Revolutionizing Pediatric Cardiac Care: A Comprehensive Review of Artificial Intelligence-Enhanced Electrocardiograms

Many thanks to our sponsor Esdebe who helped us prepare this research report.

Abstract

The landscape of pediatric cardiology is undergoing a significant transformation with the advent of Artificial Intelligence-Enhanced Electrocardiograms (AI-ECGs). These innovative diagnostic tools offer a non-invasive, cost-effective, and highly accessible modality for the early detection and characterization of a broad spectrum of cardiac conditions in children, most notably left ventricular systolic dysfunction (LVSD). This comprehensive review systematically dissects the integration of AI-ECGs within pediatric cardiac care, moving beyond mere descriptive accounts to delve into the intricate architecture of the underlying deep learning algorithms, sophisticated training methodologies employed, and robust performance metrics utilized for validation. Furthermore, it critically examines the multifaceted regulatory landscape governing these nascent technologies, alongside the complex practical and profound ethical considerations that must be meticulously navigated during their implementation across diverse and often resource-constrained clinical settings. By providing an in-depth analysis, this report aims to illuminate the profound potential of AI-ECGs to redefine diagnostic paradigms and improve clinical outcomes in pediatric cardiology, while also acknowledging the ongoing challenges that necessitate sustained research and interdisciplinary collaboration.

Many thanks to our sponsor Esdebe who helped us prepare this research report.

1. Introduction

The early and accurate diagnosis of cardiac pathologies in the pediatric population is of paramount importance, directly impacting long-term prognosis, mitigating disease progression, and reducing both morbidity and mortality. Congenital heart defects (CHDs) affect approximately 1% of live births, making them the most common birth defect, while acquired cardiac conditions, such as myocarditis, cardiomyopathies, and arrhythmias, also pose significant health burdens [1, 2]. Traditional diagnostic modalities, while highly effective and serving as clinical gold standards, frequently present inherent limitations. Echocardiography, for instance, offers detailed structural and functional insights but is resource-intensive, requiring specialized equipment, highly trained sonographers, and experienced pediatric cardiologists for interpretation, rendering it less accessible in primary care settings or underserved regions [3]. Cardiac Magnetic Resonance Imaging (cMRI) provides even more comprehensive anatomical and functional assessment but is typically reserved for complex cases due due to its high cost, prolonged scan times, and the frequent need for sedation in pediatric patients [4].

Electrocardiography (ECG) stands in stark contrast to these advanced imaging techniques. It is a venerable, non-invasive, readily available, and inexpensive diagnostic tool that captures the electrical activity of the heart [5]. However, the conventional interpretation of pediatric ECGs is notoriously challenging due to significant age-related variations in cardiac physiology, electrical axis, and waveform morphology, demanding substantial expertise and experience [6]. The recent exponential advancements in artificial intelligence (AI), particularly in the domain of deep learning, have heralded a new era for ECG analysis. By leveraging complex algorithms capable of learning intricate patterns from vast datasets, AI has begun to enhance the diagnostic capabilities of traditional ECGs, leading to the development of AI-ECGs. These sophisticated tools hold immense promise in detecting subtle electrical signatures indicative of underlying cardiac conditions, such as left ventricular systolic dysfunction (LVSD), hypertrophic cardiomyopathy, congenital long QT syndrome, and even specific genetic mutations, thereby potentially revolutionizing pediatric cardiac care by facilitating earlier diagnosis and intervention [7, 8]. The integration of AI-ECGs thus represents a paradigm shift, offering a pathway to democratize advanced cardiac screening and diagnosis, particularly in contexts where specialized pediatric cardiology resources are scarce.

Many thanks to our sponsor Esdebe who helped us prepare this research report.

2. Deep Learning Algorithms and Architecture

The power of AI-ECGs stems from their reliance on sophisticated deep learning algorithms, which are a subset of machine learning models characterized by multiple layers of processing units, enabling them to learn hierarchical representations of data. For ECG analysis, these algorithms are trained to identify subtle patterns in electrical signals that are often imperceptible to the human eye, even for experienced cardiologists. The choice of algorithmic architecture is critical, as it dictates how the model processes sequential ECG data and extracts relevant features.

2.1 Convolutional Neural Networks (CNNs)

Convolutional Neural Networks (CNNs) have emerged as the workhorse for many AI-ECG applications due to their exceptional proficiency in processing data with a grid-like topology, such as images or, in this context, one-dimensional time-series ECG signals [9]. The core innovation of CNNs lies in their convolutional layers, which apply a series of learnable filters (kernels) across the input data. Each filter is designed to detect specific local features, such as peaks, troughs, intervals, or changes in slope, within the ECG waveform [10].

2.1.1 Architectural Components

Typically, a CNN architecture for ECG analysis comprises several key layers:

• Input Layer: Receives the raw ECG signal, often preprocessed (e.g., filtered, normalized, resampled) to remove noise and standardize the input. For 12-lead ECGs, this might involve a multi-channel input.
• Convolutional Layers: These layers consist of multiple convolutional filters that slide across the ECG signal, performing dot products and generating feature maps. Each filter specializes in detecting different temporal patterns or features. The output of a convolutional layer is a set of feature maps that highlight the presence of these learned features across the ECG signal.
• Activation Functions: Non-linear activation functions, such as ReLU (Rectified Linear Unit), are applied after each convolutional layer to introduce non-linearity, enabling the network to learn more complex relationships within the data [11].
• Pooling Layers: These layers, typically Max Pooling or Average Pooling, reduce the spatial dimensions of the feature maps, thereby decreasing computational complexity, controlling overfitting, and making the learned features more robust to minor shifts or distortions in the ECG signal [12].
• Flattening Layer: After several convolutional and pooling layers, the resulting high-dimensional feature maps are flattened into a one-dimensional vector.
• Fully Connected Layers (Dense Layers): These layers receive the flattened feature vector and perform classification. Each neuron in a fully connected layer is connected to all neurons in the previous layer. The final layer typically uses a softmax activation function for multi-class classification or a sigmoid for binary classification (e.g., presence or absence of LVSD).

2.1.2 Application in Pediatric ECGs

In the context of pediatric ECGs, CNNs are particularly adept at capturing subtle temporal and spatial features of cardiac electrical activity that might indicate ventricular dysfunction. For instance, a seminal study by Mayourian et al. (2024) successfully employed a CNN to analyze ECGs from pediatric patients to detect both right and left ventricular dysfunction. Their model demonstrated high accuracy in identifying severe LVSD cases, highlighting the CNN’s ability to discern complex patterns related to altered ventricular mechanics from surface ECGs [7]. The efficacy of CNNs stems from their inherent ability to automatically learn hierarchical features, from basic wave morphologies (P, QRS, T waves) to more complex spatio-temporal dynamics across multiple leads, without requiring manual feature engineering.

