<![CDATA[Transformers have become the dominant architecture for tabular data modelling in natural language processing; however, their effectiveness for numerical tabular classification on modest sized and moderately imbalanced datasets remains unclear. This study evaluates the performance of hybrid deep learning and classical machine learning models which use the Cleveland Heart Disease dataset with 297 complete observations and was artificially constructed from 13 clinical features. The models examined include BiLSTM-CNN, Random Forest, XGBoost, Logistic Regression, and Ridge Regression. An experimental comparative approach was adopted under identical preprocessing, training conditions, and evaluation metrics, including accuracy, recall, F1-score, and Area Under the Curve (AUC). Results show that BiLSTM-CNN achieved the highest recall (0.8478), demonstrating strong minority class detection capability. Random Forest and XGBoost produced the best-balanced performance with 81.67% accuracy and the BiLSTM-CNN has the best F1-score of 0.8364, while Ridge Regression achieved the highest AUC (0.8945). This study provides empirical evidence that hybrid recurrent and ensemble models perform optimally on a small to medium sized Cleveland Heart Desease numerical tabular datasets without pre-training, offering practical guidance for Cleveland Heart Disease tabular clinical classification tasks, and no external validation was performed.]]>
Copyrights © 2026
