Enhancing Arrhythmia Diagnosis with Data-Driven Methods: A 12-Lead ECG-Based Explainable AI Model

Cardiovascular disease (CVD) is the leading cause of death in Europe and the EEUU, causing 3.9 million and 1.8 million deaths annually [1]. Traditional CVD diagnosis relies on rule-based evaluation of patient history and clinical examinations. This approach struggles with the volume and diversity of data and depends heavily on medical expertise, leading to challenges in resource-limited settings like developing countries.

The electrocardiogram (ECG) is a key, non-invasive tool for diagnosing cardiac conditions, utilizing a 12-lead setup to capture heart’s electrical activity through distinct P, Q, R, S, and T waves [2]. While ECG, especially in identifying cardiac arrhythmias, is straightforward, interpreting these signals, particularly in complex cases, remains challenging and prone to errors with serious implications [3]. Additionally, Heart Rate Variability (HRV) analysis, which examines variations in consecutive heartbeats, has emerged as a crucial technique in cardiac assessment. It assesses the autonomic nervous system’s impact on the heart by analyzing the R-R interval, the time between successive R wave peaks, and the N-N interval, the duration between consecutive QRS complexes. These measures help in understanding the cardiac system’s dynamic state [4].

Arrhythmias, a common and varied group of CVDs diagnosed using ECG, are characterized by irregular heartbeats due to improper electrical signaling, leading to abnormally fast, slow, or inconsistent heart rhythms. This work focuses on several arrhythmia classes including Atrial fibrillation (AF), Right and Left Bundle Branch Blocks (RBBB and LBBB), First-degree atrioventricular block (IAVB), and Premature Atrial and Ventricular Contractions (PAC and PVC), along with Myocardial Infarction (MI) [5]. Diagnosing arrhythmias is challenging due to: i) absence of symptoms during ECG recording; ii) high inter-patient ECG signal variability; iii) non-stationary signal morphology affected by physical state, noise, and artifacts; and iv) the need for large data volumes to avoid false diagnoses [5].

Computer-Aided Diagnosis Systems (CADS) address arrhythmia diagnosis challenges by leveraging digital technologies for the analysis of physiological and clinical data, aiding clinicians in making more informed decisions. Traditional ECG analysis techniques in CADS rely on automated detection of ECG components and classifying them based on fixed rules, but they often fall short due to outdated rules and sensitivity to imperfect ECG recordings. In the medical field, Artificial Intelligence (AI), particularly Machine Learning (ML) and Deep Learning (DL), has significantly enhanced CADS. AI combines mathematical and computer science theories to create systems capable of intelligent actions, with DL being notable for its ability to process large volumes of data through artificial neural networks. These networks perform sequential transformations to highlight crucial input features for classification and regression tasks. Modern arrhythmia diagnosis models increasingly use DL, credited for its precision in identifying ECG waveforms like QRS complexes, and P and T waves, facilitating the calculation of vital clinical measures including heart rate and axis deviation [6, 7].

AI’s potential in various fields, including healthcare, is often hindered by its ‘black box’ nature, leading to trust issues due to a lack of transparency [8]. Healthcare professionals need to understand the reasoning behind AI-recommended treatments. Without this level of explainability, AI’s adoption in healthcare can be negatively impacted. Explainable AI (XAI) addresses this challenge by providing insights into the decision-making processes of AI systems. XAI aims to make the logic behind AI algorithms clear, thereby aligning advanced AI capabilities with the healthcare sector’s need for transparent decision-making.

