Heart failure (HF) is a prevalent and life-threatening cardiovascular disorder that necessitates prompt and accurate detection to mitigate its clinical and economic burden. While previous studies have demonstrated the potential of convolutional neural networks (CNNs) in automatic HF detection using ECG signals, challenges remain regarding the generalizability and robustness of such models, particularly under variable real-world conditions. This study proposes an enhanced CNN-based framework for HF classification that incorporates advanced evaluation protocols and robustness analysis. Our approach introduces cross-database validation, noise resilience testing, and performance stratification by demographic and physiological factors to evaluate the model under more realistic clinical scenarios. We develop a deep learning architecture trained on balanced and imbalanced datasets using ECG segments from the MIT-BIH Normal Sinus Rhythm and BIDMC Congestive Heart Failure databases. The proposed model achieves outstanding performance with an accuracy of 99.31%, a sensitivity of 99.42%, and a specificity of 99.19% on cross-validation. Furthermore, the model demonstrates consistent performance across age groups, gender, and signal perturbations, highlighting its robustness and clinical applicability. These findings suggest that a comprehensive evaluation of diagnostic models is essential for their deployment in healthcare settings.