Key result
Self-supervised ECG model detects ATTR-CA with AUC 0.83, outperforming supervised learning using limited labeled data.
Why the study?
Transthyretin cardiac amyloidosis diagnosis is often delayed, and traditional supervised AI struggles with rare conditions due to limited labeled ECG data.
Does a self-supervised learning approach improve the detection of transthyretin cardiac amyloidosis from ECGs compared to conventional supervised learning?
Does a self-supervised learning approach improve the detection of transthyretin cardiac amyloidosis from ECGs compared to conventional supervised learning?
A self-supervised learning foundation model significantly outperforms traditional supervised AI in detecting transthyretin cardiac amyloidosis from ECGs, offering a scalable tool for rare disease screening with limited labeled data.
May aid ECG-based ATTR-CA screening with scarce labels; leaves open prospective validation before clinical use.
Background: The diagnosis of transthyretin cardiac amyloidosis (ATTR-CA) is often delayed, worsening outcomes. AI-enhanced ECG detection offers screening and earlier detection opportunities. Traditional AI systems, relying on supervised learning with large labeled datasets, struggle with rare conditions due to limited labeled data. We developed a self-supervised learning (SSL) approach using a foundation model (FM) pretrained on raw, unlabeled ECGs to generate meaningful signal representations, enabling accurate detection even with scarce data. Purpose: We evaluated whether this SSL approach improves ATTR-CA detection compared to conventional supervised learning, aiming to enable earlier diagnosis and better management of this life-threatening condition affecting heart function. Methods: We created an SSL framework using the Barlow Twins method, featuring augmentation, an encoder, and a projector. The FM was trained on 838,496 unlabelled 10-second 12-lead ECGs to minimise the Barlow Twins loss, generating numerical ECG representations. For ATTR-CA detection, we used a balanced dataset of 266 ATTR-CA patients (aged 65±10 years, 60% male, with common comorbidities like hypertension and heart failure) and 266 controls, following established protocols. Two models were trained: a Logistic Regression model (SSL-LR) using FM representations and a supervised model (SUP) trained directly on ECGs. Both used nested 8-fold cross-validation, with results reported as means and standard deviations. Results: The SSL-LR achieved an accuracy of 0.77 (0.75-0.79), AUC of 0.83 (0.80-0.86), significantly outperforming the supervised model (accuracy 0.68 (0.64-0.70), AUC 0.75 (0.7-0.79) ). SSL-LR performance rivals models trained on datasets 10 times larger, enabling reliable CA detection with minimal labeled data. Conclusions: This SSL approach revolutionises AI tool development for rare cardiac disease detection from ECGs, offering earlier ATTR-CA diagnosis to improve patient outcomes and reduce heart failure progression. For engineers, it demonstrates a scalable framework for data-scarce scenarios; for clinicians, it promises a practical tool for routine ECG analysis, seamlessly integrable into existing platforms for real-time screening. Future multicenter trials are needed to validate and implement this in clinical practice.
No takes yet. Share an insight, caveat, or question.
Mcauliffe et al. (2025) studied this question. Self-supervised learning ECG model detected ATTR-CA with AUC 0.83, outperforming supervised learning AUC 0.75, enabling accurate diagnosis with limited labeled data.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: