PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph-Based Pseudo-multimodal Contrastive Learning for 12-Lead ECG Representations

Mengyu Wang, Kozo Okada, Takafumi Goto, Natsuko Jinba, Hiroki Yamaya, Kiyoshi Hibi, Tomoki Hamagami

arXiv:2608.26964v1cs.LG

TL;DR

Existing ECG analysis methods often inadequately capture global waveform structure and inter-lead dependency. Graph-CMMC addresses this gap by aligning waveform and GADF representations with self-supervised contrastive learning and a shared graph structure, achieving competitive downstream performance and producing task-relevant inter-lead representations.

  • Problem

    Existing ECG methods often focus on single leads or independent one-dimensional processing, limiting capture of global waveform patterns and inter-lead dependency important for diagnosis.

  • Method

    Graph-CMMC constructs complementary waveform and GADF representations from each 12-lead recording, aligns them self-supervised, and models inter-lead dependency with a learnable shared graph structure.

  • Results

    Graph-CMMC achieves superior performance to existing self-supervised methods on multi-label coronary artery occlusion classification, approaching fully supervised learning.

  • Takeaways & Limitations

    Combining pseudo-multimodal representation learning with explicit graph-based modeling yields robust 12-lead ECG representations and task-relevant inter-lead information flow.

Abstract

from arXiv · show

12-lead electrocardiogram (ECG) is a standard, non-invasive examination widely used for diagnosing coronary artery disease, where clinical interpretation relies on comparing waveform patterns across multiple leads. However, most existing ECG analysis methods focus on single-lead signals or treat each lead independently, and typically process ECG signals as one-dimensional time-series data using CNNs or RNNs. While effective in modeling local waveform changes, such approaches have difficulty capturing inter-lead dependency and global waveform patterns essential for clinical diagnosis. To address this limitation, we propose a graph-based pseudo-multimodal contrastive learning framework called Graph-CMMC. ECG waveforms are transformed into Gramian Angular Difference Field (GADF) images to construct complementary representations of the same cardiac activity, enabling a pseudo-multimodal learning setting. Using all 12 leads, Graph-CMMC aligns waveform and GADF representations in a self-supervised manner, while a graph-based relational module is employed to model inter-lead dependency and enforce structural consistency across leads during contrastive learning. Experimental results on a multi-label coronary artery occlusion classification task demonstrate that the proposed framework achieves competitive performance compared to supervised learning methods. These results further suggest the effectiveness of using GADF as a complementary representation and incorporating explicit graph-based modeling of inter-lead dependency for learning robust 12-lead ECG representations.

I. INTRODUCTION

Graph-CMMC addresses limitations in 12-lead ECG analysis by combining complementary waveform and GADF representations with self-supervised graph-based modeling of inter-lead dependencies. It is evaluated for multi-label coronary artery occlusion classification.

  • Clinical ECG interpretation depends on both temporal waveform morphology within each lead and relationships among leads.
  • CNN- and RNN-based approaches often process leads independently, limiting their ability to capture global waveform structure and inter-lead dependency.
  • Graph-CMMC constructs complementary waveform and GADF representations from the same ECG recordings without requiring additional data.
  • Self-supervised contrastive learning aligns waveform and GADF representations across all 12 leads while enforcing structural consistency.
  • Experiments assess whether Graph-CMMC learns robust representations on a multi-label coronary artery occlusion classification task.
  • A learnable graph structure explicitly models inter-lead dependency during representation learning.

II. RELATED WORKS

Prior ECG methods include time-series, image-transformation, supervised multi-lead, reconstruction-based, and contrastive approaches. Their limitations motivate explicitly modeling inter-lead structure and complementary representations during self-supervised learning.

  • Time-series-based ECG analysis: CNNs and RNNs model ECGs as one-dimensional temporal sequences and have detected abnormalities, but commonly treat leads independently.
  • Time-series-based ECG analysis: GAF and GADF transformations produce two-dimensional images emphasizing waveform morphology and temporal correlation patterns.
  • Time-series-based ECG analysis: Prior GADF-based studies using single-lead signals, including Lead II, limit applicability to multi-lead ECG analysis.
  • Multi-lead ECG modeling: Most multi-lead ECG models use supervised, task-specific architectures to model inter-lead relationships.
  • Self-supervised ECG representation learning: Reconstruction methods such as ST-MEM recover masked one-dimensional signals but may miss global morphology and remain sensitive to local fluctuations and baseline drift.
  • Self-supervised ECG representation learning: Contrastive methods align positive pairs and separate negatives to learn transferable representations, while CMLC uses temporally aligned signals from different leads as positive pairs.
  • Self-supervised ECG representation learning: Existing contrastive methods primarily use inter-lead relationships to define positive pairs rather than imposing structured relational constraints across representation spaces.