2.2 Recurrent Neural Networks (RNNs)

Recurrent Neural Networks (RNNs) are specifically designed to process sequential data by maintaining an internal memory that allows them to learn temporal dependencies. This characteristic makes them inherently suitable for time-series analysis, which is fundamental to ECG signals where the order and timing of events are crucial for diagnosis [13].

2.2.1 Long Short-Term Memory (LSTM) Networks

A prominent type of RNN, Long Short-Term Memory (LSTM) networks, address the vanishing and exploding gradient problems commonly encountered in traditional RNNs when processing long sequences. LSTMs achieve this through a sophisticated internal structure consisting of memory cells and several ‘gates’ (input gate, forget gate, output gate) that control the flow of information into and out of the cell [14].

• Forget Gate: Decides what information to discard from the cell state.
• Input Gate: Decides what new information to store in the cell state.
• Output Gate: Decides what part of the cell state to output.

These gates allow LSTMs to selectively remember or forget information over long periods, making them highly effective at modeling long-range temporal dependencies within the ECG signal. For instance, LSTMs can learn to connect an abnormality observed early in the cardiac cycle (e.g., a specific P-wave morphology) with a later event (e.g., T-wave inversion) that together might indicate a particular condition like atrial pathology or ischemia.

2.2.2 Application in Pediatric ECGs

While RNNs, particularly LSTMs, have been successfully applied to predict cardiac events and classify arrhythmias in adult populations by learning intricate temporal dependencies in ECG data, their application in pediatric AI-ECGs has been relatively more limited compared to CNNs [15]. This is partly due to the challenges of obtaining sufficiently large, well-annotated pediatric datasets suitable for training complex temporal models and the inherent variability in pediatric ECGs across different age groups. However, their potential for discerning subtle, time-dependent variations indicative of conditions like congenital long QT syndrome or other channelopathies makes them a promising area for further exploration, particularly when focusing on specific rhythm disturbances or conduction abnormalities.

2.3 Hybrid Models

Recognizing the complementary strengths of CNNs in spatial feature extraction and RNNs in temporal dependency modeling, researchers have developed hybrid models that integrate elements of both architectures. These models aim to leverage the best of both worlds, leading to enhanced performance in complex ECG classification tasks [16].

2.3.1 CNN-LSTM Architecture

A common hybrid architecture involves stacking CNN layers before RNN layers (often LSTMs). In this configuration:

• CNN Layers: Initial CNN layers are responsible for automatically extracting robust, high-level features from local segments of the ECG signal. For example, a CNN might identify characteristic shapes of QRS complexes or ST segments.
• RNN (LSTM) Layers: The feature sequences generated by the CNN are then fed into RNN layers. These RNNs process the sequence of extracted features, learning the temporal relationships and dependencies between them. This allows the model to understand how individual features evolve over time or interact across different segments of the ECG.

For example, a hybrid CNN-LSTM model was reported to predict LVSD in pediatric patients, showing enhanced accuracy compared to models relying solely on CNNs or LSTMs [17]. This improvement can be attributed to the CNN’s ability to efficiently extract local, lead-specific features, which are then integrated by the LSTM to understand the global temporal context and inter-lead relationships, which are critical for comprehensive cardiac assessment. Other hybrid approaches might involve attention mechanisms (e.g., Transformer networks) that allow the model to focus on the most relevant parts of the ECG signal for a given diagnostic task, further improving interpretability and accuracy [18]. These models are particularly powerful when dealing with the nuanced and dynamic nature of pediatric ECGs, where both morphological details and their temporal sequencing are diagnostically significant.

Many thanks to our sponsor Esdebe who helped us prepare this research report.

3. Training Methodology and Data Considerations

The robustness, accuracy, and generalizability of AI-ECGs are inextricably linked to the meticulous design of their training methodology and the careful consideration of the underlying data. The success of deep learning models hinges on exposure to vast, high-quality, and diverse datasets during the learning phase.

3.1 Data Acquisition

The performance ceiling of any AI model is fundamentally constrained by the quality and quantity of its training data. For AI-ECGs, this translates to the need for large, clinically relevant, and accurately annotated ECG datasets. However, in the specialized domain of pediatric cardiology, the acquisition of such datasets presents significant challenges:

• Rarity of Conditions: While common in aggregate, many specific pediatric cardiac conditions, particularly rare cardiomyopathies or complex congenital heart diseases, have a lower incidence rate compared to adult cardiac diseases. This leads to fewer available ECGs for these specific conditions.
• Ethical and Practical Barriers: Obtaining pediatric patient data involves stricter ethical guidelines regarding consent (from parents/guardians) and child assent. The practicalities of long-term data collection in children, often involving multiple visits and different healthcare providers, also complicate dataset creation.
• Heterogeneity of Pediatric ECGs: Pediatric ECGs exhibit pronounced age-dependent variations. Normal ECG parameters, QRS axis, and T-wave morphology change significantly from neonates through infancy, childhood, and adolescence [6]. A robust AI model must account for this inherent biological variability, requiring datasets that comprehensively represent all relevant age groups.
• Annotation Expertise: Accurate annotation of pediatric ECGs with corresponding echocardiographic or cMRI findings (the ground truth for conditions like LVSD) requires highly specialized pediatric cardiologists. This expert labeling is time-consuming and expensive, limiting the scale of annotated datasets.

To mitigate these data scarcity issues, researchers have increasingly relied on advanced data augmentation techniques. These methods artificially expand the size of existing datasets by generating new, plausible samples from the original ones without necessarily collecting new raw data. Common techniques include:

• Synthetic Data Generation: Advanced generative models, such as Generative Adversarial Networks (GANs), can be trained to produce synthetic ECG signals that mimic the statistical properties of real pediatric ECGs. While promising, ensuring the clinical fidelity and physiological realism of synthetic data remains a challenge [19].
• Time Warping and Scaling: Distorting the ECG signal in the time domain (e.g., slightly stretching or compressing segments) or scaling its amplitude can create new variations that simulate physiological differences or measurement variability.
• Noise Injection: Adding different types of realistic noise (e.g., baseline wander, muscle artifact, power line interference) can make the model more robust to noisy clinical recordings.
• Amplitude and Baseline Shifts: Randomly shifting the amplitude or baseline of the ECG signal can help the model generalize to variations in lead placement or patient physiology.
• Lead Swapping/Permutation: For 12-lead ECGs, selectively swapping or permuting leads can help the model learn lead-independent features, although care must be taken to maintain anatomical consistency [17].

The judicious application of these techniques helps enrich training datasets, allowing models to learn more generalizable features and improve their ability to detect conditions even with limited initial real-world data.

