Clinically-aligned explainable AI for atrial fibrillation detection: A U -Net inspired multi-lead ECG analysis framework

Abstract

Background and Objective:

Deep learning offers high diagnostic accuracy but often lacks interpretability, hindering clinical adoption for Electrocardiogram (ECG) analysis. This study aims to develop a clinically interpretable, U-Net-inspired deep learning framework for Atrial Fibrillation (AFib) detection that elucidates model decision-making through morphological and rhythmic feature extraction.

Methods:

We utilized a U-Net-inspired encoder–decoder architecture to analyze 12-lead ECGs as spatially structured inputs, capturing inter-lead relationships using the PTB-XL database. Model interpretability was assessed using layer-wise Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize hierarchical feature focus. Additionally, we implemented RR interval variability assessment with heart rate stratification and group-specific threshold optimization to account for physiological variations.

Results:

On PTB-XL-Improved labels, the model achieved an Area Under the Curve (AUC) of 99.15%, Sensitivity of 97.81%, Specificity of 95.63%, and Precision (PPV) of 95.72%. Performance on the original PTB-XL annotations reached an AUC of 99.78%. Grad-CAM analysis confirmed hierarchical feature extraction, correctly localizing QRS complexes in over 97.5% of cases and P-wave and rhythm irregularities in over 93.1% of cases. Independent RR-interval variability analysis achieved AUCs of 95.5% (slow heart rate), 93.2% (normal), and 91.3% (fast) across three clinically stratified heart rate groups.

Conclusions:

This framework successfully combines high diagnostic performance with transparent, clinically aligned visual explanations. The principal contribution is not accuracy alone, but the demonstration that layer-wise explanations can be aligned with the clinical ECG reading sequence, from QRS localization to P-wave morphology. External validation remains necessary before broader clinical generalization can be claimed.

1. Introduction

Atrial fibrillation (AFIB) detection from 12-lead electrocardiogram (ECG) requires both high diagnostic accuracy and interpretable evidence that clinicians can verify. This section motivates the clinical need, summarizes gaps in current ECG-ML approaches, reviews representative high-accuracy diagnostic models and XAI methods, and introduces our clinically aligned framework and research questions.

1.1. Clinical context

AFIB represents a significant global health burden, affecting over 33 million individuals worldwide and increasing stroke risk five-fold [1]. AFIB is a cardiac arrhythmia characterized by rapid, irregular atrial contractions (see Fig. 1) that compromise effective blood flow and promote clot formation in the atria [2]. These clots can embolize to the cerebral circulation, causing ischemic stroke if left untreated [3]. Early detection and prompt intervention, with anti-coagulation therapy [4] to prevent thrombus formation, are cornerstones of arrhythmia management critical to reducing complications and mortality [5]. However, the silent and episodic nature of AFIB often leads to intermittent detection, resulting in delayed diagnoses and missed preventative opportunities [6].

Fig. 1.

Comparison of ECG signals between atrial fibrillation (AFIB) and normal sinus rhythm. (Top) AFIB signal shows irregular RR intervals with varying time distances between consecutive R-waves, characteristic of the chaotic atrial electrical activity. (Bottom) Normal sinus rhythm demonstrates consistent RR intervals with regular spacing between R-waves. In AFIB, distinct P-waves are absent, and any waveforms resembling P-waves may instead represent disorganized atrial activity. The normal rhythm, by contrast, shows clear P-waves preceding each QRS complex. The time axis represents samples over a 5000-sample window.

The electrocardiogram (ECG) is the clinical diagnostic aid for rhythm assessment, offering recordings of cardiac electrical activity from multiple anatomical perspectives, each captured by a distinct lead [7]. Yet, manual ECG interpretation is time consuming, requires specialized expertise, and is prone to human error [8]. Traditional diagnosis relies heavily on physician expertise and visual pattern recognition. However, this can result in mis-diagnosis, particularly in resource-limited settings or when dealing with subtle arrhythmic patterns. The integration of artificial intelligence into ECG analysis presents unprecedented opportunities to enhance diagnostic accuracy while maintaining clinical workflow compatibility [9]. Although AI-enhanced ECG tools show great promise, challenges related to clinical trust and interpretability remain barriers to full integration, underscoring the need for explainable automated diagnostic tools [10].

1.2. Gaps in current ECG-ML approaches

Deep learning has achieved high accuracy for automated ECG diagnosis, but its clinical adoption is hampered by the “black box” nature of high-performance models [11,12]. The lack of interpretability prevents clinicians from understanding why specific predictions are made, undermining trust and impeding integration into health care workflows. Explainable AI (XAI) is increasingly explored to validate model behavior, support patient safety, and address concerns about adopting deep-learning-based tools in clinical practice [13]. Generic explainability techniques like standard heat-maps often highlight irrelevant noise or evident morphological features rather than critical pathological features [14]. For instance, Local Interpretable Model-agnostic Explanations (LIME) does not address the unique characteristics of continuous ECG waveforms since it was designed for tabular, text, and image data [15]. Additionally, many models focus on single-lead ECGs, overlooking the diverse cardiac insights available from a full 12-lead recording.

1.3. Key technical and clinical advances

To overcome these limitations, we introduce a U-Net-inspired [16] encoder–decoder architecture that treats 12-lead ECG data as spatially structured input capturing inter-lead relationships critical for comprehensive rhythm analysis [17]. The encoder extracts fundamental ECG signals, detecting QRS complexes and other basic waveform components, while the decoder focuses on arrhythmia-specific anomalies such as missing/irregular P-waves and other AFIB markers, converting them into high-resolution attention maps interpretable by clinicians. To promote clinical trust and transparency, we integrate a layer-wise Gradient-weighted Class Activation Mapping (Grad-CAM) module that reveals a hierarchical sequence of feature extraction, beginning with core QRS detection and moving to P-wave analysis. This progression parallels the clinical ECG reading sequence and produces spatial feature maps aligned with established diagnostic markers. We also assembled and validated PTB-XL-Improved to address mislabeling issues: AFIB misclassification rates up to 4.7% [18].

Contributions.

This work advances clinically actionable AFIB diagnosis through four key methodological innovations: (1) a U-Net-inspired encoder–decoder that models inter-lead relationships in 12-lead ECG as spatially structured data; (2) layer-wise Grad-CAM visualization that addresses the black-box problem of deep learning by revealing hierarchical feature extraction aligned with clinical ECG reading steps; (3) a statistically rigorous RR-interval variability assessment (SDNN) that quantifies the irregular rhythm criterion of AFIB independent of deep-learning model outputs; and (4) PTB-XL-Improved labels that reduce diagnostic ambiguity and enhance both model reliability and clinical interpretability of results.

