Detection of cardiovascular conditions using ECG image analysis: an evaluation of DenseNet121, InceptionV3, and a hybrid deep learning model

Abstract

Cardiovascular diseases remain a leading cause of mortality worldwide, necessitating rapid and accurate diagnostic tools. Electrocardiogram (ECG) image analysis using deep learning offers a promising avenue for automated detection, yet model performance and generalizability require rigorous evaluation. This study aimed to evaluate and compare the performance of three deep learning approaches: DenseNet121, InceptionV3, and a novel hybrid model combining both architectures for the classification of paper-based ECG images into four cardiovascular condition categories. A hybrid model was developed integrating DenseNet121 and InceptionV3 via feature concatenation. All models were trained and evaluated on a dataset of paper-based ECG images using a strictly held-out test set and k-fold cross-validation. Performance was assessed using test accuracy, per-class performance, and statistical significance was determined via McNemar’s test. DenseNet121 and InceptionV3 achieved identical test accuracies of 59.45%. The hybrid model significantly outperformed both individual networks, attaining the highest test accuracy of 72.35% (McNemar’s test, p < 0.001). The difference between the two single-network models was not statistically significant ( p = 0.90). While the hybrid model demonstrated superior generalization and more balanced per-class performance, it exhibited greater cross-validation fold-to-fold variability compared to the individual models. The proposed hybrid model effectively leverages complementary feature extraction mechanisms to improve classification accuracy in ECG image analysis. Although its performance is lower than some previously reported results, the results highlight the potential of hybrid architectures while underscoring the need for validation on larger, more diverse datasets to ensure stability and generalizability in clinical applications.