3.2 Transfer Learning

Transfer learning is a powerful machine learning paradigm particularly valuable in domains with limited task-specific data, such as pediatric cardiology [20]. It involves leveraging knowledge gained from training a model on a large source dataset (often from a related but different domain) and then fine-tuning this pre-trained model on a smaller, target-specific dataset.

3.2.1 Principles of Transfer Learning

The underlying principle is that features learned from a large, general dataset are often transferable to more specific tasks. For ECG analysis:

• Pre-training: A deep learning model (e.g., a CNN) is initially trained on a massive dataset of adult ECGs for a broad classification task (e.g., detecting various adult arrhythmias or structural heart diseases). During this phase, the lower layers of the neural network learn fundamental features common to all ECG signals, such as basic wave morphologies, interval durations, and segment characteristics.
• Feature Extraction or Fine-tuning: Once pre-trained, the model can be adapted for the pediatric task. In a feature extraction approach, the pre-trained layers are ‘frozen’ (their weights are not updated), and only new classification layers are added and trained on the pediatric dataset. In a fine-tuning approach, the entire pre-trained model (or at least the higher layers) is re-trained with a smaller learning rate on the pediatric dataset. This allows the model to adapt its learned features specifically to the nuances of pediatric ECGs while retaining the foundational knowledge acquired from the adult data [21].

3.2.2 Benefits and Application in Pediatric AI-ECGs

Transfer learning offers several significant benefits for pediatric ECG analysis:

• Overcoming Data Limitations: It directly addresses the scarcity of large, annotated pediatric datasets by allowing models to leverage learned features from more abundant adult ECG datasets. This drastically reduces the amount of pediatric data required for effective training.
• Improved Generalization: Models pre-trained on diverse adult populations are often more robust and less prone to overfitting when fine-tuned on smaller pediatric datasets.
• Faster Convergence: Training a pre-trained model typically requires fewer epochs and converges faster than training a model from scratch, saving computational resources and time.

Yang et al. (2025) demonstrate the effectiveness of this approach, where models pre-trained on adult ECG data were adapted to predict LVSD in pediatric congenital heart disease, showing robust performance despite the inherent differences between adult and pediatric cardiac physiology and ECG characteristics [17]. This highlights the potential of domain adaptation strategies, where a model trained on one domain (adult ECGs) can be effectively adapted to another related domain (pediatric ECGs) with tailored adjustments.

3.3 Model Validation

Rigorous and comprehensive model validation is an indispensable step to ensure that AI-ECGs are not only accurate on the training data but also generalize reliably to unseen clinical data and perform consistently across diverse patient populations. Without robust validation, models risk being overfitted, biased, or clinically unreliable.

3.3.1 Internal Validation

Internal validation assesses a model’s performance within the dataset from which it was developed. Common techniques include:

• K-fold Cross-validation: The dataset is partitioned into ‘k’ equally sized folds. The model is trained ‘k’ times, each time using ‘k-1’ folds for training and the remaining fold for validation. The results are then averaged across all ‘k’ iterations. This method provides a more stable and less biased estimate of model performance than a single train-test split, particularly for smaller datasets [22].
• Stratified Sampling: When splitting data for training and validation (or during cross-validation), it is crucial to use stratified sampling to ensure that the distribution of key characteristics (e.g., age groups, gender, disease prevalence) is maintained across all subsets. This prevents the model from being exposed to an imbalanced representation of conditions during training or testing.
• Hold-out Sets: While k-fold provides good estimates, a final hold-out test set, completely separate from the data used for training and hyperparameter tuning, is essential for an unbiased evaluation of the final model’s performance.

3.3.2 External Validation

External validation is arguably the most critical step in establishing the clinical utility and generalizability of an AI-ECG model. It involves evaluating the model’s performance on entirely independent datasets collected from different institutions, geographical locations, patient populations, and often using different ECG machines or acquisition protocols. This assesses the model’s robustness and its applicability across diverse clinical settings, addressing potential issues like data shift or domain drift that can occur when models are deployed in new environments [23]. Challenges in external validation include:

• Data Heterogeneity: Variations in patient demographics, disease prevalence, concomitant medications, and recording standards across different sites can significantly impact model performance.
• Logistical Complexity: Multi-center studies for external validation are resource-intensive, requiring extensive collaboration, data sharing agreements, and standardized data collection protocols.

Successfully passing external validation provides strong evidence for a model’s clinical readiness. For example, a model trained at a tertiary academic center must demonstrate equivalent performance when tested in a community hospital or a primary care clinic with a different patient demographic and prevalence of cardiac conditions. The inclusion of diverse populations, including various ethnic groups and socioeconomic backgrounds, is also vital to detect and mitigate potential algorithmic biases.

Many thanks to our sponsor Esdebe who helped us prepare this research report.

4. Performance Metrics and Efficacy

The evaluation of AI-ECGs extends beyond simple accuracy to encompass a suite of performance metrics that reflect their diagnostic utility, reliability, and clinical relevance. These metrics provide a nuanced understanding of how well an AI model can identify cardiac conditions in pediatric patients, considering the inherent variability and diagnostic challenges.

4.1 Diagnostic Accuracy

Diagnostic accuracy is a foundational metric, but it comprises several components that offer different insights into a model’s capability. For classification tasks (e.g., detecting LVSD), common metrics derived from a confusion matrix (True Positives, True Negatives, False Positives, False Negatives) include:

• Sensitivity (Recall): The proportion of actual positive cases (e.g., children with LVSD) that are correctly identified by the AI-ECG. High sensitivity is crucial for screening tools to minimize false negatives and ensure that affected individuals are not missed.
• Specificity: The proportion of actual negative cases (e.g., children without LVSD) that are correctly identified as negative. High specificity helps reduce unnecessary follow-up tests and anxiety for healthy children.
• Positive Predictive Value (PPV or Precision): The proportion of positive predictions made by the AI-ECG that are actually correct. A high PPV indicates that a positive test result is highly reliable.
• Negative Predictive Value (NPV): The proportion of negative predictions made by the AI-ECG that are actually correct. A high NPV implies that a negative test result reliably rules out the condition.
• F1-Score: The harmonic mean of precision and recall, providing a balanced measure of the model’s accuracy, especially useful when dealing with imbalanced datasets (where one class is much more prevalent than the other).
• Accuracy: The overall proportion of correct predictions (both true positives and true negatives) among the total number of cases. While intuitive, it can be misleading in imbalanced datasets.
• Area Under the Receiver Operating Characteristic Curve (AUC-ROC): This metric plots the true positive rate (sensitivity) against the false positive rate (1 – specificity) at various threshold settings. An AUC value closer to 1.0 indicates excellent discriminatory power, meaning the model can effectively distinguish between positive and negative cases across a range of operational thresholds. It is a robust measure for evaluating binary classifiers, particularly in imbalanced datasets [24].
• Area Under the Precision-Recall Curve (AUC-PRC): Plots precision against recall at various thresholds. This curve is often more informative than AUC-ROC for highly imbalanced datasets, where the number of negative instances greatly outweighs the number of positive instances, as it focuses on the performance on the positive class [25].