1.4. Related works

To contextualize our approach, this subsection summarizes prior work on deep learning for ECG arrhythmia detection and on explainable AI for ECG interpretation. We then briefly connect these themes to representative multi-channel physiological-signal studies that motivate our multi-lead modeling and clinically verifiable explanations.

Recent work on multi-channel physiological signals (e.g., EEG and sEMG) highlights the value of learning relationships across channels and combining signal processing with end-to-end deep models [19]. Similar themes apply to multi-lead ECG analysis, where generalization and interpretability remain key concerns. For example, mutual-information-based connectivity has been used to build EEG channel graphs for attention-based fatigue detection [20] and to derive normalized connectivity features for efficient classification [21]. Nonlinear decomposition plus temporal–spatial deep learning has also been reported for single-channel sEMG motion recognition [22].

1.4.1. Deep learning approaches for ECG arrhythmia detection

Early machine learning techniques for ECG arrhythmia classification using generic features and simple classifiers (k-Nearest Neighbors, Support Vector Machines) often struggled to achieve high accuracy, with reported performance around 70 to 80% [23]. However, more sophisticated feature engineering approaches that incorporate clinical knowledge have demonstrated substantially higher accuracy, particularly when applied to 12-lead ECG configurations. Jekova et al. developed atrioventricular synchronization metrics (RR-interval and PQ-interval variability, PQ-amplitude) combined with a dense neural network classifier, achieving 93.8% balanced accuracy (91.8% sensitivity, 95.8% specificity) on over 71,000 ECG records from six databases [24]. Tuboly et al. employed RR-interval Poincaré plot analysis with P-wave detection on average majority cycles, achieving 96.87 to 100% accuracy depending on arrhythmia subtype, including challenging cases with frequent premature beats [25]. Soni et al. utilized QT-interval variability and heartbeat features with error-correcting output codes, reporting 89.6% accuracy (100% sensitivity, 90% specificity) [26]. These approaches achieve high performance by explicitly modeling clinically meaningful ECG interval-based behavioral characteristics. Nevertheless, handcrafted feature design requires domain expertise and may not capture all subtle patterns that end-to-end deep learning can discover automatically [27,28]. Moreover, simple classifier performance degrades substantially in the presence of signal noise or short ECG segments [29].

In contrast, deep learning has emerged as a more powerful tool, showing greater robustness and achieving state-of-the-art predictions for arrhythmia detection [30,31]. Architectural innovations have been key to this success. One-dimensional Convolutional Neural Networks (1D-CNNs) have proven particularly effective at learning morphological variations directly from raw ECG signals, outperforming other architectures like RNNs in certain contexts [32]. For instance, specific 1D-CNN models have achieved accuracies of 98.28% for classifying five arrhythmia types and up to 99.84% on 5-second ECG segments [33,34]. Recent CNN advances include a 1D-CNN with squeeze-and-excitation attention and focal loss that achieved 99.48% accuracy on MIT-BIH without data augmentation [35], and a complex-valued CNN (C-AFNet) that transforms ECG segments into Gramian Angular Field images and achieves 95.63% accuracy on the CPSC2021 AF dataset with Grad-CAM and Integrated Gradient explanations [36]. Phukan et al. proposed a resource-efficient 1D-CNN that combines P-wave and RR-interval features for AF detection, achieving 99.44% accuracy with a 4.33 MB model suitable for edge deployment [37]. Accordingly, the central gap is no longer simply whether high discrimination can be achieved, but whether high-performing models can also provide clinically verifiable explanations of what waveform regions drive their predictions. A broader comparison of recent feature-engineering and deep learning methods is provided in Supplementary Table A.2.

Recurrent Neural Networks (RNNs), particularly Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) variants, are well-suited for capturing temporal dependencies in ECG data, with some studies reporting near-perfect PPV on curated single-lead datasets [38, 39]. Hybrid models that combine CNNs for feature extraction with RNNs for sequence modeling have also shown strong results, with a sequential CNN-Bi-LSTM model achieving an F1 score of 82% [40]. Dong and Xie recently proposed a multimodal CNN-LSTM with cross-attention that fuses time- and frequency-domain ECG features, achieving 98.97% accuracy on MIT-BIH with SMOTE-based rebalancing [41]. In the ResNet family, Li et al. introduced MGCNet, which combines a ResNet backbone with bidirectional gating and supervised contrastive learning for robust cross-dataset AF detection on AFDB and CPSC2021 [42]. Recently, attention mechanisms and transformers have been explored for ECG arrhythmia detection. Nankani et al. reported an F1 of 87% for AFIB with a transformer architecture [43], and Vision-Transformer-style encoders and hybrid CNN-transformer models have since followed. In parallel, self-supervised and contrastive pre-training have shown promise for improving generalization under distribution shift, exemplified by the supervised contrastive approach described above and broader pre-training strategies on large unlabeled ECG corpora surveyed in [44]. These two directions are largely orthogonal to our contribution: they optimize the representation space rather than enforce clinically aligned, waveform-level explanations, and token-based attention can make direct beat-level localization less straightforward than layer-wise convolutional feature maps. The choice of a convolutional encoder–decoder over transformer-based alternatives is methodological rather than incidental. Grad-CAM is mathematically defined over convolutional feature maps and their gradients with respect to spatial activations, and does not transfer directly to token-based attention architectures, which require fundamentally different attribution techniques such as attention rollout or attention flow. These techniques yield qualitatively different explanations and do not preserve the layer-wise, waveform-aligned hierarchy that constitutes the central interpretability contribution of this study. A convolutional backbone is therefore integral to the proposed clinical-alignment framework rather than a substitutable design element.

1.4.2. Explainable AI in ECG

The deployment of obscure high-performing deep learning models in clinical settings necessitates interpretability to ensure trust, regulatory compliance, and clinical acceptance. Regulatory bodies like the FDA are increasingly requiring explainability features in AI-based medical devices [45]. It is critical that these models can provide clear rationales for their predictions, empowering physicians to make more informed, data-driven decisions that enhance the quality of patient care [46].

To address this “black box” problem, a variety of XAI techniques are being adapted for ECG analysis. Saliency maps and Gradient-weighted Class Activation Mapping (Grad-CAM) are popular methods used to visualize which parts of the ECG waveform a model focuses on when making a diagnosis [47]. Such methods have been used to confirm that a model learns clinically relevant features, such as role of P-waves in AFIB detection [48] or to provide lead-wise interpretability in multi-lead ECG systems [49]. Other research has focused on generating clinician-friendly language explanations that have been subsequently verified by cardiologists [50].

