An explainable AI for trustworthy detection of atrial fibrillation on reduced lead ECGs in mobile applications

Abstract Introduction Artificial intelligence (AI), particularly deep learning (DL), has demonstrated high performance in various diagnostic problems, including the detection of paroxysmal atrial fibrillation (AF) from electrocardiograms (ECGs). In mobile applications, these approaches can help to detect AF at an early stage and thus prevent possible secondary diseases. However, due to their black-box nature, DL approaches lack explainability and interpretability. Furthermore, 12-lead ECGs are the clinical standard, but mobile devices usually only provide 1–3 leads, which are similar but not identical to Einthoven leads and differ in morphology. Purpose We apply a recently introduced explainable ECG analysis architecture (xECGArch) to single-lead ECGs to verify its performance in mobile applications. xECGArch can learn rhythmic and morphological characteristics using two parallel convolutional neural networks of different dimensions (Fig. 1). When combined with methods from explainable AI, the classification can be traced back to rhythmic and morphological characteristics. Methods We used a subset of 9 854 ECGs (n(AF) = n(non-AF) = 4 927) from 4 public databases (Chapman-Shaoxing, CPSC2018, Georgia, and PTB-XL). Noise was removed using a cascade of high-pass filtering at 0.3 Hz and discrete wavelet transform. For each lead, we trained xECGArch on single-lead ECGs of 10 s length to classify AF vs. non-AF using 90% of the ECGs with 5-fold cross-validation and tested it on the remaining randomly selected ECGs (n(AF) = 507, n(non-AF) = 479). Consistent data splitting was ensured for all leads. Model explanations were generated using deep Taylor decomposition. Results Both models achieved minimum 91.1% accuracy for each ECG lead (Fig. 1). On lead I, the morphology model outperformed the rhythm model with 92.8% vs. 91.5% accuracy, while the rhythm model slightly outperformed the morphology model on leads II and III with 94.2%–95.0% accuracy vs. 91.1%–94.0%. Both models achieved maximum accuracy of 95.0% (rhythm) and 94.0% (morphology) on lead II. By combining both models, xECGArch reached 93.7%–95.1% accuracy. The model explanations confirm that the rhythm model primarily considers QRS complexes as relevant, while the morphology model focuses on fibrillatory waves (Fig. 1). Conclusions Both rhythm and morphology models achieved reliable results across all leads, competing with the state-of-the-art for automatic AF detection from single-lead ECGs. The combination of both models in xECGArch further increased the accuracy. xECGArch is based on the medical reading of ECGs by considering rhythm and morphology characteristics and is therefore interpretable by design. The model explanations are consistent with the diagnostic criteria for AF (Fibrillatory waves and absolute arrhythmia) across leads. With high accuracy on different single-lead ECGs and improved interpretability, xECGArch provides a trustworthy method for ECG analysis in mobile applications.