AI-ECGs have shown remarkable diagnostic accuracy in pediatric applications. For example, a significant study by Niaz et al. (2024) reported an impressive AUC of 0.93 for detecting severe LVSD (defined as LVEF ≤ 35%) and 0.88 for moderate LVSD (LVEF < 50%) within a large cohort of 10,142 pediatric patients [8]. These figures suggest that AI-ECGs possess excellent discriminatory capabilities, making them highly valuable as screening or diagnostic aids.

4.2 Generalization Across Demographics

The successful deployment of AI-ECGs in clinical practice hinges on their ability to generalize effectively across the diverse spectrum of the pediatric population. Children present a dynamic physiological landscape, with significant developmental changes impacting cardiac electrical activity. Research efforts are therefore focused on assessing the efficacy of AI-ECGs across various pediatric demographics, including:

• Age Groups: From neonates and infants (0-1 year) to toddlers (1-3 years), preschoolers (3-5 years), school-aged children (5-12 years), and adolescents (12-18 years). ECG characteristics, such as heart rate, QRS axis, and T-wave morphology, undergo substantial physiological changes with age. An AI model must be robust enough to handle these normal developmental variations while still accurately identifying pathology [6].
• Sex: While less pronounced than age-related differences, subtle sex-based variations in ECG parameters exist even in pediatric populations. Models should perform consistently for both sexes without inherent bias.
• Racial and Ethnic Backgrounds: Biological variations, genetic predispositions, and differences in environmental factors across various racial and ethnic groups can influence cardiac parameters. Ensuring equitable performance across all groups is a critical ethical and clinical imperative to avoid exacerbating health disparities [26].
• Underlying Conditions: Pediatric cardiac conditions are incredibly diverse, ranging from simple shunts to complex single-ventricle physiology. An AI-ECG designed for LVSD detection might perform differently in a child with a complex congenital anomaly versus one with acquired cardiomyopathy. Comprehensive validation across various underlying pathologies is necessary.

While existing research indicates that AI-ECGs can accurately detect LVSD in children of different ages and sexes, suggesting a degree of inherent robustness and generalizability, further granular studies are imperative. Specifically, performance in highly specific subgroups, such as neonates and infants (where ECG interpretation is particularly challenging and the stakes are often higher), children with genetic syndromes, or those with very rare conditions, requires dedicated investigation and validation. The goal is to ensure that the AI-ECG does not exhibit ‘cold spots’ of poor performance for any particular subgroup, which could lead to missed diagnoses or inappropriate referrals.

4.3 Comparison with Traditional Methods

The true value proposition of AI-ECGs becomes apparent when compared against existing gold standard diagnostic modalities and conventional ECG interpretation. While not intended to completely replace traditional methods, AI-ECGs offer distinct advantages:

• Echocardiography: While highly detailed and indispensable for structural diagnosis and precise functional quantification, echocardiography is expensive, requires specialized equipment and expertise, and is often time-consuming. AI-ECGs, in contrast, are non-invasive, widely accessible, and offer immediate results, making them ideal for initial screening or triaging in resource-limited settings or primary care. They can identify patients who truly need a follow-up echocardiogram, thus optimizing resource utilization and reducing unnecessary referrals.
• Conventional ECG Interpretation by Pediatric Cardiologists: Human expert interpretation, while nuanced, is subject to inter-observer variability, fatigue, and the inherent difficulty of pediatric ECGs. AI-ECGs provide an objective, standardized analysis that can detect patterns too subtle or complex for human perception. This is particularly true for conditions like specific genetic cardiomyopathies or subtle forms of ventricular dysfunction where diagnostic criteria can be challenging to apply consistently across all ages [7, 8]. The AI acts as a sophisticated ‘second reader’ or an ‘enhanced initial screen’.
• Cost-Effectiveness: The infrastructure for standard ECG acquisition is already ubiquitous in most healthcare systems. Integrating AI analysis primarily involves software, making it a highly cost-effective addition to existing diagnostic pathways, especially when considering the potential to prevent costly hospitalizations or advanced imaging due to delayed diagnosis.
• Speed and Automation: AI-ECGs provide near-instantaneous analysis, automating a traditionally manual and often time-consuming process. This speed can be critical in acute care settings or high-volume clinics.

In essence, AI-ECGs do not seek to supplant the deep clinical judgment of a pediatric cardiologist or the detailed imaging capabilities of an echocardiogram. Instead, they serve as a powerful complementary tool, offering a rapid, non-invasive, and cost-effective alternative that can provide comparable or even superior performance for specific diagnostic tasks, particularly in screening for conditions like LVSD. This makes them exceptionally valuable for broadening access to cardiac screening, improving diagnostic efficiency, and enabling earlier intervention, especially in regions lacking specialized pediatric cardiology infrastructure.

Many thanks to our sponsor Esdebe who helped us prepare this research report.

5. Regulatory Landscape and Ethical Considerations

The integration of AI-ECGs into routine clinical practice, particularly in the vulnerable pediatric population, necessitates navigating a complex and evolving regulatory landscape coupled with profound ethical considerations. These frameworks are crucial for ensuring patient safety, diagnostic reliability, and equitable access to these transformative technologies.

5.1 Regulatory Approval

The development and deployment of AI-ECGs as medical devices require stringent regulatory oversight to guarantee their safety, efficacy, and clinical reliability. Different jurisdictions have established various pathways for the approval of AI-based diagnostic tools.

5.1.1 United States: Food and Drug Administration (FDA)

In the United States, the FDA is the primary regulatory body overseeing medical devices, including Artificial Intelligence/Machine Learning (AI/ML)-based Software as a Medical Device (SaMD). The FDA has established several regulatory pathways relevant to AI-ECGs:

• De Novo Classification Request: This pathway is for novel devices that have no predicate device and are classified as low-to-moderate risk. If an AI-ECG is deemed a novel technology without a substantially equivalent device already on the market, it would likely follow this pathway, requiring robust clinical evidence of safety and effectiveness [27].
• 510(k) Premarket Notification: This is the most common pathway for medical devices and applies when a new device is ‘substantially equivalent’ to a legally marketed predicate device. For AI-ECGs that perform similar functions to existing ECG interpretation software but with AI enhancements, this pathway might be applicable, requiring a demonstration of substantial equivalence in performance and safety [28].
• Premarket Approval (PMA): This is the most stringent pathway, typically reserved for high-risk devices that sustain or support human life, are of substantial importance in preventing impairment of human health, or present a potential unreasonable risk of illness or injury. While less likely for a diagnostic AI-ECG, complex AI systems with significant clinical impact might be considered under PMA.