However, applying XAI to time-series data such as ECGs poses unique challenges. Several studies have compared attribution techniques, including LIME, Shapley values, and gradient-based methods, to identify the most robust XAI methods for ECG interpretation [52]. Beyond ECG, related XAI strategies have been explored across medical domains: Shravya et al. integrated Grad-CAM alignment into the training loss to enforce saliency consistency in lung CT classification [53], while Anand et al. combined GAN-based augmentation with post-hoc LIME and SHAP heatmaps validated against dermatologist annotations for skin lesion classification [54]. In segmentation, Rajasekar et al. employed attention-enhanced generative autoencoders for lung imaging without incorporating XAI into the classification pathway itself [55], and Elakkiya et al. addressed extreme class imbalance in microRNA classification using a hybrid DNN with structural interpretability [56]. Our work differs from these approaches in that we operate on 1D multi-lead ECG time series rather than 2D images or tabular features, we validate that successive encoder–decoder layers correspond to successive stages of the clinical diagnostic sequence using quantitative beat-level fiducial comparisons rather than enforcing class-level saliency consistency or applying post-hoc model-agnostic explanations, and we complement deep learning explanations with an independent RR-interval variability analysis that provides model-free clinical evidence.

2. Methods

To bridge computational modeling with clinical interpretation, this section describes the full pipeline used in our study. We first introduce the dataset and expert label refinement, then describe signal pre-processing and the encoder–decoder model, and finally present the training configuration and interpretability analyses used for clinical validation.

2.1. Dataset and expert label refinement

Our study utilized the PTB-XL database, a comprehensive collection of 21,799 twelve-lead ECG recordings sampled at 500 Hz [57], containing 1599 AFIB cases and 20,200 non-AFIB cases. Because PTB-XL uses multi-label SCP diagnostic codes, we treated any recording carrying an AFIB or atrial flutter annotation as AFIB-positive; all remaining recordings were assigned to the non-AFIB control group, yielding a binary classification scheme that enforces sharp diagnostic boundaries while preserving clinical relevance.

To improve annotation quality, the database underwent expert reannotation as detailed by Marzocchi et al. [18]. Recordings with paced rhythm (297) were excluded, and a consensus protocol involving two independent expert cardiologists resolved 157 inter-annotator disagreements, yielding an overall AFIB annotation error rate of approximately 4.7% in the original labels. The resulting PTB-XL-Improved dataset contained 1529 AFIB and 19,973 non-AFIB recordings (Supplementary Table A.3).

The annotation quality assessment revealed pronounced variability, raising the possibility that poor labels could not only degrade algorithmic performance but also mislead XAI explanations. To examine this risk, we systematically evaluated whether annotation fidelity affected classification metrics (see Table 2). This underscores the critical need for high-quality annotations in clinical AI applications.

Table 2.

Comparative performance analysis of AFIB detection methods on the PTB-XL-Improved test set and the original PTB-XL test set. All methods were trained and evaluated as 12-lead ECG classifiers using the same patient-level train–test split. For the encoder–decoder, repeated training across 13 random seeds yielded empirical 95% confidence intervals within ±0.1% for the reported metrics.

Method	Dataset	Set	AUC	Sensitivity	Specificity	F1-Score	PPV
Encoder–decoder (ours)	PTB-XL-Improved	Test	99.15*	97.81	95.63*	96.75*	95.72*
CNN-BiLSTM [40,41]	PTB-XL-Improved	Test	97.15	95.62	90.62	93.29	91.06
ResNet [42,51]	PTB-XL-Improved	Test	97.27	98.75*	87.50	93.49	88.76
CNN [31,35–37]	PTB-XL-Improved	Test	79.18	82.19	60.31	74.08	67.44
Encoder–decoder (ours)	Original PTB-XL	Test	99.78*	99.01*	98.02	98.52*	98.04*
CNN-BiLSTM [40,41]	Original PTB-XL	Test	98.84	97.69	92.85	95.38	93.17
ResNet [42,51]	Original PTB-XL	Test	99.69	97.03	98.84*	95.15	93.33
CNN [31,35–37]	Original PTB-XL	Test	94.29	86.14	90.10	87.88	89.69

2.2. Clinically informed signal pre-processing

Our pre-processing pipeline was designed to replicate the systematic signal conditioning that experienced physicians implicitly apply during visual ECG interpretation (Fig. 2) [51]. This approach ensures that our computational framework operates on signals that align with clinical practice standards. We selected the filter cutoff and baseline correction method based on established ECG pre-processing practice and prior work [51], while prioritizing interpretability analyses over parameter ablations. We first applied a zero-phase Butterworth low-pass filter at 50 Hz (order 4) to suppress power-line and muscle noise while avoiding phase distortion via forward-reverse filtering [51]. Next, for each lead we estimated and subtracted the slow baseline drift using LOESS (Locally Estimated Scatterplot Smoothing). LOESS is a non-parametric method that fits low-degree polynomials to small, overlapping segments of the signal, modeling slow non-linear trends such as respiration- or electrode-induced drift.

Fig. 2.

Clinically-informed ECG signal pre-processing pipeline. The figure shows 12-lead ECG signals before (left column) and after (right column) the two-stage pre-processing approach. Each row represents a different ECG lead (I, II, III, aVR, aVL, aVF, V1-V6). The pre-processing consists of: (1) zero-phase Butterworth low-pass filtering at 50 Hz to remove power-line and muscle noise, and (2) LOESS-based baseline drift correction to eliminate respiratory and electrode-induced artifacts. The processed signals (right) demonstrate reduced noise and corrected baseline drift while preserving the diagnostic morphological features essential for atrial fibrillation detection. Pre-processing parameters follow Xie et al. [51].

2.3. Encoder–decoder network with skip connections for multi-lead ECG fusion

Our approach treats the 12-lead ECG as spatial data with a dimension of leads crossed by their time series, enabling the network to learn important patterns in the signal that are fundamental to clinical ECG interpretation. Since all 12 leads are recorded simultaneously, we preserve the temporal structure of each lead using 1D convolutions along the time axis. Consequently, we apply Conv1D layers along the temporal axis to capture the sequential dynamics of cardiac signals. This multi-lead fusion strategy aligns with evidence from multi-channel physiological signal analysis showing that preserving inter-channel structure improves accuracy and interpretability [20–22].

The initial convolutional layers serve a unique function in our framework. They transform the original 12 leads into learned feature maps that combine information across leads. For example, the first convolutional layer expands from 12 channels to 32, producing 32 learned feature maps that capture different aspects of inter-lead relationships. We refer to these as virtual leads in the sense of learned representations, not as clinically defined leads, and we do not claim direct clinical interpretability of individual feature maps.

