TLDR: SelfMIS is a new framework that significantly improves myocardial infarction (heart attack) detection from single-lead ECGs. It addresses limitations of previous methods by directly aligning single-lead and multiple-lead ECG representations in the latent space using a ‘self-cutting’ strategy. This approach enriches single-lead representations, enabling better inference of global cardiac context from local signals. SelfMIS achieves superior diagnostic performance, maintains a simpler architecture, and is more computationally efficient, especially robust in low-data environments, making it promising for mobile health applications.
Myocardial infarction, commonly known as a heart attack, remains a leading cause of death globally. Electrocardiograms (ECGs) are standard tools for diagnosis, but traditional methods often rely on complex twelve-lead ECGs, which require specialized equipment and trained personnel, limiting their use in urgent, out-of-hospital settings. The rise of mobile devices capable of recording single-lead ECGs, like the Apple Watch, offers a promising solution for early screening. However, accurately detecting heart attacks from a single-lead ECG is challenging due to its limited view of the heart’s electrical activity.
Current approaches to enhance single-lead ECG diagnosis often involve converting single-lead signals into multiple-lead ones using generative models. While these models can produce visually similar signals, a significant problem arises: the generated signals often have a large ‘gap’ in their underlying data representation (latent space) compared to real multiple-lead ECGs. This mismatch can severely degrade diagnostic performance. Other methods focus on ‘alignment learning’ to make models robust to signal variations, but this doesn’t directly address the core challenge of inferring a global cardiac state from a limited local signal.
To overcome these limitations, researchers have introduced a novel framework called SelfMIS (Self-alignment learning to improve Myocardial Infarction detection from Single-lead ECG). This approach offers a simple yet highly effective way to improve heart attack detection from single-lead ECGs. Instead of relying on complex data augmentations or signal-level generation, SelfMIS directly aligns single-lead ECGs with their corresponding multiple-lead segments in the latent space.
How SelfMIS Works
The core idea behind SelfMIS is straightforward. It uses a ‘self-cutting’ strategy to create pairs of single-lead and multiple-lead ECGs from the same recording. For instance, it takes a full multiple-lead ECG and extracts the Lead I signal to form a ‘positive pair’. These pairs are then fed into two separate encoders: one for single-lead and one for multiple-lead ECGs. Crucially, during the training process, the multiple-lead ECG encoder is ‘frozen’, meaning its parameters are not updated. This forces the single-lead ECG encoder to learn representations that align with the already discriminative feature space of the multiple-lead encoder.
This design shifts the learning objective from simply making the model robust to variations to enriching the single-lead representation. Essentially, the single-lead ECG encoder is trained to infer the broader cardiac context from its limited local signal, leading to more informative embeddings that significantly boost detection performance.
Key Advantages and Results
SelfMIS demonstrates several significant advantages:
- Superior Performance: Experiments show that SelfMIS consistently outperforms existing discriminative and generative methods across nine different myocardial infarction types. It achieves an overall average macro AUC (a key performance metric) exceeding 80.0, a level not reached by other methods.
- Simpler Architecture and Efficiency: Despite its superior performance, SelfMIS maintains a simpler architecture and lower computational overhead compared to many other models. It uses fewer parameters and significantly less GPU memory during pre-training, making it more practical for real-world applications.
- Robustness in Low-Data Settings: The framework shows strong generalization capabilities even with limited training data. SelfMIS performs remarkably well with only 25% or 50% of the training data, often matching or surpassing the performance of other models trained on 100% of the data.
- Augmentation-Free Paradigm: By using a self-cutting strategy, SelfMIS avoids manual data augmentations, which can sometimes distort the semantic information of ECG signals. This ensures semantic consistency between single-lead and multiple-lead representations.
The research paper, available at https://arxiv.org/pdf/2509.19397, highlights that the alignment mechanism is the most critical component of SelfMIS. The ability to transfer diagnostic knowledge from a data-rich modality (multiple-lead ECG) to a data-sparse one (single-lead ECG) is key to its success. Even when trained from scratch without pre-trained weights, SelfMIS still achieves respectable diagnostic performance, underscoring the effectiveness of enforcing representational consistency.
Also Read:
- PPGFlowECG: Transforming PPG Signals into Diagnostic-Grade ECG
- DACL: Enhancing Biosignal Analysis with Diffusion-Powered Contrastive Learning
Conclusion
SelfMIS represents a significant step forward in myocardial infarction detection from single-lead ECGs. Its innovative direct latent space alignment, coupled with a simple and efficient design, offers a powerful tool for improving cardiac diagnostics. This approach holds immense promise for advancing ECG diagnostic performance, particularly for mobile devices, making timely heart attack detection more accessible and widespread.
