the best overall balance across metrics when compared against 11 baseline classifiers. Gradient Boosting and Gaussian Naïve Bayes achieved competitive AUC values, and Gaussian Naïve Bayes in fact posted a slightly higher ROC-AUC of 94.45% versus 94.01% for the proposed model. The researchers explicitly declined to interpret these small ROC-AUC differences as evidence of statistical superiority, instead framing their model’s advantage as balanced performance: it recorded the highest accuracy at 91.07% and the highest F1-score at 91.71% while maintaining strong threshold-independent discrimination. The recall of 91.21% indicates the model detected most heart disease cases, while the specificity of 90.91% shows it also excluded non-disease cases effectively, a balance the authors describe as clinically reasonable because excessive emphasis on either side inflates false positives or misses patients.

Ablation analysis dissected the contribution of each component. Interaction modelling and mutual-information-based feature selection progressively refined the predictive representation, ensemble blending improved overall discrimination, and threshold optimisation lifted accuracy and F1-score without altering the ranking-based ROC-AUC and PR-AUC. The variant without threshold optimisation achieved 89.29% accuracy, with ROC-AUC and PR-AUC identical to the full model, underscoring that ranking quality and the final operating decision are distinct properties. To reduce dependence on a single data split, the team also ran repeated stratified validation in which all preprocessing, feature selection, blending, and threshold selection were fitted within the training and validation portions only. Across runs, the model achieved mean values of 85.85% accuracy, 87.33% F1-score, 92.93% ROC-AUC, and 92.80% PR-AUC, with relatively small standard deviations, and the mean selected threshold of 0.462 closely matched the value used in the independent test evaluation.

Uncertainty and calibration received dedicated treatment. Bootstrap resampling produced a mean accuracy of 91.04% with a 95% confidence interval spanning 86.31% to 95.24%, while the F1-score interval ranged from 87.01% to 95.48% and ROC-AUC from 89.64% to 97.51%; the authors note that the interval widths also reflect the limited size of the test set. The model achieved a Brier score of 0.0908, with a calibration intercept of 0.3591, a calibration slope of 1.2717, and a calibration-in-the-large value of 0.3066. The observed event rate of 0.5417 exceeded the mean predicted probability of 0.5071, indicating mild overall underprediction, and the reliability curve stayed reasonably close to the ideal diagonal with some deviation in the intermediate probability range. The framework also grouped test predictions by confidence: 70 samples were assigned to the confident heart-disease group, 62 to the confident non-heart-disease group, and 36, or 21.43%, were flagged as uncertain, a triage mechanism the authors suggest could route ambiguous cases to physician review, additional ECG interpretation, or laboratory testing.

Explainability analysis tied the model’s behaviour back to clinical reasoning. Chi-square tests confirmed that all categorical attributes were statistically significant at p < 0.05, with ST_Slope, ChestPainType, and ExerciseAngina showing the strongest associations with heart-disease status, a pattern the authors call clinically plausible given the link between ST-segment changes and cardiovascular stress response. For a representative test sample, the SHAP waterfall explanation identified ST_Slope_Flat and ST_Slope_Up as dominant positive contributors, while MaxHR, ChestPainType_ASY, and cholesterol-related values reduced the prediction, yielding a final output of approximately f(x) = 0.527 against an expected value of 0.499, a borderline case. The LIME explanation complemented this with rule-like intervals such as ST_Slope_Up ≤ 0 and MaxHR > 0.73. The authors caution that SHAP and LIME quantify feature contributions to predictions and should not be read as evidence of causal relationships or clinical risk causation.

The study’s limitations are stated plainly. Because the work rests on a single public dataset, the authors advise caution regarding broader clinical generalisability and call for external validation on independent multi-centre cohorts to confirm robustness across different populations, acquisition settings, and clinical environments. Future directions include integrating additional biomarkers and multi-modal data, prospective deployment in clinical workflows, adaptive learning mechanisms to update the model with new data, and fairness and bias-mitigation strategies to ensure equitable predictions across patient subgroups. Within those boundaries, the framework’s policy significance statement positions it as a tool for transparent, data-driven heart disease risk assessment that may improve early screening, patient prioritisation, and clinical decision support while maintaining clinician oversight, rather than as a replacement for clinical judgement.

Subject of Research: An interpretable, calibrated machine-learning framework for heart disease risk prediction from tabular clinical data

Article Title: Feature-weighted calibrated tabular explainable AI for heart disease prediction

Article References: Asad, M., Islam, M. T., Khan, M. T. R., Aregay, E., Belay, R., Hossain, M. R., Rone, P. D., Mondal, P., & Jahan, S. (2026). Feature-weighted calibrated tabular explainable AI for heart disease prediction. Discover Artificial Intelligence, 6(1), Article 1642. https://doi.org/10.1007/s44163-026-02360-3

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02360-3

Keywords: heart disease, explainable AI, machine learning, SHAP, LIME, probability calibration, ensemble learning, cardiovascular disease, risk prediction, feature selection, mutual information, clinical decision support