The encoder comprises four convolutional blocks with progressively increasing channel dimensions, each followed by max pooling for temporal down-sampling; the decoder mirrors this structure with transposed convolutions, and skip connections from matching encoder stages preserve fine-grained temporal detail (architectural hyperparameters are detailed in Fig. 3). This multi-scale design enables the network to capture both local features (individual cardiac cycles) and global patterns (rhythm irregularities). At the bottleneck, a 512-channel layer captures high-level feature representations. We refer to this model as U-Net-inspired because it follows the symmetric encoder–decoder structure with skip connections, adapted to 1D temporal ECG signals.

Fig. 3.

Multi-scale ECG analysis and interpretability pipeline. (Top) U-Net-inspired encoder–decoder: 12-lead ECG input (5000 × 12) passes through four encoder blocks (yellow, filters 32–256) with max pooling, a 512-filter bottleneck (pink), and four decoder blocks (blue, 256–32) with skip connections. A classification head (GAP + FC + sigmoid) produces binary AFIB prediction. (Middle) Grad-CAM analysis: (A.1) early-layer attention on QRS complexes; (A.2) late-layer attention on rhythm irregularities and P-wave absence. (Bottom) Independent HRV pipeline: (B.1) R-peak extraction and rhythm-bound detection; (B.2) RR-interval variability with SDNN threshold classification (orange = above threshold).

The encoder–decoder architecture with skip connections ensures that this learned multi-lead spatial information is preserved and refined throughout the network hierarchy. The encoder captures features at multiple temporal scales, while the decoder reconstructs these features with enhanced spatial resolution. This architectural design facilitates hierarchical feature abstraction, enabling the network to learn complex AFIB patterns that may not be apparent from raw signals. The decoder produces interpretable feature maps that can be analyzed for clinical relevance. Finally, a classification head (comprising global average pooling, dropout with P = 0.5 for regularization, and a fully connected layer) aggregates these spatial–temporal features and produces a sigmoid-activated binary prediction for AFIB presence.

2.4. Training configuration

We employed a stratified 80/20 train–test split to ensure representative distribution of AFIB and control cases in both subsets. PTB-XL contains multiple ECG recordings per patient in a subset of cases, so the split was performed at the patient level using the provided patient_id to prevent patient overlap between training and test sets. To address class imbalance, we implemented random under-sampling of the control group to match the AFIB case count, creating a balanced training dataset that prevents bias toward the majority class. This choice supports our primary goal of clinically aligned interpretability by ensuring the model is not dominated by the non-AFIB class during training when learning and visualizing AFIB-relevant features. The composition of each subset and the resulting class ratios are shown in Table 1.

Table 1.

Composition of the stratified patient-level 80/20 split used for model development on PTB-XL-Improved. The AFIB:Non-AFIB ratio illustrates the severe class imbalance that motivated our under-sampling strategy.

Subset	AFIB	Non-AFIB	Total	AFIB : Non-AFIB
Training before under-sampling	1223	15,978	17,201	1 : 13.1
Training after under-sampling	1223	1223	2446	1 : 1
Held-out test set	306	3995	4301	1 : 13.1

Network optimization utilized the Adam optimizer with an initial learning rate of 5 × 10⁻⁴ and binary cross-entropy loss function. We employed a ReduceLROnPlateau scheduler that reduced the learning rate by a factor of 0.5 when validation loss plateaued for 5 consecutive epochs, enabling adaptive convergence. Training proceeded for 50 epochs with a batch size of 8, selected to balance computational efficiency with gradient stability.

Model performance was evaluated using comprehensive metrics: AUC (Area Under the Receiver Operating Characteristic Curve), F1-score, sensitivity, specificity, PPV, and NPV, defined as:

$$\begin{gathered} \text{Sensitivity}=\frac{TP}{TP+FN},\text{Specificity}=\frac{TN}{TN+FP},\mathrm{PPV}=\frac{TP}{TP+FP}, \\ \mathrm{NPV}=\frac{TN}{TN+FN},\text{F}1=\frac{2\cdot \mathrm{PPV}\cdot \text{Sensitivity}}{\mathrm{PPV}+\text{Sensitivity}} \end{gathered}$$

where TP, TN, FP, and FN denote true positives, true negatives, false positives, and false negatives, respectively. AUC summarizes the tradeoff between sensitivity and specificity across all classification thresholds. To validate our approach against established methods, we implemented baseline comparisons including a CNN-based AFibNet baseline [31], a CNN-BiLSTM baseline [40], and a multi-branching ResNet baseline [51]. All baseline models were trained as 12-lead ECG classifiers on the same patient-level training split, using the same pre-processing and sampling strategy, and were evaluated on the same held-out test set used for the encoder–decoder. These metrics provide a holistic assessment of classification performance relevant to clinical decision-making and enable robust comparison with state-of-the-art methods. To assess the statistical significance of performance improvements, we employed the DeLong test [58] to compare ROC curves between our method and each baseline on the same test set, with significance threshold set at P < 0.05.

2.5. Clinically aligned interpretability framework

We outline the interpretability framework used in our study. The central assumption is that the encoder–decoder network extracts ECG features in a hierarchical manner. Early encoder layers detect fundamental QRS complexes and beat boundaries. Intermediate layers (deeper encoder and bottleneck) digest beat-by-beat information to capture rhythm regularity and RR-interval patterns. Decoder layers reconstruct fine-grained attention over P-wave morphology and atrial activity relevant to AFIB. This assumption is grounded in the progressive down-sampling and up-sampling architecture, where each stage operates at a different temporal resolution. To validate this hierarchical feature extraction, we employ layer-wise Grad-CAM analysis across all leads, which visualizes the encoder–decoder model’s internal feature hierarchy. In addition, we introduce an independent RR-interval variability assessment (SDNN) as a separate, purely statistical method for AFIB analysis that quantifies rhythm irregularity directly from fiducial R-peak annotations without relying on the deep-learning model. These two methods produce separate results (reported independently in Sections 3.2 and 3.3) and are not combined into a single decision, but rather provide complementary diagnostic evidence from different analytical perspectives.

2.5.1. Layer-wise grad-CAM analysis

Clinical ECG interpretation follows a systematic hierarchical approach, where physicians first identify basic structures like QRS complexes [59], then analyze rhythm patterns, and finally examine subtle morphological features for diagnostic decision-making. We hypothesize that deep neural networks learn hierarchical ECG features. Early layers detect basic structures like QRS complexes, while deeper layers capture complex arrhythmia patterns.

To validate this clinical alignment hypothesis, we implemented layer-wise Grad-CAM analysis across all convolutional layers in our architecture. The implementation utilizes forward and backward hooks on all Conv1D layers to capture both activations and gradients during inference. The framework computes gradients and activations for each lead, enabling detailed analysis of attention patterns across the 12-lead spatial configuration. All attention maps are normalized to the [0, 1] range for consistent visualization and quantitative comparison.