III. PROPOSED METHOD

Graph-CMMC constructs complementary waveform and GADF views of each 12-lead ECG, then aligns them during self-supervised pretraining. GADF emphasizes global waveform structure while the downstream model uses the waveform encoder.

  • Graph-CMMC is a self-supervised framework for 12-lead ECG representation learning that uses raw waveforms and GADF images.
  • A. Pseudo-multimodal: Pseudo-multimodal representation combines two complementary encodings of the same cardiac activity rather than adding multimodal inputs downstream.
  • 1) GADF Transform:: GADF transforms each waveform into a two-dimensional representation emphasizing global morphology and temporal correlation patterns.
  • 1) GADF Transform:: The GADF transform normalizes the signal, maps values to angular values, and computes sine-based pairwise angular differences.
  • 1) GADF Transform:: The generated GADF image is resized to 64 × 64 for this study.

B. Graph-based Module

The graph-based module models each ECG lead as a node and exchanges information across leads. A shared adjacency structure learned from waveform representations also constrains the GADF branch.

  • Each ECG lead is represented as a graph node, with an adjacency matrix encoding inter-lead dependency.
  • Graph convolution performs relational message passing on lead-level features extracted by the modality-specific encoders.
  • The adjacency matrix is learned in the waveform branch, shared with the GADF branch, and not independently optimized there.
  • Waveform representations provide the primary basis for relational structure, while GADF serves as a complementary derived view.
  • A shared graph structure encourages both representations to preserve consistent lead-wise dependency patterns during learning.

C. Contrastive Learning

Graph-CMMC aligns waveform and GADF representations from the same ECG recording while separating representations from different recordings. Its symmetric NT-Xent objective preserves inter-lead dependency through the shared graph structure.

  • Graph-CMMC uses SimCLR-style contrastive learning to align waveform-based and GADF-based representations.
  • The projected branch representations are aggregated and mapped into a shared embedding space through a lightweight MLP projection head.
  • Representations from the same ECG recording form positive pairs, while representations from different recordings in the mini-batch form negative samples.
  • The NT-Xent loss uses waveform–GADF positive pairs and separates non-corresponding pairs within the batch.
  • The symmetric contrastive objective averages waveform-to-GADF and GADF-to-waveform losses, using cosine similarity and a temperature parameter.
  • Optimizing the objective aligns the two representations while preserving inter-lead dependency through the shared graph structure.

IV. EXPERIMENT

The study evaluates Graph-CMMC on real clinical 12-lead ECG recordings for multi-label coronary artery occlusion classification. Experiments use a two-stage pretraining and downstream classification setup with a 6:2:2 data split.

  • The dataset contains real clinical ECG recordings collected from Yokohama City University Medical Center.
  • Recordings from 1,068 patients are annotated with coronary artery occlusion information determined using coronary angiography.
  • The study selects 3,641 fixed-window ECG samples, each containing 2,000 points corresponding to approximately four cardiac cycles.
  • The downstream task is multi-label classification of occlusion in LAD, LCX, RCA, and LMT arteries.
  • Data are split into training, validation, and test sets at a 6:2:2 ratio without enforced patient-level separation.
  • Experiments consist of self-supervised pretraining followed by downstream multi-label classification.

1) Self-supervised Pretraining:

Self-supervised pretraining uses unlabeled training data to learn representations, followed by frozen-encoder downstream classification with an MLP. Performance is assessed using four multi-label metrics averaged across three test runs.

  • Self-supervised pretraining uses the entire training set without label information.
  • The pretrained waveform encoder is frozen and paired with an MLP classifier for downstream multi-label classification.Only the classification head is optimized during downstream training.
  • Performance is evaluated with Macro F1-score, Subset accuracy, Overall accuracy, and Jaccard index.Macro F1 averages per-label F1-scores, while Subset accuracy requires an exact label-set match.
  • Test results are averaged over three experimental runs.

4) Comparison methods:

The comparison includes supervised and self-supervised baselines using matched data splits, architectures, and classification heads. The self-supervised methods differ in whether they use inter-lead pairing, cross-representation alignment, or both.

