ECG-XPLAIM achieved 95.5% sensitivity, 96.9% specificity, and 98.9% AUROC for AFib detection externally, with high performance and interpretability across arrhythmia types.
Does ECG-XPLAIM accurately detect arrhythmias and provide explainable outputs compared to existing AI models in large-scale ECG datasets?
ECG-XPLAIM provides high diagnostic accuracy for arrhythmia detection across diverse datasets while offering explainable outputs via Grad-CAM to enhance clinical trust.
Tasa de eventos absoluta: 0% vs 0%
Abstract Background Automated electrocardiogram (ECG) analysis using deep learning has shown promising potential in arrhythmia detection. However, many of the proposed models lack robustness, interpretability or clinical trust. Explainable artificial intelligence (XAI) techniques, such as Gradient-weighted Class Activation Mapping (Grad-CAM), enhance interpretability by highlighting ECG regions influencing model predictions. Purpose To develop ECG-XPLAIM, an eXPlainable Locally-adaptive Artificial Intelligence Model for arrhythmia classification from Electrocardiograms, evaluate its diagnostic performance and interpretability, and compare it with existing models. Methods ECG-XPLAIM was designed with an Inception-inspired convolutional neural network (CNN) architecture. The model was trained on MIMIC-IV, a large publicly available ECG dataset comprising more than 800,000 ECG samples, and evaluated both internally, on a held-out subset of MIMIC-IV, and externally, on PTB-XL, another large-scale ECG dataset of more than 25,000 samples. Sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC) were computed for multiple arrhythmia classification tasks. Performance was benchmarked against a baseline CNN, an advanced gated recurrent unit (GRU) network, and a pre-trained model developed externally. Grad-CAM-derived attention maps were generated to assess interpretability. Results ECG-XPLAIM demonstrated high performance in internal validation, with all metrics exceeding 90% across all tasks. External validation on PTB-XL confirmed the model's generalisability, achieving a micro-averaged sensitivity of 95.5%, specificity of 96.9% and ROC-AUC of 98.9% in distinguishing between atrial fibrillation (AFib), sinus tachycardia and non-arrhythmia samples. For conduction disturbances, ECG-XPLAIM achieved a sensitivity of 82.1%, a specificity of 95.2% and an AUROC of 95.0%, in detecting right bundle branch block (RBBB), left bundle branch block (LBBB) and left anterior fascicular block (LAFB) from normal ECGs. Micro-averaged sensitivity, specificity and AUROC were 69.1%, 86.4% and 87.8% for long QT syndrome (LQT), 77.3%, 97.3% and 89.5% for Wolff-Parkinson-White (WPW) syndrome and 96.0%, 98.8% and 99.3% for pacemaker detection, respectively. ECG-XPLAIM maintained comparable or superior performance relative to baseline and external models (summary metrics in Figure 1), while Grad-CAM visualisations revealed relevant waveform regions influencing predictions, enhancing model explainability (Figure 2). ECG-XPLAIM architecture and trained weights are available online. Conclusion ECG-XPLAIM achieves a strong balance between diagnostic accuracy and interpretability, demonstrating its potential as a valuable tool for AI-assisted ECG analysis. By offering both high-performance classification and explainable outputs, it can enhance clinical decision-making supporting both experts and non-experts in arrhythmia detection.Figure 1.Summary performance metrics Figure 1.Grad-CAM visualisations
Pantelidis et al. (Sat,) reported a other. ECG-XPLAIM achieved 95.5% sensitivity, 96.9% specificity, and 98.9% AUROC for AFib detection externally, with high performance and interpretability across arrhythmia types.