Cross-Modal Contrastive Learning of ECG and Angiography Representations for Severe Stenosis Classification

Technical Summary: Cross-Modal Contrastive Learning of ECG and Angiography Representations for Severe Stenosis Classification

Problem Statement
Coronary artery stenosis is a critical cardiovascular condition where severe, untreated cases significantly increase the risk of heart failure and death. The current gold standard for diagnosis is coronary X-ray angiography. However, this procedure is invasive, resource-intensive, and carries a small mortality risk; consequently, it is typically reserved for patients with a high pre-test probability of disease based on symptoms. This approach leaves asymptomatic patients at risk of remaining undiagnosed as the disease progresses. While Electrocardiograms (ECGs) are fast, cheap, non-invasive, and routinely acquired, they have historically provided limited information for stenosis diagnosis. Patients with severe stenosis often present with normal ECGs, and previous studies indicate that ECGs alone are insufficient for reliable stenosis risk stratification without auxiliary data.

Methodology: StenCE Framework
To address the lack of stenosis-specific signals in ECGs, the authors propose StenCE, a contrastive pretraining framework designed to distill knowledge from angiography into ECG representations. The approach follows a multi-stage pipeline:

1. Data Alignment: The framework utilizes a hospital dataset of 3,414 patients (4,067 matched pairs) where each patient has a 12-lead ECG and a corresponding X-ray angiogram. The angiogram data includes annotations for blockage severity in specific coronary segments, aggregated to patient-level labels.
2. Dual-Encoder Architecture:
   • Angiography Encoder: Based on the SegmentMIL model, this component processes multi-view angiography data. It employs a view encoder and a transformer-based "views aggregator" to generate study-level, artery-level, and segment-level predictions. A new CLS query token is introduced to extract a global study-level representation.
   • ECG Encoder: The framework evaluates two backbone options: OTIS (a 7.1M-parameter transformer pretrained on general time-series data) and EchoingECG (a 128M-parameter 1D-ResNet pretrained on echocardiogram-text-ECG triplets).
3. Contrastive Pretraining: The core mechanism employs Cross-Modal Contrastive Learning (CLIP loss). The ECG encoder is trained to align its feature representations with those of the angiography encoder in a shared embedding space.
   • The angiography encoder is kept frozen during this phase to preserve its learned stenosis features.
   • The ECG encoder, the views aggregator (specifically the CLS token output), and linear projection layers are trained.
   • A Supplementary Stenosis Loss is applied to the angiography side (weighted at $W_{Sten} = 0.3$) to maintain the model's ability to classify stenosis severity directly from the angiogram during pretraining.
4. Fine-Tuning: After pretraining, the angiography encoder is discarded. The pretrained ECG encoder is fine-tuned on downstream tasks (stenosis classification and other cardiac abnormalities) using only ECG inputs, with labels derived from the associated angiography studies.

Key Contributions

1. Novel Framework: The introduction of StenCE, the first framework to align ECG representations with multi-view angiography features via contrastive learning to enable stenosis detection from ECGs alone.
2. Performance Breakthrough: The demonstration that severe stenosis can be identified solely from ECGs with high performance, overcoming the historical limitation that ECGs lack specific stenosis signals.
3. Generalizability: Validation of the pretraining framework not only on stenosis classification across varying severity thresholds but also on a separate dataset (EchoNext) for detecting various cardiac abnormalities (e.g., LVEF, aortic stenosis, mitral regurgitation).

Experimental Results
The authors evaluated the models using full fine-tuning and linear probing across multiple severity thresholds (e.g., $\ge 100\%$ blockage vs. $0\%$, $\ge 90\%$ vs. $0\%$).

• Severe Stenosis Detection: The best-performing model, EchoingECG-StenCE (full fine-tuning), achieved an AUC of 0.822 for the most severe threshold ($=0 | =100$). This represents an 11% improvement over the base EchoingECG model and is the first reported high-performance result for ECG-only severe stenosis classification.
• Linear Probing: The benefits of StenCE were even more pronounced in the linear probing setting, where EchoingECG-StenCE achieved an AUC of 0.816 (a 17% improvement over the base model) and OTIS-StenCE improved by 8% over the base OTIS.
• Severity Trends: As expected, performance decreased for less severe thresholds (AUC dropped to 0.704 for $\ge 90\%$ and approached random chance for $\ge 70\%$), indicating the model is currently most effective for detecting the most critical cases.
• Ablation Studies: Results confirmed that keeping the angiography encoder frozen and using a moderate stenosis supervision weight ($W_{Sten} = 0.3$) yielded the optimal balance between contrastive alignment and task-specific feature retention.

Significance and Claims
The paper claims that StenCE successfully enables the extraction of stenosis-relevant signals from ECGs by leveraging the rich, diagnostic information present in angiography. The authors state that their models are the first to achieve high performance in severe stenosis classification using only ECG inputs.

However, the authors maintain a modest tone regarding clinical deployment. They explicitly acknowledge that while the results support early diagnosis and risk stratification, the current performance levels (particularly for less severe cases) are not yet sufficient for direct use in clinical practice. The study highlights a selection bias in the dataset (patients already undergoing angiography due to symptoms) and notes the absence of clinical risk factors in the model input. The primary significance lies in demonstrating the viability of cross-modal pretraining to bridge the gap between non-invasive screening (ECG) and invasive gold-standard diagnosis (angiography), paving the way for future developments in preliminary coronary artery assessment.