TLDR: A new post-training strategy significantly improves the performance of ECG foundation models like ECGFounder. By incorporating stochastic depth to reduce signal redundancy and preview linear probing for better classification head initialization, the method bridges performance gaps compared to task-specific models. Experiments on the PTB-XL benchmark show substantial gains in AUROC and AUPRC, particularly in data-scarce scenarios, making these models more stable and clinically applicable.
Electrocardiography (ECG) is a fundamental, non-invasive tool used globally for screening, diagnosing, monitoring, and predicting risks associated with cardiovascular diseases. It captures the heart’s electrical activity, providing crucial insights into its function and overall health. In recent years, artificial intelligence, particularly deep learning, has significantly advanced ECG analysis, with deep neural networks demonstrating expert-level accuracy in detecting various cardiac abnormalities.
Foundation models have emerged as a powerful approach in AI, offering generalizable and transferable models trained on vast datasets that can be adapted for diverse tasks. In the ECG field, models like ECGFounder have shown broad applicability. However, despite their potential, these foundation models often exhibit performance gaps when fine-tuned for specific diagnostic tasks, especially when compared to specialized, task-specific models. This limitation has raised concerns about their practical use in real-world clinical settings.
To address this challenge, researchers have proposed a new post-training strategy designed to enhance ECGFounder, a leading ECG foundation model. This strategy aims to bridge the performance gap that persists even after extensive pre-training on millions of ECG recordings and subsequent fine-tuning on target data. The core idea is that an effective post-training approach can significantly improve the model’s adaptability and accuracy.
The proposed strategy is built on two key insights. First, ECG signals inherently contain redundant information, meaning neighboring time points and heartbeat cycles are often predictable from each other. To tackle this, the researchers incorporated ‘stochastic depth,’ a technique that reduces redundancy and makes the model more robust. Second, while pre-training provides a valuable starting point, the final classification layer of the model is typically initialized randomly. To optimize this, ‘preview linear probing’ is introduced during post-training to better prepare this classification layer, thereby boosting performance for specific tasks.
Experiments conducted on the PTB-XL benchmark, a comprehensive public dataset for ECG analysis, demonstrated the effectiveness of this new approach. The post-training strategy, referred to as ECGFounder-PT, significantly improved upon the baseline fine-tuning strategy. For instance, it showed improvements of 1.2%–3.3% in macro AUROC (Area Under the Receiver Operating Characteristic Curve) and 5.3%–20.9% in macro AUPRC (Area Under the Precision-Recall Curve) across various classification tasks, including all-71 labels, rhythm-12, diagnostic-44, and subclass-23.
Furthermore, ECGFounder-PT not only outperformed the original ECGFounder and another foundation model, HuBERT-ECG-BASE, but also surpassed several recent state-of-the-art task-specific and advanced models. This highlights the potential of ECG foundation models when equipped with an effective post-training strategy to achieve balanced generalization across diverse ECG classification tasks.
A notable finding was the strategy’s stability and efficiency, particularly in scenarios with limited training data. When using only 10% of the available training data, the proposed method achieved a 9.1% improvement in macro AUROC and a remarkable 34.9% improvement in macro AUPRC. This makes the strategy highly valuable for clinical practice where access to large, labeled datasets can be scarce.
An ablation study confirmed the critical contributions of stochastic depth and preview linear probing to the enhanced performance. These components were identified as the most influential factors, with their removal leading to substantial performance degradation. The study also showed that the proposed post-training strategy, even with random initialization, could achieve performance comparable to the baseline strategy that leveraged extensive pre-training.
Also Read:
- Bridging ECG and AI: Introducing ECG-aBcDe for Enhanced Medical Diagnosis
- Guiding Acoustic Scene Classification with Entropy for Better Generalization
In conclusion, this research underscores the importance of post-training strategies in enhancing the clinical applicability of ECG foundation models. By addressing existing performance gaps, this work paves the way for more robust and adaptable AI tools in cardiovascular diagnostics. For more details, you can refer to the full research paper here.
