Multi-label ECG Classification

• Multi-label ECG classification is the automated assignment of multiple co-occurring cardiac diagnosis labels using advanced deep learning techniques.
• It leverages methods like attention-based temporal CNNs, hybrid CNN-RNNs, and state-space models to capture multi-scale features and address label imbalance.
• The approach enhances clinical decision-making through robust feature extraction, interpretable prototype reasoning, and efficient risk stratification in diverse ECG datasets.

Multi-label ECG classification refers to the automated assignment of multiple, potentially co-occurring diagnostic labels to electrocardiogram (ECG) recordings. Unlike traditional single-label frameworks—where each ECG is associated with exactly one diagnosis—multi-label approaches are designed to recognize multimorbidity: the simultaneous presence of multiple cardiac conditions. This paradigm is essential in clinical environments, reflecting the reality that many patients present with overlapping rhythm, conduction, ischemic, and morphological disturbances. Multi-label classification leverages modern deep learning architectures, attention mechanisms, domain adaptation, and interpretability strategies to advance reliable, accurate diagnosis and clinical decision support.

1. Problem Definition and Clinical Significance

Multi-label ECG classification models predict a binary vector $\mathbf{y} \in \{0,1\}^C$ for each recording, where $C$ denotes the total number of clinically relevant cardiac disorders under consideration. These can range from rhythm and conduction issues (e.g., atrial fibrillation, bundle branch block) to ischemic pathologies (e.g., ST-elevations, myocardial infarction) and rare congenital defects, with substantial label imbalance and frequent label co-occurrence. The clinical importance is heightened by the significantly increased mortality associated with multimorbidity, especially in cardiovascular disease populations (Prabhakararao et al., 2023). Multi-label frameworks thus support comprehensive risk stratification and enable holistic screening in routine ECG analysis and longitudinal patient monitoring.

2. Algorithmic Frameworks and Representative Architectures

The multi-label ECG domain encompasses several major algorithmic strategies, distinguished by their approach to feature extraction and label prediction.

2.1 Attention-Based Temporal CNNs

The ATCNN framework (Prabhakararao et al., 2023) employs an ensemble of binary classifiers, each comprising temporal convolutional layers to capture multi-scale ECG dynamics. Spatial and temporal attention modules focus feature extraction both within and across leads:

• Temporal attention weights $\alpha_m$ highlight diagnostically relevant time windows for each lead.
• Spatial attention weights $\beta$ reweight the contribution of individual leads, facilitating channel redundancy pruning and interpretable optimal-subset selection.

2.2 Hybrid CNN-RNN Architectures

Hybrid models utilize convolutional blocks for local morphology and recurrent layers for rhythm/context (Jafari et al., 25 Jan 2026):

• Empirical evidence demonstrates that a single-layer BiLSTM appended to a CNN backbone achieves minimal Hamming loss and optimal micro-F1, with diminishing returns (and overfitting risks) from deeper or stacked RNNs.
• Attention gating after RNNs enables interpretable feature-wise channel aggregation.

2.3 Prototype-Based Reasoning

ProtoECGNet (Sethi et al., 11 Apr 2025) introduces a multi-branch architecture with rhythm (1D CNN/global prototype), morphology (2D CNN/local prototype), and global abnormality (2D CNN/global prototype) branches. Learning is guided by clustering, separation, diversity, and co-occurrence-aware contrastive losses. The resulting prototypes can be projected onto real ECG segments, supporting explicit, case-based interpretability without sacrificing model performance.

2.4 Multi-Resolution Fusion and State-Space Models

• MRM-Net (Huang et al., 2024) incorporates parallel dual-resolution attention branches and a mutual-learning mechanism, where channel attention (CA), spatial attention (SA), and KL-based mimicry facilitate robust multi-scale feature capture.
• State-space models (e.g., 1DCNN-ECG-Mamba (Jiang et al., 14 Oct 2025), Mamba 2 (Chen, 4 Oct 2025)) implement selective SSM blocks for efficient global sequence modeling, outperforming transformer encoders in handling lengthy ECG inputs.

