Deep learning methods exhibit excellent performance in diagnosing cardiovascular diseases from ECGs. Nevertheless, their “black-box” nature contributes to their limited integrations into clinical scenarios because a lack of interpretability hinders their broader adoption. To overcome this limitation, researchers from the Institute of Biomedical Engineering, TU Dresden, developed xECGArch, a deep learning architecture designed specifically for ECG analysis. This architecture uniquely separates short-term (morphological) and long-term (rhythmic) ECG features using two independent Convolutional Neural Networks (CNNs), aiming to satisfy clinical requirements for accurate and transparent automated analysis.
The study utilizes databases such as PTB-XL, Georgia-12-Lead, China Physiological Signal Challenge 2018 (CPSC2018), and Chapman-Shaoxing, offering an extensive 12-lead ECG database. A focus on single-lead ECGs ensures compatibility with portable devices and optimum efficacy in AF detection.
The xECGArch integrates two independent 1D CNNs that capture both the morphological and rhythmic patterns in the ECG signals. The short-term network specializes in examining beat-level attributes in a 0.6-second window, while the long-term network focuses on wider rhythmic information in the 10-second ECG recording. Both networks also contain features such as Global Average Pooling (GAP) that enhance model efficiency and performance.
xCPRGArch achieved superior performance in AF detection, with the best short-term model earning a 94.18% F1 score and the best long-term model receiving a 95.13% F1 score. Combining the outputs of both models further improved overall performance to a 95.43% F1 score. The investigation used interpretation methods, such as Deep Taylor Decomposition (DTD), Integrated Gradients (ITG), and Layer-wise Relevance Propagation (LRP), to demonstrate the model’s decision-making process. The short-term model emphasized P waves and F waves, while the long-term model focused on irregular R peaks, showing the relevance of different ECG features in the classification.
In conclusion, by leveraging distinct temporal ECG features, the xECGArch’s combined short- and long-term CNNs improve the detection of AF. The architecture exceeds many existing methods, achieving a 95.43% F1 score making it effective for interpreting model decisions in ECG analysis. Further application of xECGArch may extend beyond ECG data to other biosignals, significantly contributing to big data cardiac screening through automated and reliable diagnostics.