For validating the model’s explainability, we designed hypothesis-driven validation tests to assess whether our network learns clinically relevant features in a hierarchical manner consistent with expert ECG analysis as follows:

1. Early Layer Analysis: QRS Detection Capability. We investigate whether early encoder layers consistently highlight QRS complexes in both AFIB and control ECGs, reflecting the fundamental role of QRS detection in clinical ECG analysis. For each recording in the test set (318 non-AFIB and 316 AFIB cases), we computed Grad-CAM activation maps from the first encoder layer and used PTB-XL+ fiducial annotations to score beat-level localization. A QRS complex was counted as correctly localized if the Grad-CAM peak within that beat fell inside the annotated QRS window (QRS-onset to QRS-offset). A recording was counted as a successful QRS detection only if all QRS complexes in that recording were correctly localized; otherwise it was marked as missed.

2. Deep Layer Analysis: Rhythm and Morphological Pattern Recognition. We evaluate whether decoder layers emphasize the P-wave region in both control and AFIB cases, corresponding to the advanced pattern recognition performed by experienced clinicians. Using the same fiducial annotations, we assessed whether the final decoder layer’s attention peak for each beat fell within the P-onset to S-peak window, which encompasses the atrial depolarization segment critical for AFIB diagnosis. A recording was counted as successful morphological detection only if all beats in that recording satisfied this localization criterion; otherwise it was marked as missed.

To systematically verify our hypothesis, we leveraged the PTB-XL+ dataset’s fiducial point annotations [60], which provide beat-level labels for R-peaks, P-wave onset/offset, QRS onset/offset, and T-wave offset across all 12 leads. The validation was performed on the entire test set (634 recordings: 318 non-AFIB, 316 AFIB). For each recording, Grad-CAM maps were computed at every convolutional layer, and the peak attention regions were compared against the corresponding fiducial windows to determine which specific ECG segments the model prioritized for diagnostic decision-making.

2.5.2. RR interval analysis

AFIB is also characterized by irregular ventricular depolarization, which manifests as increased variability in successive RR intervals [61]. Quantifying this variability offers a direct, physiologically grounded feature for rhythm classification and complements our broader framework by translating annotated fiducial points into a clinically interpretable metric.

Using PTB-XL+ fiducial annotations [60], we extracted R-peak locations for each recording and computed the corresponding beat-to-beat intervals $r_i=t_{i+1}-t_i$ (milliseconds), thereby preserving the original temporal resolution of the ECG. For each patient, the RR interval series was summarized by the standard deviation of RR intervals (SDNN, in milliseconds) [62].

To derive an interpretable, clinically aligned decision rule, we implemented a one-sided threshold model in which the threshold U was learned directly from the training data using binary cross-entropy loss. After training, the resulting U* defined a simple rule: if the RR variability (SDNN) for a patient exceeded U*, the patient was flagged as potentially having irregular beats; otherwise, the rhythm was classified as normal.

Based on clinically established heart rate ranges for bradycardia [63], normal rhythm [64], and tachycardia [65], we further stratified the analysis into three groups according to heartbeats per 10 seconds: slow (≤9 b/10 s), normal (10 to 16 b/10 s), and fast (≥17 b/10 s). Separate thresholds were learned for each group to capture rate-dependent variability patterns in our statistics.

3. Results

Guided by clinical AFIB criteria (irregular RR intervals and absent or disorganized P-wave activity), we first report classification performance and baseline comparisons on both original PTB-XL labels and PTB-XL-Improved labels, then quantify clinical alignment using fiducial-annotation-based attention localization and RR-interval variability (SDNN) analysis.

3.1. Classification performance

Our proposed model achieved higher accuracy than baseline methods on the PTB-XL-Improved test set. As detailed in Table 2, all models were trained as 12-lead ECG classifiers on the same patient-level training split and evaluated on the same held-out test set, enabling a direct comparison under identical data conditions. On the PTB-XL-Improved test set, our encoder–decoder model achieved an AUC of 99.15%, surpassing the CNN-BiLSTM [40], ResNet [51], and standard CNN [31]. DeLong test confirmed that all performance improvements were statistically significant (p < 0.01 for all baseline comparisons), demonstrating that the superiority of our method is not due to random variation. Across 13 repeated runs with different random seeds, all reported metrics (AUC, sensitivity, specificity, F1, PPV) remained within an empirical 95% confidence interval of approximately ±0.1%, supporting the stability of the reported results. Furthermore, when evaluated on the original PTB-XL test set using the same split, our encoder–decoder model achieved an even higher AUC of 99.78% and an F1-score of 98.52%, outperforming all other methods listed. These results demonstrate that our pre-processing pipeline and encoder–decoder model architecture effectively capture the complex spatial and temporal patterns characteristic of atrial fibrillation while maintaining high diagnostic accuracy.

To verify that the encoder–decoder model generalizes beyond the training distribution, we monitored training and validation loss throughout the optimization process. As shown in Fig. 4, both curves converge smoothly and track closely, with early stopping triggered at epoch 28.

Fig. 4.

Training and validation loss curves showing convergence without overfitting. Early stopping at epoch 28 (red dashed line) where validation loss reached minimum. Close tracking between training and validation losses with stable validation around 0.15–0.20 after epoch 20 indicates good generalization.

3.2. Clinical interpretability validation

The goal here is to evaluate whether the model’s learned features correspond to established practices in ECG interpretation. We begin with qualitative observations from figures that compare attention maps across layers in AFIB and sinus rhythm. We then provide quantitative validation of two separate components: (1) layer-wise attention localization against fiducial annotations for QRS and P-onset-to-S-peak regions, and (2) independent RR-interval discrimination using SDNN thresholds. The objective is to demonstrate that the model consistently highlights physiologically meaningful cues, while the RR-interval analysis separately quantifies rhythm irregularity, thereby supporting its application in clinical decision-making.

3.2.1. Layer-wise feature hierarchy

Our layer-wise Grad-CAM analysis revealed a clinically meaningful feature hierarchy that closely resembles the methodical diagnostic process used by expert cardiologists when analyzing ECG recordings. Early network layers consistently demonstrated sharp, focused attention on QRS complexes across both AFIB and control cases (see Fig. 5), confirming our hypothesis that fundamental cardiac anatomical features occur at shallow network depths. Under the recording-level criterion defined in Table 3, early layers successfully localized QRS regions in 317 of 318 non-AFIB recordings and 308 of 316 AFIB recordings.

Fig. 5.

