This study proposes an advanced methodology for the classification of electrocardiogram (ECG) signals into normal and abnormal categories, leveraging deep learning techniques and statistical features. Specifically, deep features extracted from a Siamese Network structure are combined with an extra trees classifier for decision-making layers. In addition to deep features, fundamental statistical features such as entropy, mean, variance, and skewness are incorporated to enrich the classification process.
The Siamese Network architecture captures intricate patterns within ECG signals, providing rich representations crucial for accurate classification. Complemented by the extra trees classifier, which effectively handles class imbalances and enhances classification performance, our approach ensures robust categorization. Siamese structures in ECG classification enhance adaptability, capturing nuanced patterns for reliable real-world implementation.
Furthermore, we introduce a novel approach for reconstructing ECG signals from conventional ECG images into signal format using a bespoke Deep Neural Network (DNN), thereby expanding the feature space and enhancing the discriminative power of the features. This method(YAMA) achieved an Signal-to-Noise Ratio(SNR) of 3.7 during our private validation stage. During the testing phase, our system achieves a SNR of -18.10, as per the official test scores of the PhysioNet Challenge 2024, when reconstructing ECG signals from image representations using only statistical-based features.
When employing statistical-based features and the extra trees classifier, our system(YAMA) achieves F-measure scores of 0.61 and 0.51, respectively, on the official test scores of the PhysioNet Challenge 2024. Moreover, the incorporation of deep features significantly boosts classification performance, highlighting the effectiveness of our methodology, as evidenced by an F-measure of 0.98 on the initial private validation set. The implications of our research extend to clinical practice, offering a reliable solution for automated ECG analysis from ECG image reconstruction to initial stages fast and reliable ECG classification.
Keywords: Electrocardiogram (ECG), Deep Features, Extra Trees Classifier, Statistical Features, Image Reconstruction