Crucially, the FDA has also recognized the unique challenges of AI/ML-based SaMD, particularly their adaptive capabilities (learning from real-world data post-market). In 2019, the FDA issued a discussion paper outlining a proposed regulatory framework for ‘Good Machine Learning Practice’ (GMLP) and a ‘Total Product Lifecycle’ (TPLC) approach, emphasizing pre-defined performance metrics, data management practices, and plans for managing modifications to ensure safety and effectiveness of continuously learning algorithms [29]. This means AI-ECGs must not only demonstrate initial safety and efficacy but also have a clear strategy for post-market surveillance, validation of updates, and transparency in their operational changes.

5.1.3 Global Regulatory Landscape

Similar regulatory bodies exist globally, such as the European Medicines Agency (EMA) in the European Union, the Medicines and Healthcare products Regulatory Agency (MHRA) in the UK, and Health Canada. These bodies are also grappling with appropriate regulatory frameworks for AI/ML in healthcare, generally emphasizing comprehensive clinical validation, risk management, data governance, and post-market vigilance [30]. Harmonization of these international standards is an ongoing effort to facilitate global adoption and innovation while ensuring consistent patient protection.

5.2 Ethical Considerations

The deployment of AI-ECGs, especially in the sensitive context of pediatric care, gives rise to profound ethical questions that extend beyond mere regulatory compliance. These considerations are vital to fostering trust, ensuring equitable access, and safeguarding patient well-being.

5.2.1 Data Privacy and Security

The handling of sensitive pediatric health data demands the most stringent measures to protect patient confidentiality. Pediatric health information, by its nature, often carries lifelong implications and requires special protections. Key considerations include:

• Informed Consent and Assent: Obtaining truly informed consent from parents/guardians for data collection and use in AI model training is complex. For older children, their assent (agreement to participate) is also crucial. This needs to clearly outline how data will be used, anonymized, and protected.
• Anonymization and De-identification: While aiming for complete anonymization, the sheer volume and granularity of medical data can sometimes lead to re-identification risks. Robust de-identification techniques, including pseudonymization and aggregation, are critical.
• Regulatory Compliance: Adherence to data protection regulations like HIPAA (Health Insurance Portability and Accountability Act) in the US and GDPR (General Data Protection Regulation) in the EU is mandatory. These regulations impose strict requirements on how personal health information (PHI) is collected, stored, processed, and shared [31].
• Cybersecurity: Implementing state-of-the-art data encryption, secure data storage (e.g., in cloud environments with robust security protocols), access controls (role-based access), and regular security audits are vital to prevent unauthorized access, data breaches, and cyber-attacks. The vulnerability of digital health records necessitates a proactive and adaptive cybersecurity posture.

5.2.2 Algorithmic Bias and Fairness

AI models are trained on historical data, and if this data contains systemic biases or is unrepresentative of the target population, the AI model can inadvertently learn and perpetuate these biases, leading to disparities in diagnostic accuracy and healthcare outcomes [32]. In pediatric cardiology, this is a particularly acute concern:

• Sources of Bias: Bias can originate from various sources:
  • Sampling Bias: If training datasets predominantly consist of data from specific demographic groups (e.g., urban populations, certain ethnic groups, or only patients from tertiary care centers), the model may perform poorly on underrepresented groups.
  • Labeling Bias: Inaccurate or inconsistent diagnostic labels (ground truth) applied by human annotators can introduce bias. For instance, if certain conditions are more frequently misdiagnosed in specific patient groups, the AI might learn this faulty association.
  • Measurement Bias: Differences in ECG acquisition protocols or equipment calibration across various clinical sites can introduce subtle biases.
• Consequences: Algorithmic bias can lead to:
  • Underdiagnosis/Overdiagnosis: Disparities in diagnostic accuracy for different racial, ethnic, or socioeconomic groups, leading to missed diagnoses for some and unnecessary interventions for others.
  • Exacerbation of Health Disparities: If AI-ECGs perform less well for vulnerable populations, they could inadvertently widen existing health inequities.
• Mitigation Strategies: Continuous monitoring, auditing, and updating of AI models are essential. This involves:
  • Diverse Data Collection: Actively seeking and incorporating data from a wide range of pediatric populations, including various age groups, sexes, ethnicities, and socioeconomic backgrounds.
  • Fairness Metrics: Employing specific fairness metrics (e.g., demographic parity, equalized odds) during model development and evaluation to ensure equitable performance across different subgroups [33].
  • Bias Detection and Debiasing Techniques: Using techniques like adversarial debiasing or re-weighting training data to reduce the impact of biased samples.
  • Transparency and Auditing: Making model outputs and decision-making processes transparent to allow for human oversight and auditing, particularly for high-stakes decisions.

5.2.3 Algorithmic Transparency and Explainability (XAI)

Many deep learning models operate as ‘black boxes,’ providing a diagnosis without clearly articulating the reasoning behind it. This lack of transparency poses a significant ethical challenge, particularly in healthcare where trust and accountability are paramount.

• The Black Box Problem: Clinicians need to understand why an AI-ECG made a particular diagnosis to effectively integrate it into their decision-making process, confirm its validity, and maintain professional responsibility. A ‘yes/no’ answer without justification is insufficient for clinical adoption [34].
• Interpretability and Explainability: Ensuring that AI models are interpretable and that their decisions can be understood by human experts is essential. This field, known as Explainable AI (XAI), seeks to develop methods that can shed light on the internal workings of AI models. Techniques include:
  • Saliency Maps: Visualizing which parts of the ECG signal (e.g., specific leads, time points, or waveform features) most strongly influenced the model’s prediction [35].
  • LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations): Model-agnostic methods that explain individual predictions by approximating the complex model with a simpler, interpretable model locally around the prediction [36, 37].
  • Attention Mechanisms: Architectures incorporating attention mechanisms allow the model to explicitly highlight features or segments of the ECG that are most relevant for a given task.
• Accountability: In cases of diagnostic error, understanding the AI’s reasoning is crucial for determining accountability. Clear guidelines on responsibility (e.g., between the AI developer, healthcare provider, and institution) are needed.

5.2.4 Equitable Access and Digital Divide

While AI-ECGs promise to democratize access to advanced cardiac screening, their deployment could inadvertently exacerbate existing health disparities if not carefully managed. The ‘digital divide’ – disparities in access to technology, internet connectivity, and digital literacy – must be considered. Strategies to ensure equitable access include developing lightweight models that can run on low-cost devices, providing necessary infrastructure in underserved areas, and ensuring that training and support are available to healthcare professionals in all settings.

In summary, the journey from AI-ECG development to widespread clinical adoption requires a holistic approach that not only focuses on technical performance but also rigorously addresses the intricate regulatory requirements and the profound ethical responsibilities inherent in deploying powerful AI systems in pediatric healthcare.