In recent literature, advancements in arrhythmia classification using DL have been notable, especially with the application of Convolutional Neural Networks (CNNs). A diverse range of studies has utilized various DL architectures, showing significant progress in ECG signal analysis. Chen et al. [9] combined a CNN with RestNet-34 layers and bi-directional LSTM, achieving an accuracy of 0.81 on 12-lead ECG samples. Their study, while promising, highlighted the need for balanced datasets. Cheng et al. [10] used a modified 1-D CNN on the MIT-BIH arrhythmia database, focusing on compressed ECG signals suitable for wearable devices. Gao et al. [11]‘s approach involved a 4-layer LSTM model, achieving an accuracy of 0.992, demonstrating robustness against noise and normal ECG beat dominance. Niu et al. [12] introduced a novel DL method based on adversarial domain adaptation, while Romdhane et al. [13] focused on enhancing minority class classification accuracy. Wang et al. [14, 15] employed 1-D CNNs and continuous wavelet transform techniques, demonstrating effectiveness against noise. Yildirim et al. [16] developed a 1D-CNN suitable for mobile and cloud computing applications due to its efficiency. Zhang [17] and Zhang et al. [18] used 1D-CNN networks, showing the superiority of 12-lead over single-lead ECGs. Rai et al. [19] tested a CNN + LSTM ensemble approach, improving minority class accuracy. Finally, Toma et al. [20] presented a parallel approach combining RNN and 2D CNN, effectively capturing temporal and spatial ECG signal characteristics. These studies illustrate the advancements and diversity in DL applications for arrhythmia detection, with a trend towards optimizing network architectures and inputs for enhanced classification accuracy. However, the integration of explainable AI (XAI) remains a largely unexplored area, crucial for the clinical applicability of these models.

In our study, we developed an explainable model for detecting cardiac arrhythmias using 12-lead ECG signals. We explored two approaches: one using raw ECG signals as input for various DL architectures like ResNet, VGG, and DenseNet, and another combining raw ECG signals with HRV features in a hybrid model. We thoroughly evaluated the performance of these classifiers. Furthermore, we assessed the explainability of the most effective model by analyzing the significance of different leads in arrhythmia classification and presenting case examples that illustrate the ECG signal segments influencing the predictions.

The remainder of this paper is organized as follows: Section 2 outlines the dataset, the DL algorithms, training and testing methodologies, and the explainability technique used in building the arrhythmia detector. Section 3 details the evaluation results, including the predictive performance of the DL approaches (using raw ECG and the hybrid method with HRV features) and the explainability analysis of the most effective model. Section 4 discusses these results, and Sect. 5 concludes the paper with key findings.

This section details the classification outcomes of two methods examined in our study: DL with raw ECG signal, and a hybrid method combining DL with HRV and raw ECG signals. For each method, we highlight the algorithm with the highest performance, based on AUROC scores, and present its confusion matrix and ROC curve.

Moreover, we delve into the approach that yielded the most accurate classification, emphasizing explainability. This involves analyzing the significance of different ECG leads in identifying arrhythmia classes and providing explanations for individual instances, specifically highlighting the ECG segments that contributed to arrhythmia classification.

ML and DL are increasingly important in CVD detection, leveraging ECGs as key data sources. ECGs are crucial for improving diagnostic accuracy in data-driven predictive models for CVD [29]. These technologies not only facilitate early CVD detection, leading to better health outcomes, but also help address the demand for skilled cardiologists in ECG data analysis. Recent research in clinical cardiology suggests that ML and DL, especially in combination with other methods, provide superior predictive power for cardiovascular or overall mortality compared to traditional clinical or imaging techniques alone [7]. DL’s strength lies in its ability to capture temporal signal variations and autonomously learn from complex inputs like ECG signals, provided there’s enough high-quality training data. This learning bypasses the need for predefined feature processing, offering an end-to-end solution that minimizes errors in feature calculation, thereby enhancing model accuracy [30‐34].

In our study, we developed an explainable model for detecting cardiac arrhythmias using 12-lead ECG signals by exploring two different approaches: one using raw ECG signals as input for various DL architectures, and another combining raw ECG signals with HRV features in a hybrid model. We identify the best-performing model in the multiclass arrhythmia prediction by assessing the AUROC for each arrhythmia category and their average. If classifiers exhibit equal AUROC values, we also evaluate the F1-score, an important metric in multiclass classification scenarios.

In our evaluation of DL models for arrhythmia detection, both ResNet models outperformed DenseNet and VGG16, with ResNet34 and ResNet50 showing similar effectiveness. However, ResNet34 marginally leads as the optimal model with an AUROC of 0.98 and an F1-score of 0.83, compared to ResNet50 (AUROC: 0.98, F1-score: 0.81), DenseNet (AUROC: 0.96, F1-score: 0.77), and VGG16 (AUROC: 0.95, F1-score: 0.76). By inspecting the different ROC curves and the confusion matrix for individual arrhythmia classifications, ResNet34 excels in detecting SNR, LBBB, RBBB, PAC, and STD, while ResNet50 is superior in classifying AF, IAVB, PVC, and STE.

