Introduction
Artificial Intelligence (AI), particularly deep learning (DL), has demonstrated high performance in diagnostic issues, including the detection of cardiovascular diseases (CVDs) from the electrocardiogram (ECG). In long-term monitoring, AI approaches can improve early detection of CVDs and increase the chance of detecting paroxysmal cases. However, due to their black-box nature, the decision making of DL lacks explainability and their self-learned features lack interpretability. Nevertheless, both are essential for physicians to assess AI-based diagnostic recommendations, which is a prerequisite for clinical integration.
Purpose
We demonstrate the capabilities of the self-learning DL architecture for explainable ECG analysis (xECGArch) in detecting atrial fibrillation (AFib), atrial flutter (AFlut), and 1st-degree AV block (AVB-I). xECGArch combines 2 models, with one focusing on rhythm and the other on morphology. Methods from explainable AI (xAI) generate explanations for the individual decisions of both models. Their fused representation (xFuseMap) enables interpretability in line with clinical knowledge.
Methods
xECGArch was trained to detect AFib, AFlut, and AVB-I using lead II ECGs from 5 public databases (Chapman-Shaoxing/Ningbo, CPSC2018, Georgia, PTB, PTB-XL). The CVD data was balanced with reference ECGs, containing 90% pathological and 10% normal sinus rhythm ECGs (n(AFib)=n(non-AFib)=4,927, n(AFlut)=n(non-AFlut)=8,374, n(AVB-I)=n(non-AVB-I)=3,530). We used the middle 10 s of each ECG, discarding shorter ECGs. The training was conducted on randomly selected 90% of the ECGs in a 5-fold cross-validation with testing on the remaining ECGs. Model explanations were generated using deep Taylor decomposition, the most reliable in a systematic comparison of 13 xAI methods.
Results
The rhythm and morphology models achieved 92.4%−95.0% accuracy for AVB-I and AFib, increasing to 94.1%−95.3% using xECGArch (Tab. 1). For AFlut, xECGArch reached 91.1% accuracy caused by a lower precision of 85.5%. The model explanations (Fig. 1) confirm that the rhythm model focuses on QRS complexes, especially for AFib detection, while the morphology model focuses on clinically relevant morphology features, like fibrillatory or flutter waves and the absence of P waves in AFib, or the PQ duration in AVB-I.
Conclusions
Both rhythm and morphology models reliably detect AFib, AFlut, and AVB-I from single-lead ECGs, competing with state-of-the-art methods. Their combination in xECGArch further improves performance. xECGArch is inspired by the medical reading of ECGs by considering rhythm and morphology characteristics and is therefore interpretable by design. The explanations align with diagnostic criteria for AFib, AFlut, and AVB-I. With high accuracy and improved interpretability, xECGArch provides a trustworthy method for automated ECG analysis and meets the prerequisite for integration into clinical routine for diagnostic support.