Many thanks to our sponsor Esdebe who helped us prepare this research report.

6. Practical Challenges and Implementation Roadblocks

Beyond the theoretical and ethical considerations, the successful integration of AI-ECGs into the complex, real-world clinical environment of pediatric cardiology faces several practical hurdles. Addressing these challenges is paramount for realizing the full potential of these transformative technologies.

6.1 Data Privacy and Security in Practice

While ethical principles mandate data privacy, the practical implementation in a dynamic healthcare ecosystem presents ongoing challenges:

• Interoperability and Data Exchange: Healthcare systems are often fragmented, with different Electronic Health Record (EHR) systems that may not seamlessly communicate. Securely integrating AI-ECG platforms with various EHRs while maintaining data privacy and integrity is a significant technical challenge. Data transfer protocols must be robust and compliant with strict regulations [38].
• Cloud vs. On-Premise Deployment: Deciding whether to deploy AI models and store sensitive data in cloud environments or on-premise servers involves trade-offs between scalability, cost, security, and regulatory compliance. Cloud solutions offer flexibility but raise concerns about data residency and third-party access, while on-premise solutions demand significant IT infrastructure investment and expertise.
• De-identification Challenges: While techniques exist, ensuring complete and irreversible de-identification of large, complex medical datasets, especially those with longitudinal data, remains a non-trivial task. The risk of re-identification, however small, necessitates continuous vigilance and advanced cryptographic techniques.
• Insider Threats: Despite external cybersecurity measures, insider threats (e.g., unauthorized access by staff) remain a concern. Strict access controls, regular audits, and staff training on data privacy protocols are essential.

6.2 Algorithmic Bias in Real-World Data

Detecting and mitigating algorithmic bias is more complex in real-world clinical data than in controlled research settings:

• Dynamic Nature of Bias: Bias is not static; it can emerge or change as the model interacts with new, real-world data streams. Continuous monitoring for performance disparities across different demographic groups is essential, which requires robust data collection on patient demographics during deployment.
• Proxy Bias: Datasets might not directly contain sensitive attributes (e.g., race, socioeconomic status) due to privacy concerns, but the AI model might still infer these attributes from other available data (e.g., zip code, insurance status) and perpetuate proxy biases [39]. Identifying and mitigating such indirect biases is a significant challenge.
• Feedback Loops: If an AI system is used for screening and affects who receives further diagnostic work-up, it can create feedback loops where diagnostic data becomes biased towards the population that the AI already ‘favors’ for referral, further entrenching the bias.
• Defining ‘Fairness’: There are multiple mathematical definitions of fairness (e.g., equal accuracy, equalized odds, demographic parity), and optimizing for one may come at the expense of another. Clinicians, ethicists, and AI developers must collaborate to define the most appropriate fairness metrics for specific clinical applications and target populations.

6.3 Integration into Clinical Workflow

The seamless integration of AI-ECGs into existing clinical workflows is crucial for user adoption and maximizing their impact. This involves addressing several logistical and human factors:

• Interoperability with EHRs and PACS: AI-generated reports and interpretations must be seamlessly integrated into Electronic Health Records (EHRs) and Picture Archiving and Communication Systems (PACS) to be readily accessible to clinicians. This requires standardized data formats (e.g., DICOM, HL7 FHIR) and robust application programming interfaces (APIs) [40].
• Alert Fatigue: If the AI system generates too many false positives or unhelpful alerts, clinicians may become desensitized or even ignore its recommendations, leading to ‘alert fatigue’ and undermining trust. The alert system must be intelligently designed, prioritized, and clinically actionable.
• Training and Education: Healthcare professionals, including pediatric cardiologists, general pediatricians, and primary care physicians, need comprehensive training on how to interpret AI-generated results, understand the model’s limitations, and integrate AI insights into their clinical decision-making. This involves not only technical training but also fostering a mindset of collaborative intelligence between human and AI.
• Trust and Acceptance: Clinician skepticism or over-reliance on AI can both be detrimental. Building trust requires demonstrating consistent accuracy, transparency (explainability), and clear communication regarding the AI’s capabilities and limitations. Over-reliance, where clinicians might blindly follow AI recommendations without critical appraisal, could lead to diagnostic errors when the AI is wrong [41].
• Workflow Optimization: The AI-ECG solution must fit naturally into existing clinical pathways without adding significant burden or disruption. For example, if it requires additional data entry or steps, adoption will be low. Ideally, it should automate tasks, reduce turnaround times, and free up clinician time.

6.4 Resource Constraints in Low-Resource Settings

While AI-ECGs promise increased accessibility, significant resource constraints can impede their implementation in low- and middle-income countries (LMICs) or rural areas:

• Infrastructure Limitations: Reliable internet connectivity, stable power supply, and adequate computational resources (e.g., powerful servers for running complex deep learning models) are often lacking. This makes cloud-based AI solutions challenging and limits the deployment of resource-intensive on-premise models.
• Cost of Technology: Even if AI software is affordable, the upfront cost of compatible ECG machines, computers, and IT infrastructure can be prohibitive for poorly funded clinics.
• Human Resources and Expertise: A lack of trained IT personnel, data scientists, and even healthcare providers familiar with digital tools can hinder deployment and maintenance. Training programs need to be tailored to local contexts.
• Data Scarcity for Local Training: If models need to be fine-tuned or re-trained on local data to account for population-specific variations, the lack of robust, labeled datasets in these settings becomes a significant barrier.

Innovative solutions are required to overcome these barriers, such as developing lightweight, ‘edge-AI’ models that can operate on low-cost, off-the-shelf devices (e.g., smartphones, tablets) with limited computational power and intermittent internet connectivity. This might involve model compression techniques, quantization, or specialized hardware optimization. Furthermore, collaborative initiatives to build regional data repositories and provide open-source AI tools adapted for local needs are essential for equitable global adoption.

Many thanks to our sponsor Esdebe who helped us prepare this research report.

7. Future Directions

The field of AI-ECGs in pediatric cardiology is rapidly evolving, with ongoing research pushing the boundaries of diagnostic capability and clinical utility. Several key areas represent promising future directions.

7.1 Model Interpretability and Explainable AI (XAI)

As AI-ECGs become more sophisticated, enhancing their interpretability will be crucial for broader clinical acceptance and to foster trust between clinicians and AI systems. Moving beyond ‘black box’ predictions, future AI-ECGs will need to provide clear, actionable explanations for their diagnostic outputs [34].