The hybrid approach, combining ResNet34 with Heart Rate Variability (HRV) features, did not enhance performance in any arrhythmia category, including overall average. Notably, it showed reduced accuracy and precision, especially in AF and RBBB categories. This suggests that instead of providing ResNet34 with additional informative features, the HRV integration might have impeded the network’s learning, leading to a decline in classification performance. Thus, the integration of HRV features appears to compromise rather than improve the model’s efficacy in arrhythmia classification.

This study’s results indicate that the employed DL algorithms have the potential to aid clinicians and healthcare professionals in detecting cardiac conditions that might otherwise be missed or diagnosed later through specialist evaluations or echocardiography. Leveraging both short-term and long-term learning, these methods enable early detection of cardiovascular diseases (CVD), facilitating timely treatment initiation. This early intervention can lead to improved health outcomes, while delays or missed diagnoses could exacerbate health conditions.

ML and DL tools offer precise predictions but their ‘black box’ nature poses significant interpretability challenges, hindering clinician acceptance due to difficulties in understanding decision-making processes. This lack of transparency is especially problematic in clinical practice, where clinicians need clarity for effective decision-making. The interpretability and application of these advanced models, particularly in identifying critical features, remain complex, potentially limiting their utility in settings without computer assistance. This opacity significantly impedes the adoption of AI models by healthcare professionals who require comprehensible explanations of AI-derived results. In response, the emergence of XAI is a critical development, aiming to demystify AI models and outcomes to enhance accessibility and user trust. XAI facilitates the identification of key features influencing predictions and explores causal relationships between features and clinical outcomes. Despite its importance, research focusing on the understandability and trust in ML models, particularly those aiding in the diagnosis or prognosis of CVD, is still scarce, highlighting a vital area for future exploration.

In our study, we employed the SHAP technique to analyze explainability at both global and individual levels using the 12-lead ECG data. Globally, SHAP enabled us to perform a lead-specific relevance analysis for each arrhythmia category, guiding clinical experts to focus on the most influential leads identified. For example, the V2 lead is highlighted as the most significant feature in six out of nine arrhythmia categories, with avR, V1, V3, and V4 also being important. This information is valuable for researchers focusing on specific arrhythmia predictions, as it suggests prioritizing these leads in ECG monitoring. Interestingly, this contrasts with the common use of ECG lead II in arrhythmia detection, which our model found to be less relevant, highlighting a divergence from established practices noted in related literature.

The significant advantage of our explainable model lies in the synergy created by merging global and individual explainability approaches. Clinicians can use features deemed important globally to scrutinize individual arrhythmia predictions. This process allows them to verify if the ECG segments identified by SHAP as influential align with established clinical knowledge. Essentially, the model facilitates a comprehensive cross-verification, enabling clinicians to confirm the clinical validity of globally relevant features by examining their contribution to specific arrhythmia classifications on a case-by-case basis.

This study’s primary limitation is its reliance on a single database for constructing the prediction model, potentially affecting both the model’s generalizability and the validity of explainability analysis. Future research should include additional databases like MIT-BIH to validate and benchmark our findings. Additionally, our study focused exclusively on ECG signal characteristics, excluding other potentially informative patient data such as demographic details, medical history, and laboratory results, due to database constraints. Expanding future models to datasets with comprehensive patient information is recommended. Finally, our study lacks clinical validation of the explainability results, a common challenge in CVD prediction models using XAI. Future efforts should aim to correlate XAI findings with established clinical knowledge to enhance their medical relevance.

This study developed a deep learning (DL) based prediction model for nine arrhythmia classes using 12-lead ECG signals and tested various DL architectures including ResNet, DenseNet, and VGG16. ResNet34 was identified as the most effective model, achieving an AUROC of 0.98 and an F1-score of 0.826. A hybrid approach combining raw ECG signals with Heart Rate Variability (HRV) features was also tested, but it did not outperform the raw signal model. Additionally, we conducted an explainability analysis using SHAP within the XAI framework. This analysis pinpointed key ECG leads and signal segments critical for arrhythmia prediction, enhancing the model’s transparency and potential usability for clinical professionals. The study highlights the significance of explainable models in cardiovascular disease (CVD) prediction, promoting their acceptance in clinical settings due to the clarity of model decisions.