Complementary Grad-CAM attention and RR-interval variability evidence. (A) AFIB case (Patient 05 631): early-layer attention on QRS complexes (red), late-layer attention on irregular rhythm and absent P-waves (blue); SDNN = 81.5 ms exceeds the learned threshold of 65.4 ms. (B) Normal sinus rhythm (Patient 10 829): regular QRS attention (red), consistent rhythm patterns with preserved P-waves (blue); SDNN = 30.9 ms remains below threshold. Green dots mark R-peaks. The two analyses are independent, complementary sources of evidence.

Table 3.

Recording-level and beat-level localization results for the layer-wise Grad-CAM validation. A recording was counted as successful only if all annotated beats satisfied the corresponding localization criterion. Beat-level metrics were computed by pooling beats across all evaluated recordings under the conservative assumption that each failed recording contained exactly one mis-localized beat. Because each beat produces exactly one Grad-CAM peak and has exactly one fiducial window, sensitivity equals PPV by construction.

Localization assay	Group	Successful recs	Mean beats/rec	Sens = PPV (%)
QRS (early encoder)	Non-AFIB	317/318 (99.7%)	11.9	99.97
	AFIB	308/316 (97.5%)	14.9	99.82
P-onset to S-peak (decoder)	Non-AFIB	296/318 (93.1%)	11.9	99.42
	AFIB	315/316 (99.7%)	14.9	99.98

Mid-level layers exhibited broadening attention patterns encompassing the PR segment and QRS complexes, with AFIB cases showing irregular attention spacing reflecting characteristic variable R-R intervals, while control cases maintained more consistent temporal patterns aligned with normal sinus rhythm. This demonstrates the network’s ability to learn rhythm-specific features at intermediate processing stages.

Deep layers predominantly exhibited attention patterns spanning from the T-wave offset to the Q-wave peak, indicating targeted search for the P-wave region. AFIB cases frequently showed fragmented or absent P-wave attention patterns, reflecting the arrhythmia’s underlying sinoatrial node dysfunction [65]. Under the same recording-level criterion, deeper layers localized the P-onset-to-S-peak region in over 93% of non-AFIB and over 99% of AFIB recordings (Table 3). Fig. 5 compares attention patterns between the first and last layers, showing early-layer localization of QRS regions versus late-stage capture of subtler P-wave irregularities. This hierarchical progression supports our clinical alignment hypothesis, demonstrating that the network learns diagnostic features in a sequence that parallels the clinical ECG reading approach.

3.3. RR interval boundary analysis

To assess the effectiveness of the RR variability-based approach, we evaluated the one-sided threshold model on the patient-level SDNN distribution, stratifying the cohort into three heart rate groups: slow (≤9 beats/10 s), normal (10 to 16 beats/10 s), and fast (≥17 beats/10 s). Performance within each heart rate group remained high, with AUCs of 95.5% for the slow group, 93.2% for the normal group, and 91.3% for the fast group, indicating that a simple interpretable cutoff on RR variability can provide strong discriminative power between AFIB and non-AFIB patients (Fig. 6). The learned SDNN thresholds were U* = 96.8 ms (slow), U* = 65.4 ms (normal), and U* = 46.1 ms (fast) (Fig. 6).

Fig. 6.

Heart rate variability analysis across different heart rate categories. (A) Distribution of heartbeat counts per 10-second segments, categorized into slow (<60 bpm), normal (60–100 bpm), and fast (>100 bpm) heart rates. (B–D) SDNN (standard deviation of RR intervals) distributions for slow, normal, and fast heart rate groups, respectively. Blue histograms represent non-AFIB cases, orange histograms represent AFIB cases. Dashed red vertical lines indicate the learned optimal thresholds (U) for each heart rate category: 96.8 ms (slow), 65.4 ms (normal), and 46.1 ms (fast). Heart rate groups follow clinically established ranges for bradycardia [63], normal rhythm [64], and tachycardia [65]. AFIB patients exhibit significantly higher SDNN values across all categories, with distinct optimal thresholds for each group.

However, some limitations were observed: false positives occurred in non-AFIB patients with high RR variability due to other arrhythmias or noise, while false negatives appeared in AFIB cases with relatively low variability. Importantly, some cases demonstrated irregular RR intervals without corresponding P-wave abnormalities, or vice versa. For this reason, we do not combine the two analyses into a single integrated decision rule; instead, we present morphological Grad-CAM attention maps and independent RR variability analysis as complementary evidence for clinicians.

Taken together, the layer-wise Grad-CAM localization and the independent RR-interval variability analysis provide two complementary lines of evidence. The former demonstrates that the encoder–decoder has learned a clinically coherent hierarchy of ECG features, while the latter confirms that a simple, model-free rhythm statistic can independently discriminate AFIB from non-AFIB patients across all heart rate strata. These converging findings motivate the discussion that follows.

4. Discussion

4.1. Clinical significance

Our clinically aligned explainable AI framework advances the integration of explainable AI in clinical ECG analysis between high-performance deep learning models and clinical decision-making requirements for AFIB detection. By consistently focusing attention on specific ECG segments (e.g., the R, T, and P wave regions), our method’s clinical importance is supported by several consistent findings.

First, the expert validation process introduces a dataset that has undergone partial clinical validation to address the 4.7% misclassification problem inherent in automated labeling systems [18]. This level of annotation quality is essential for developing clinically reliable AI systems, as diagnostic errors propagated from training data can have serious patient safety implications. Furthermore, our XAI framework helps identify when models focus on clinically irrelevant features, improving diagnostic reliability.

The spatial analysis capabilities of our encoder–decoder architecture enable comprehensive utilization of multi-lead ECG information, capturing inter-lead morphological relationships that are fundamental to clinical ECG interpretation [7]. Traditional single-lead approaches might miss critical diagnostic information available across the complete 12-lead configuration, particularly subtle P-wave abnormalities and spatial patterns of atrial depolarization [47].

The layer-wise interpretability analysis reveals a hierarchical feature extraction process that aligns with the systematic approach used in clinical ECG interpretation. The shallow layer analysis reveals that early layers primarily respond to key morphological features of the cardiac cycle, particularly QRS complexes, which serve as temporal landmarks for rhythm analysis [46]. This behavior aligns with the first step in clinical ECG interpretation, where clinicians establish fundamental cardiac rhythm and identify individual cardiac cycles before proceeding to more sophisticated pattern recognition. The decoder focuses attention on atrial irregularities like missing P-waves and erratic patterns that signal AFIB. This alignment between model behavior and clinical diagnostic patterns enhances physician confidence in AI-assisted diagnosis and facilitates seamless integration into existing clinical workflows (see Fig. 7).

Fig. 7.