• Explainable Outputs: Developing models that not only provide a risk score or diagnosis but also highlight specific ECG features (e.g., ‘elevated ST segment in V2-V4’, ‘notched P-wave in lead II’, ‘prolonged QTc interval’) that led to the prediction. Visual saliency maps that pinpoint critical regions of the ECG waveform or specific leads influencing the decision are invaluable [35].
• Causal Inference: Moving towards AI models that can infer causal relationships rather than just correlations. Understanding ‘why’ a particular ECG pattern causes a specific cardiac outcome would significantly deepen clinical understanding and improve diagnostic confidence.
• Interactive Explanations: Allowing clinicians to interact with the AI model to ask ‘what-if’ questions or explore alternative interpretations could enhance their understanding and facilitate shared decision-making with patients.
• Clinical Justification: Integrating clinical guidelines and expert knowledge into the interpretability framework to provide justifications that align with established medical reasoning, bridging the gap between raw data interpretation and clinical wisdom.

7.2 Real-World Validation and Prospective Studies

While promising results have emerged from retrospective analyses, the ultimate test of AI-ECGs lies in their performance during real-world, prospective clinical trials. This is essential for confirming generalizability, assessing clinical impact, and securing regulatory approval.

• Large-Scale, Multi-center Studies: Conducting studies across diverse clinical environments, patient populations, and geographical regions to validate model performance, identify sources of variability, and confirm generalizability. This includes primary care clinics, emergency departments, and various types of pediatric cardiology centers.
• Pragmatic Clinical Trials: Designing trials that test AI-ECGs within routine clinical workflows rather than highly controlled research settings. This provides more realistic data on clinical utility, workflow integration, and impact on patient outcomes [42].
• Impact on Patient Outcomes: Beyond diagnostic accuracy, future studies must evaluate whether AI-ECGs actually lead to improved patient outcomes, such as earlier intervention, reduced morbidity, prevention of sudden cardiac death, or improved quality of life. Cost-effectiveness analyses in real-world settings are also critical.
• Longitudinal Studies: Assessing the long-term impact of AI-ECG screening and diagnosis on pediatric patient cohorts, including disease progression, need for interventions, and overall health trajectories.

7.3 Continuous Learning and Adaptive AI Systems

The static ‘train-then-deploy’ model for AI systems is increasingly being challenged by the dynamic nature of clinical data and evolving medical knowledge. Future AI-ECGs will likely incorporate continuous learning frameworks.

• Adaptive Algorithms: Developing AI models that can continually learn and update themselves from new incoming clinical data, adapting to new disease presentations, changes in population characteristics, or improved diagnostic criteria. This requires robust mechanisms for data curation, model monitoring, and controlled updates to ensure performance stability and prevent ‘model drift’ [29].
• Federated Learning: A decentralized machine learning approach where models are trained locally on individual institutional datasets, and only the learned parameters (not raw data) are shared and aggregated to create a global model. This addresses data privacy concerns while allowing for continuous learning across multiple institutions without centralizing sensitive pediatric data [43].
• Active Learning: AI systems could identify ‘uncertain’ cases where their confidence in a diagnosis is low and automatically flag them for expert human review. The expert’s feedback on these challenging cases can then be used to selectively retrain and improve the model, making the learning process more efficient and targeted.
• Integrated Learning from Multi-modal Data: Combining ECG data with other clinical information, such as electronic health records (EHRs), demographic data, genetic markers, and even point-of-care ultrasound images, could significantly enhance diagnostic accuracy and predictive power. AI models capable of integrating and interpreting such multi-modal data streams are a promising frontier.

7.4 Personalization and Predictive Analytics

Moving beyond population-level diagnoses, AI-ECGs could contribute to personalized medicine in pediatric cardiology.

• Individualized Risk Prediction: Predicting an individual child’s specific risk for developing a certain cardiac condition or experiencing an adverse event based on their unique ECG patterns, genetic profile, and clinical history.
• Therapeutic Guidance: Aiding in the selection of optimal therapies or dosages by predicting individual responses to different treatments based on detailed physiological and electrical markers from ECGs.
• Early Disease Trajectory Prediction: Identifying children at risk of rapid disease progression or those who might benefit most from early preventative interventions. This could include predicting the likelihood of developing LVSD in children with specific genetic mutations or congenital anomalies before symptoms arise.

Many thanks to our sponsor Esdebe who helped us prepare this research report.

8. Conclusion

Artificial Intelligence-Enhanced Electrocardiograms (AI-ECGs) represent a truly transformative advancement in the field of pediatric cardiology, offering a powerful, non-invasive, cost-effective, and highly accessible modality for the early and accurate diagnosis of a myriad of cardiac conditions, particularly left ventricular systolic dysfunction. The underlying deep learning architectures, predominantly Convolutional Neural Networks and hybrid CNN-RNN models, demonstrate an unprecedented ability to extract subtle, clinically significant patterns from ECG signals that often elude conventional human interpretation. Rigorous training methodologies, bolstered by data augmentation and transfer learning, are steadily addressing the inherent challenges of limited and diverse pediatric datasets, leading to models with impressive diagnostic accuracy and promising generalizability across varied demographics.

However, the path to widespread clinical integration is multifaceted and demands careful navigation. Significant practical challenges persist, including the stringent requirements for data privacy and security, the critical imperative to detect and mitigate algorithmic bias to ensure equitable healthcare, and the complex task of seamlessly integrating these advanced tools into existing clinical workflows without causing alert fatigue or disrupting established practices. Furthermore, the substantial resource constraints prevalent in low-income and rural settings necessitate innovative deployment strategies to avoid exacerbating health disparities and ensure truly equitable access.

The ethical dimensions surrounding AI-ECGs are equally profound, encompassing informed consent, algorithmic transparency, and accountability, all of which are paramount to fostering trust among clinicians, patients, and the public. Looking ahead, the future directions for AI-ECGs are exciting and expansive, focusing on enhancing model interpretability through Explainable AI (XAI) techniques, conducting robust real-world validation through large-scale prospective studies, and developing continuous learning frameworks that allow models to adapt and improve over time. The ultimate goal is to move towards truly personalized and predictive pediatric cardiac care, where multi-modal data integration and individualized risk prediction become standard practice.

With sustained interdisciplinary research, meticulous validation, thoughtful regulatory engagement, and a steadfast commitment to ethical principles, AI-ECGs possess the profound potential to significantly enhance pediatric cardiac care. They can democratize access to advanced diagnostics, optimize resource allocation, and ultimately improve the long-term health outcomes for countless children worldwide, particularly in underserved regions. The current trajectory suggests that AI-ECGs are poised to become an indispensable component of the pediatric cardiologist’s toolkit, marking a new era in the vigilance and management of cardiac health in children.

Many thanks to our sponsor Esdebe who helped us prepare this research report.

References

[1] Hoffman, J. I. E., & Kaplan, S. (2000). The incidence of congenital heart disease. Journal of the American College of Cardiology, 36(6), 1890-1900.