2.5 Semi-Supervised and Knowledge Distillation Paradigms

• ECGMatch (Zhou et al., 2023) leverages a mean-teacher/student pipeline, multi-level augmentation, neighbor agreement modeling, and label correlation alignment to propagate diagnostic knowledge under extreme label scarcity.
• MVKT-ECG (Qin et al., 2023) uses inter-lead multi-view contrastive learning and multi-label knowledge distillation, substantially improving single-lead classifier accuracy by transferring features extracted from 12-lead teachers.

3. Feature Extraction and Preprocessing Techniques

Robust feature representation is critical for multi-label ECG classification. Several contemporary approaches are:

• Dilated Temporal Convolutions (ATCNN): Capturing waveforms at multiple scales, critical for discriminating P-waves, QRS complexes, and ST/T abnormalities (Prabhakararao et al., 2023).
• Wavelet and Scatter Transforms: The SCATTER-ResNet (Oppelt et al., 2020) employs theoretically motivated wavelet decompositions for multi-scale, stable feature extraction, with trainable adaptation of the entire classification pipeline.
• Domain-Specific Augmentation: ECGMatch (Zhou et al., 2023) deploys channel permutation, time flipping, signal dropout, and additive noise to diversify data and ensure generalization under limited annotations.
• Time-Frequency Analysis: xLSTM-ECG (Kang et al., 14 Apr 2025) converts signals into STFT cubes, enhancing separation of cardiac bands and improving robustness to noise.

Preprocessing is standardized to bandpass filtering (0.5–45 Hz), per-lead z-score normalization, length cropping/padding, and, in image-based tasks, binarization and artifact removal (Nguyen et al., 19 Feb 2025).

4. Training Mechanisms, Loss Functions, and Evaluation Metrics

Loss Functions

• Multi-label binary cross-entropy (BCE) forms the primary task loss (Prabhakararao et al., 2023, Jiang et al., 14 Oct 2025, Huang et al., 2024).
• Weighted BCE or focal loss is typical for addressing severe class imbalance (Chen, 4 Oct 2025, Wong et al., 2020).
• Prototype and contrastive losses supplement clustering, separation, and diversity objectives in interpretable models (Sethi et al., 11 Apr 2025).

Optimization

• Adam and AdamW optimizers are predominant, occasionally incorporating cosine annealing or learning-rate plateau reduction (Jiang et al., 14 Oct 2025, Chen, 4 Oct 2025).
• Early stopping on validation objective (AUROC, F1) mitigates overfitting (Kang et al., 14 Apr 2025, Jafari et al., 25 Jan 2026).

Metrics

Standard evaluation includes Hamming loss, micro/macro-F1 scores, mean average precision (MAP), AUROC/AUPRC, and subset accuracy:

• Hamming loss quantifies per-label misclassification rate (Jafari et al., 25 Jan 2026, Chen, 4 Oct 2025).
• Macro-F1 treats all classes equally, highlighting rarer condition performance (Chen, 4 Oct 2025).
• Subset accuracy measures exact multi-label vector matches (Prabhakararao et al., 2023, Jafari et al., 25 Jan 2026).

5. Dataset Landscape and Benchmarking

Prominent ECG classification datasets include:

Dataset	#Records	Leads	Label Structure	Domain
PTB-XL	21,837	12	71 SCP-ECG codes	Adult, CVD
CPSC 2018	~6,877	12	9 rhythm/morph classes	Arrhythmia
ZZU-pECG	3,716	9/12	19 pediatric CVDs	Pediatrics
PhysioNet/CinC	43,101	12	27 multi-label classes	Mixed
ICM/ICM-data	3,607+	1	5 rhythm/morph classes	Device
Georgia/Chapman	40,258	12	7–19 CVD classes	Mixed, US

Class imbalance and multi-label prevalence range from 23 % in PTB-XL (Prabhakararao et al., 2023) to 28.6 % in broader PTB-XL stratifications (Jafari et al., 25 Jan 2026) and 19-label pediatric cohorts (Chen, 4 Oct 2025). Children’s datasets emphasize the granularity necessary for age-adaptive and multi-class screening (Chen, 4 Oct 2025).