Layer-wise visualization of U-Net attention dynamics during ECG interpretation for atrial fibrillation detection. Reading order: early encoder layers begin at top-left, progress through mid-level layers to the bottleneck at bottom-center, and continue through decoder layers ending at top-right. Red highlighting indicates high attention weights, demonstrating progressive feature learning from basic waveform detection to complex arrhythmia pattern recognition across the network hierarchy.

4.2. Technical innovations

Our encoder–decoder framework reinterprets multi-lead ECGs as spatially coherent data rather than independent time series, preserving inter-lead correlations that single-lead methods often overlook. Skip connections and multi-scale feature extraction enable simultaneous abstraction of high-resolution morphological details and lower-resolution rhythm dynamics, while the decoder reconstructs these representations at full spatial resolution to generate clinically interpretable attention maps that expose P-wave irregularities characteristic of atrial fibrillation [47].

Our fiducial-annotation-based evaluation confirmed that early network layers localized QRS regions in 99.7% of non-AFIB recordings (317/318) and 97.5% of AFIB recordings (308/316), while deeper layers localized P-onset-to-S-peak regions in 93.1% of non-AFIB and 99.7% of AFIB recordings. This structured validation confirms that network attention aligns with established physiological features rather than spurious correlations [46], providing clinicians with interpretable markers that guide focused review of diagnostic features [65].

Across both the original PTB-XL test set and the PTB-XL-Improved test set, our encoder–decoder model achieved AUCs above 99%, with recall and precision in the high 90 s, and it outperformed the CNN-BiLSTM [40], ResNet [51], and CNN [31] baselines evaluated under the same split and pre-processing protocol. DeLong testing further indicated that these improvements were statistically significant (P < 0.01 for all baseline comparisons). These results are also competitive with recent state-of-the-art methods evaluated on other ECG benchmarks, including attention-augmented 1D-CNNs [35], complex-valued CNNs with GAF representations [36], edge-optimized CNNs with P-wave and RR features [37], multimodal CNN-LSTM architectures with cross-attention [41], and ResNet-based contrastive learning frameworks for cross-dataset AF generalization [42]. Direct numeric comparison across these studies is not straightforward because they employ different datasets, segment durations, lead configurations, and evaluation protocols. Our controlled same-split comparison ensures that the reported improvements over the CNN, CNN-BiLSTM, and ResNet families reflect genuine architectural gains rather than protocol differences. These findings highlight our architecture’s ability to capture inter-lead relationships and generate attention maps that pinpoint clinically relevant ECG features without sacrificing sensitivity or specificity (see Table 2).

4.3. Comparison with previous research

Our systematic layer-wise analysis delivers more clinically interpretable and reliable explanations compared to previous model-agnostic approaches like LIME, which have demonstrated limitations when applied to time-series data [66]. This hierarchical feature extraction process parallels the clinical diagnostic sequence, where shallow layers systematically focus on fundamental cardiac structures while deeper layers emphasize diagnostic patterns, demonstrating the model’s hierarchical feature learning (see Fig. 7).

It is also important to interpret these results in the context of prior high-performing feature-engineering studies rather than framing the comparison as deep learning versus weak traditional methods. Clinically informed approaches based on RR variability, P-wave analysis, PQ features, or QT variability have already demonstrated strong AFIB discrimination, in some cases exceeding 90% accuracy [24–26]. To provide a fair and comprehensive landscape, Supplementary Table A.2 presents a broader comparison that includes these feature-engineering studies alongside our same-split deep learning baselines and recent state-of-the-art architectures from different benchmarks. The present work therefore aims to complement, rather than dismiss, such methods by combining strong discrimination with a layer-wise explanation framework and an independent SDNN-based rhythm analysis. To further enhance reliability, we addressed data quality concerns through PTB-XL-Improved labels, which underwent targeted partial relabeling [18] to provide enhanced annotation fidelity compared to synthetic data augmentation methods commonly employed in prior studies [32]. This comprehensive approach, combining architecturally-driven interpretability with higher-quality annotations, creates a robust foundation for clinically interpretable AFIB analysis while external validation remains a necessary next step.

4.4. Edge cases

Grad-CAM attention maps demonstrate that our model generalizes effectively beyond the regions emphasized during training, consistently localizing clinically relevant abnormalities in challenging ECG morphologies. For instance, in Patient #128 (Fig. 8a), the decoder’s attention map precisely highlights the absence of the T-wave, despite its atypical presentation. In Patient #271, the first encoding layer accurately detects an anomalously shaped QRS complex (Fig. 8b), and the second encoding layer further refines this localization to the correct QRS complex position (Fig. 8c). These results indicate that our architecture captures both fine-grained morphological features and broader temporal sequence information, maintaining robustness in the face of irregular and edge case signals.

Fig. 8.

Edge-case saliency analysis demonstrating model robustness. (a) Patient #128: Decoder attention highlights the absent T-wave despite its atypical presentation, confirming the model’s ability to detect waveform absence rather than only waveform presence. (b) Patient #271: First encoder layer detects a region where an R-peak is missing due to recording artifacts. (c) Patient #271: Second encoder layer correctly localizes the remaining R-waves despite the misplaced beat, demonstrating generalizability and robust cardiac cycle understanding even under artifact-corrupted conditions.

4.5. Limitations and future work

Despite these promising results, all performance and interpretability metrics were derived solely from PTB-XL, so external generalizability remains unproven and any clinical applicability claims should be interpreted within this internal evaluation setting. Future work will evaluate the encoder–decoder framework on independent cohorts such as CPSC [67], Chapman-Shaoxing [68], and clinical partner data. A practical challenge is that these external public datasets do not currently provide the beat-level fiducial annotations required for our present layer-wise localization assay, so an external validation pipeline will require either automated fiducial extraction or new annotation workflows.

We also note that random under-sampling reduces the diversity of non-AFIB training examples and may affect generalization and the stability of learned decision thresholds. Future work will explore alternative imbalance strategies such as class-weighted loss functions, focal loss, and balanced sampling without discarding control recordings.

We did not perform a direct single-lead-vs-multi-lead ablation here because clinical AFIB interpretation relies on inter-lead morphology and the 12-lead configuration is the operationally relevant input. The lead-count comparison itself has been examined separately in our companion work on myocardial infarction detection [69], and the analogous AFIB ablation on PTB-XL with the present encoder–decoder is a planned follow-up.