[2] Lipshultz, S. E., et al. (2013). The Future of Pediatric Cardiology: Part 1: Cardiovascular Disease in Childhood. Journal of the American College of Cardiology, 62(15), 1361-1372.

[3] Geva, T. (2018). Echocardiography in congenital heart disease: an update. Current Opinion in Cardiology, 33(5), 455-463.

[4] Powell, A. J. (2017). Magnetic Resonance Imaging in Pediatric Cardiology. Pediatric Cardiology, 38(1), 177-185.

[5] Kligfield, P., et al. (2007). Recommendations for the Standardization and Interpretation of the Electrocardiogram: Part I: The Electrocardiogram and Its Technology A Scientific Statement From the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society. Circulation, 115(9), 1306-1327.

[6] Davignon, A., et al. (1979). Normal ECG standards for infants and children. Pediatric Cardiology, 1(1), 123-141.

[7] Mayourian, J., La Cava, W. G., de Ferranti, S. D., Mah, D., Alexander, M., Walsh, E., & Triedman, J. K. (2024). Detection of Right and Left Ventricular Dysfunction in Pediatric Patients Using Artificial Intelligence–Enabled ECGs. Journal of the American Heart Association, 13(11), e035201. (pmc.ncbi.nlm.nih.gov)

[8] Niaz, T., et al. (2024). Revolutionizing pediatric cardiac care through AI-enabled ECGs. Mayo Clinic Proceedings, 100(11), 2025–2035. (mayoclinic.org)

[9] Kiranyaz, S., et al. (2021). Convolutional Neural Networks for ECG-Based Arrhythmia Detection: A Review. IEEE Reviews in Biomedical Engineering, 14, 1-17.

[10] Rajpurkar, P., et al. (2017). Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature Medicine, 23(1), 849-856.

[11] Glorot, X., & Bengio, Y. (2010). Understanding the difficulty of training deep feedforward neural networks. Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics (AISTATS 2010), 9, 249-256.

[12] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). ImageNet Classification with Deep Convolutional Neural Networks. Advances in Neural Information Processing Systems (NIPS 2012), 25, 1097-1105.

[13] Lipton, Z. C., et al. (2015). A critical review of recurrent neural networks for sequence learning. arXiv preprint arXiv:1506.00019.

[14] Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory. Neural Computation, 9(8), 1735-1780.

[15] Acharya, U. R., et al. (2017). Deep convolutional neural network for automated detection of myocardial infarction using ECG signals. Information Sciences, 385-386, 325-337.

[16] Pustozerov, E., et al. (2019). Hybrid CNN-RNN Architecture for ECG Signal Classification. Proceedings of the 2019 International Conference on Information Technology and Nanotechnology (ITNT), 1-4.

[17] Yang, Y., Peracchio, L., Mayourian, J., Triedman, J. K., Miller, T., & La Cava, W. G. (2025). Robust AI-ECG for Predicting Left Ventricular Systolic Dysfunction in Pediatric Congenital Heart Disease. arXiv preprint. (arxiv.org)

[18] Zhang, D., Yuan, X., & Zhang, P. (2020). Interpretable Deep Learning for Automatic Diagnosis of 12-lead Electrocardiogram. arXiv preprint. (arxiv.org)

[19] Gu, S., et al. (2020). ECG Data Augmentation Using Generative Adversarial Networks. 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), 3217-3220.

[20] Pan, S. J., & Yang, Q. (2010). A Survey on Transfer Learning. IEEE Transactions on Knowledge and Data Engineering, 22(10), 1345-1359.

[21] Ghassemi, M., et al. (2020). A review of transfer learning for electrocardiogram analysis. Artificial Intelligence in Medicine, 107, 101899.

[22] Kohavi, R. (1995). A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection. Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, 2, 1137-1143.

[23] Luo, C., et al. (2023). External validation of AI models in healthcare: A systematic review. npj Digital Medicine, 6(1), 1-14.

[24] Hanley, J. A., & McNeil, B. J. (1982). The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology, 143(1), 29-36.

[25] Saito, T., & Rehmsmeier, M. (2015). The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLoS One, 10(3), e0118432.

[26] Chen, Y. (2025). A Benchmark Study of Deep Learning Methods for Multi-Label Pediatric Electrocardiogram-Based Cardiovascular Disease Classification. arXiv preprint. (arxiv.org)

[27] U.S. Food & Drug Administration. (2017). De Novo Classification Request. Retrieved from fda.gov

[28] U.S. Food & Drug Administration. (2020). 510(k) Premarket Notification. Retrieved from fda.gov

[29] U.S. Food & Drug Administration. (2021). Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan. Retrieved from fda.gov

[30] European Medicines Agency. (2020). EMA reflection paper on the use of artificial intelligence in the medicinal product lifecycle. Retrieved from ema.europa.eu

[31] General Data Protection Regulation (GDPR). (2016). Regulation (EU) 2016/679 of the European Parliament and of the Council. Retrieved from gdpr-info.eu

[32] Obermeyer, Z., et al. (2019). Dissecting racial bias in an algorithm used to manage the health of millions of Black patients. Science, 366(6468), 447-453.

[33] Barocas, S., Hardt, M., & Narayanan, A. (2019). Fairness and Machine Learning: Limitations and Opportunities. MIT Press.

[34] Ghassemi, M., et al. (2021). The Future of Trustworthy AI in Health. The Lancet Digital Health, 3(3), e135-e137.

[35] Jiang, R., et al. (2024). Deep Learning–Augmented ECG Analysis for Screening and Genotype Prediction of Congenital Long QT Syndrome. JAMA Cardiology, 9(3), 234–242. (jamanetwork.com)

[36] Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). ‘Why Should I Trust You?’: Explaining the Predictions of Any Classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 1135-1144.

[37] Lundberg, S. M., & Lee, S. I. (2017). A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems (NIPS 2017), 30, 4765-4774.

[38] Vest, J. R., & Gamm, L. D. (2010). An integrative review of electronic health records and healthcare quality. Journal of the American Medical Informatics Association, 17(5), 570-580.

[39] Zhang, B. H., & Neff, G. (2019). ‘AI and data bias: The problem with biased data’. Journal of Big Data, 6(1), 1-15.

[40] Automated ECG interpretation. (2025). In Wikipedia. (en.wikipedia.org)

[41] Wiens, J., et al. (2019). The Future of Healthcare: Artificial Intelligence. Journal of the American Medical Association (JAMA), 321(18), 1817-1818.

[42] Tunis, S. R., et al. (2011). Practical Clinical Trials: Increasing the Value of Clinical Research for ‘Real World’ Care. Journal of the American Medical Association (JAMA), 306(11), 1255-1259.

[43] Rieke, N., et al. (2020). The future of digital health with federated learning. npj Digital Medicine, 3(1), 1-13.