  • All comparison methods use the same dataset split, encoder architecture, and downstream classification head.This setup is intended to ensure a fair evaluation.
  • The supervised baseline trains a 2-layer 1D CNN encoder and MLP classifier end-to-end with labeled data.
  • CMLC treats temporally aligned signals from different leads of the same patient as positive pairs and signals from other patients as negatives.
  • CMMC aligns waveform-based and GADF-based representations from the same ECG recording while treating different recordings as negatives.Unlike Graph-CMMC, CMMC considers each lead independently and does not model inter-lead relationships with graph structures.
  • CMMC+CMLC jointly applies cross-modal contrastive learning and multi-lead contrastive learning.

C. Results

Graph-CMMC is evaluated in encoder-only and graph-enhanced downstream settings against supervised and self-supervised methods. It performs best among self-supervised methods, while the supervised baseline remains the upper-bound reference.

  • The evaluation compares Graph-CMMC (d1), encoder-only transfer, with Graph-CMMC (d2), which adds learned-adjacency graph aggregation before classification.The comparison tests whether pretrained inter-lead dependency remains effective downstream.
  • The fully supervised baseline achieves the highest performance and serves as an upper-bound reference.Table I reports Macro F1-score, Subset Accuracy, Overall Accuracy, and Jaccard index.
  • Among self-supervised methods, Graph-CMMC achieves the best performance across all evaluation metrics.
  • Graph-CMMC (d2) outperforms Graph-CMMC (d1), providing additional gains beyond encoder-only transfer.
  • CMMC improves over CMLC, but CMLC+CMMC does not outperform CMMC alone.The results suggest that combining multi-lead and cross-modal contrastive objectives does not necessarily produce additive improvements.
  • Graph-CMMC (d2) achieves higher F1-scores than Graph-CMMC (d1) for LAD, LCX, LMT, and RCA and outperforms other self-supervised baselines across classes.The results indicate benefits across multiple coronary regions.

1) Feature space analysis of Inter-Lead Representations:

Inter-lead embedding analysis shows that Graph-CMMC learns compact and stable relationships, with directional dependencies and four learned communities. These communities do not strictly match clinical anatomical groupings, although the structure remains informative for the downstream task.

  • 1) Feature space analysis of Inter-Lead Representations:: Inter-lead analysis computes 66 pairwise Euclidean distances from 12 lead-wise embeddings and summarizes their distributions with four statistics.
  • 1) Feature space analysis of Inter-Lead Representations:: Graph-CMMC achieves the smallest mean distance and lowest standard deviation among all methods.
  • 1) Feature space analysis of Inter-Lead Representations:: The learned structure is interpreted as compact and stable without trivial collapse or excessive separation of specific leads.
  • 1) Feature space analysis of Inter-Lead Representations:: Graph-CMMC produces a narrower interquartile range and fewer extreme values than other methods, indicating more consistent inter-lead relationships across samples.
  • 2) Visualization of learned Inter-lead dependency in Graph-CMMC:: Graph-CMMC learns a nonsymmetric adjacency matrix, capturing directional dependencies among leads.The asymmetry reflects differing spatial orientations of the 12 ECG leads.
  • 2) Visualization of learned Inter-lead dependency in Graph-CMMC:: Louvain clustering partitions the 12 leads into four learned communities: {I, aVF}, {aVL, V1, V3, V6}, {aVR, V4, V5}, and {II, III, V2}.
  • 2) Visualization of learned Inter-lead dependency in Graph-CMMC:: The learned communities do not strictly correspond to textbook anatomical groupings, but the adjacency structure remains informative for the downstream task.

V. CONCLUSION

Graph-CMMC combines pseudo-multimodal waveform–GADF learning with explicit graph modeling of inter-lead dependency for 12-lead ECG representation learning. It improves downstream performance and yields compact, stable representations, while learned dependencies remain distinct from predefined anatomical groupings.

  • Graph-CMMC constructs complementary waveform and GADF representations, aligns them through contrastive learning, and models inter-lead dependency with a learnable graph structure.
  • Compared with existing self-supervised methods, Graph-CMMC achieves superior performance on multi-label coronary artery occlusion classification, approaching fully supervised learning.
  • Graph-CMMC produces more compact and stable inter-lead representations than comparison methods.
  • The learned adjacency matrix reveals directional inter-lead dependencies reflecting task-relevant information flow, but these structures do not strictly match predefined anatomical groupings.
  • The framework is presented as a step toward bridging data-driven ECG representation learning and multi-lead relational reasoning in clinical practice.
  • Future work will test larger, more diverse datasets with patient-wise splitting and investigate graph designs, architectural ablations, downstream strategies, and domain knowledge.
Loading 2608.26964v1…