Finally, we note that Grad-CAM does not constitute a formal faithfulness guarantee. However, the fundamental problem with black-box models is the complete absence of insight into what internal logic drives predictions. Our layer-wise Grad-CAM analysis addresses this by demonstrating that, unlike other high-performing models that remain entirely opaque, the encoder–decoder has developed a structured, hierarchical understanding of ECG signals (progressing from QRS localization in early layers to P-wave and rhythm analysis in deeper layers) that parallels the clinical diagnostic sequence. This alignment, validated quantitatively against fiducial annotations at the beat level (Table 3), provides evidence that the model’s internal representations are clinically meaningful rather than relying on spurious correlations. Grad-CAM should therefore be interpreted as supportive evidence of meaningful feature learning, not as proof of causality. Additional faithfulness tests such as input perturbation and random-weight sanity checks are planned for future work. The choice of Grad-CAM is grounded in prior cross-method evidence. A recent comparative study of LIME, SHAP, and Grad-CAM for ECG-based AFIB detection under matched clinical-alignment criteria [70], and our broader systematic review of explainable AI in electrocardiography [44], both find that Grad-CAM produces more clinically coherent, waveform-aligned localizations on 1D ECG inputs than LIME or SHAP, the latter two having been designed primarily for tabular and image data and frequently failing to align with QRS, P-wave, or RR-interval landmarks. This motivates the methodological choice in the present study.

We also acknowledge that probabilistic calibration metrics such as the Expected Calibration Error and Brier score were not evaluated in this study. Although our focus was on clinically verifiable explainability rather than risk calibration, such metrics are important for clinical deployment and will be addressed in future work.

Additionally, while our heart rate stratification follows clinically established ranges for bradycardia [63], normal rhythm [64], and tachycardia [65], the globally optimized RR-interval thresholds within each group may not fully capture inter-patient variability, so we will develop personalized or hierarchical thresholding approaches including meta-learning and Bayesian modeling to adapt bounds on a per-case basis.

5. Conclusion

We developed a U-Net-inspired encoder–decoder architecture that successfully combines high-performance atrial fibrillation detection with clinically meaningful explanations. Our model achieved AUCs of 99.15% on PTB-XL-Improved labels and 99.78% on original PTB-XL annotations, outperforming the CNN, CNN-BiLSTM, and ResNet baselines evaluated in this study and comparing favorably with prior high-performing ECG methods reported in the literature.

Layer-wise Grad-CAM analysis revealed hierarchical feature extraction that aligns with clinical diagnostic patterns, with early layers detecting QRS complexes and deeper layers emphasizing P-wave morphology critical for AFIB diagnosis. Independent RR-interval variability analysis (SDNN) provided AUCs of 95.5% (slow heart rate), 93.2% (normal), and 91.3% (fast) across clinically stratified heart rate groups, offering complementary rhythm-based evidence that does not depend on the deep learning model. The targeted expert relabeling process addressing the 4.7% misclassification rate in public datasets [18], combined with our systematic interpretability framework, represents a significant step toward clinically interpretable AFIB analysis. While external validation remains necessary, our findings demonstrate the feasibility of explainable deep learning models that align with clinical ECG interpretation practices.

Appendix. Supplementary material

This supplement contains three tables. Table A.1 reports the learned SDNN thresholds and AUC for each heart rate group on the full (unbalanced) cohort. Table A.2 provides a broader comparison of AFIB detection methods across feature-engineering and deep learning studies. Table A.3 summarizes dataset composition before and after expert reannotation.

Table A.1.

Group-wise SDNN thresholds and performance of the RR-interval variability baseline on the full (unbalanced) cohort.

Heart rate group	Learned threshold U* (ms)	AUC (%)
Slow (≤9 beats/10 s)	96.8	95.5
Normal (10 to 16 beats/10 s)	65.4	93.2
Fast (≥17 beats/10 s)	46.1	91.3

Table A.2.

Broader comparison of AFIB detection methods, including feature-engineering approaches and recent deep learning architectures. Direct numeric comparison across studies should be interpreted with caution because datasets, lead configurations, segment lengths, class definitions, and evaluation protocols differ. Our same-split baselines (rows marked with †) provide a controlled comparison under identical conditions.

Study	Method	Dataset (leads)	Key metric	XAI	Year
Feature-engineering approaches
Jekova et al. [24]	AV-sync (RR, PQ variability) + DNN	6 databases (12-lead), 71k+ records	93.8% bal. Acc	No	2022
Tuboly et al. [25]	RR Poincaré + P-wave detection	Lead I, AF vs. non-AF arrhythmias	96.9%–100% Acc	No	2021
Soni et al. [26]	QT variability + ECOC classifier	MIT-BIH AF	89.6% Acc, 100% Sens	No	2022
Same-split baselines on PTB-XL-Improved (this study) †
CNN† [31]	1D-CNN (AFibNet)	PTB-XL-Improved (12-lead)	79.18% AUC	No	2021
CNN-BiLSTM† [40]	CNN + Bidirectional LSTM	PTB-XL-Improved (12-lead)	97.15% AUC	No	2022
ResNet† [51]	Multi-branching ResNet	PTB-XL-Improved (12-lead)	97.27% AUC	No	2024
Ours †	Encoder–decoder + Grad-CAM + SDNN	PTB-XL-Improved (12-lead)	99.15% AUC	Yes	2026
Recent deep learning methods (different benchmarks, for context)
Guhdar et al. [35]	1D-CNN + SE attention + focal loss	MIT-BIH, PTB (2-lead)	99.48% Acc	No	2025
Mahali et al. [36]	Complex-valued CNN (C-AFNet)	CPSC2021 (Lead I/II)	95.63% Acc, 94.15% F1	Grad-CAM, IG	2026
Phukan et al. [37]	1D-CNN + P-wave/RR features	5 databases (1/2/12-lead)	99.44% Acc	No	2025
Dong & Xie [41]	CNN-LSTM + cross-attention	MIT-BIH, PTB (single-lead)	98.97% Acc	Cross-attn map	2026
Li et al. [42]	ResNet + contrastive learning	AFDB, CPSC2021 (single-lead)	SOTA cross-dataset	No	2026

Table A.3.

Dataset composition before and after expert reannotation.

	Original PTB-XL	PTB-XL-Improved
AFIB recordings	1599	1529
Non-AFIB recordings	20,200	19,973
Excluded (paced rhythm)	–	297
Total	21,799	21,502

Code availability

Training, Grad-CAM, and RR-interval analysis code: https://github.com/Luo-Innovation-Lab/AFIB_PTBXL.

Data availability

The PTB-XL database is freely available through PhysioNet (https://physionet.org/content/ptb-xl/), a repository maintained by the MIT Laboratory for Computational Physiology. Access to this dataset requires completion of PhysioNet’s credentialing process and agreement to data use terms prohibiting re-identification attempts and requiring compliance with applicable privacy regulations. All analysis code has been made publicly available at https://github.com/Luo-Innovation-Lab/AFIB_PTBXL to promote reproducibility and transparency.