Adding deterministic, clinically motivated indicators to deep neural networks for heart disease prediction failed to demonstrate an advantage over using standard clinical predictors alone, according to an ablation study conducted on the UCI Statlog (Heart) dataset. In an evaluation of 270 patient records, researchers observed that while feature augmentation slightly altered single-split test metrics, five-fold stratified cross-validation restricted to the training partition consistently favoured the baseline clinical-only model across all five evaluated performance metrics.
The investigation addresses an expanding domain within biomedical informatics, where heart disease prediction models increasingly combine machine learning techniques with structured clinical knowledge and explainable artificial intelligence. However, the study notes that performance claims on small benchmark datasets can be difficult to interpret when feature construction, data partitioning protocols, and metric calculations are not fully reproducible across independent evaluations.
Conducted by Priyadharshini U and Vijayan R from the Vellore Institute of Technology in Vellore, India, the controlled study evaluated thirteen clinical predictors alongside eight deterministic indicators. Because the original OWL and SWRL ontology, terminology mappings, source code, and historical prediction artifacts were entirely unavailable, the eight indicators were derived directly from the original thirteen predictors. Consequently, the research evaluated deterministic feature augmentation rather than formal automated ontology reasoning.
To establish rigorous experimental controls, the researchers applied a fixed stratified 80:20 train–test split using random seed 42 across the 270 records in the UCI Statlog dataset. This partitioning assigned 216 records to the training set and isolated 54 records to serve as the test set. A five-fold stratified cross-validation procedure was strictly restricted to the 216-record training partition to ensure that validation cycles did not leak test information.
On the held-out 54-record test set, the clinical-only deep neural network achieved 83.33 percent accuracy, 80.00 percent precision, 83.33 percent recall, an F1-score of 81.63 percent, and a ROC-AUC of 0.867. In contrast, the augmented deep neural network incorporating the eight engineered indicators achieved 79.63 percent accuracy, 72.41 percent precision, 87.50 percent recall, an F1-score of 79.25 percent, and a ROC-AUC of 0.879.
The authors emphasised that the single-split differences observed between the two models are purely descriptive and should not be interpreted as evidence of genuine model improvement. Although the augmented model showed higher test recall and ROC-AUC on the single test split, its accuracy, precision, and F1-score fell below those of the clinical-only model, and all five mean cross-validation metrics across the training folds systematically favoured the clinical-only architecture.
The research paper also noted that an original SHAP explainability figure could not be validated against the reported input features and was consequently removed from the study, with no quantitative SHAP findings claimed in the revision. The authors concluded that the findings support cautious, rigorous evaluation of deterministic feature engineering in medical machine learning and do not establish clinical effectiveness or any independent predictive value for ontology reasoning.
Subject of Research: Evaluation of deterministic feature engineering in heart disease prediction models
Article Title: A controlled ablation study of clinically motivated engineered features for heart disease prediction on the Statlog dataset
Article References: Priyadharshini, U., & Vijayan, R. (2026). A controlled ablation study of clinically motivated engineered features for heart disease prediction on the Statlog dataset. Frontiers in Artificial Intelligence, 9, Article 2010348. https://doi.org/10.3389/frai.2026.2010348
Image Credits: AI Generated
DOI: 10.3389/frai.2026.2010348
Keywords: Statlog heart dataset, heart disease prediction, ablation study, deep neural network, deterministic feature engineering, clinical predictors, feature augmentation, stratified cross-validation, model evaluation, reproducibility, ROC-AUC, F1-score, explainable artificial intelligence, SHAP analysis, ontology reasoning, Vellore Institute of Technology, benchmarking
News Source: Frances Kline. (November 20, 2026). Clinical Predictors Outperform Engineered Features in Heart Disease Model Cross-Validation. Scienmag.