6. Interpretability, Robustness, and Clinical Integration

Attention Visualization:

Spatial and temporal attention mechanisms enable ranking and selection of diagnostically pivotal leads and signal segments (Prabhakararao et al., 2023, Huang et al., 2024). Such visualizations align with clinical conventions (e.g., emphasizing V1/V2 in bundle branch block) and directly inform interpretability.

Prototype Reasoning:

The ProtoECGNet pipeline (Sethi et al., 11 Apr 2025) grounds decisions in similarity to representative ECG segments:

• Structured clinician review confirms prototype representativeness and clarity.
• Multi-branch reasoning mirrors the human clinical workflow (rhythm, morphology, diffuse abnormalities), supporting trust and clinical adoption.

Noise Robustness and Domain Adaptation:

Pipelines tailored to device artifacts and suboptimal acquisition (e.g., ICM data, scanned paper ECGs) employ semi-supervised clustering, artifact-specific preprocessing (CEEMD), and image-based augmentation to maintain sensitivity and specificity under real-world constraints (Bleich et al., 2023, Nguyen et al., 19 Feb 2025).

Clinical Deployment:

Efficient architectures (single BiLSTM (Jafari et al., 25 Jan 2026), MVKT-ECG (Qin et al., 2023)) and inference latency (≤10 ms/record) target edge devices, smartwatches, and resource-limited settings. Attention weights, prototypes, and channel gating support post-hoc reasoning and highlight actionable risk criteria for clinicians.

7. Current Challenges and Future Directions

• Class Imbalance and Rare Disease Recognition:

Under-recognition of infrequent diagnoses persists; solutions include class-weighted/focal losses, oversampling, and few-shot learning (Chen, 4 Oct 2025, Wong et al., 2020).

• Feature Fusion and Multi-Scale Learning:

Integrative techniques such as dual-resolution attention and feature complementary modules (MRM-Net (Huang et al., 2024)) or xLSTM fusion (Kang et al., 14 Apr 2025) facilitate robust rhythm and morphology detection.

• Semi-Supervised and Cross-Domain Generalization:

ECGMatch (Zhou et al., 2023) demonstrates strong cross-population performance using only 1–5 % labeled data, essential for deployment in heterogeneous clinical environments.

• Interpretable AI:

Prototype-based learning and multi-view knowledge transferring models bridge the explainability gap, supporting transparent diagnostics and structured clinician review (Sethi et al., 11 Apr 2025, Qin et al., 2023).

• Validation and Real-World Impact:

Multi-center, age-stratified, and federated benchmarks (ZZU-pECG (Chen, 4 Oct 2025), PTB-XL) are necessary for transition from research to clinical adoption, with ongoing studies on external domain robustness, continuous adaptation, and multimodal integration.

---

Multi-label ECG classification stands at the intersection of advanced temporal modeling, multi-scale feature fusion, and interpretable representation learning. Recent frameworks—including attention-based TCNNs, hybrid CNN-RNNs, selective state-space models, and prototype-guided architectures—attain state-of-the-art accuracy while progressively addressing key issues of label co-occurrence, multimodality, rarity, and interpretability (Prabhakararao et al., 2023, Jafari et al., 25 Jan 2026, Huang et al., 2024, Sethi et al., 11 Apr 2025, Qin et al., 2023, Zhou et al., 2023, Jiang et al., 14 Oct 2025, Chen, 4 Oct 2025). The ongoing expansion of robust benchmarks, semi-supervised pipelines, and clinically aligned reasoning mechanisms continues to drive the evolution of multi-label ECG classification as an essential technology in digital cardiology.