xLSTM-ECG: Multi-label ECG Classification via Feature Fusion with xLSTM

TL;DR

This work targets automated, multi-label classification of 12-lead ECGs to aid cardiovascular diagnostics. It introduces xLSTM-ECG, which couples STFT-based time-frequency features with a dual-module xLSTM (sLSTM for short-term and mLSTM for long-range memory) and layer-wise feature fusion to capture local and global patterns across leads. Evaluations on PTB-XL show state-of-the-art accuracy and AUC, while extended experiments on the Georgia dataset demonstrate robustness across cohorts; ablation and co-occurrence analyses provide insights into model behavior and data labeling. The approach offers a scalable, clinically relevant framework for accurate ECG interpretation and highlights extended-memory architectures as a viable path for complex multi-lead biomedical signals.

Abstract

Cardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, highlighting the critical need for efficient and accurate diagnostic tools. Electrocardiograms (ECGs) are indispensable in diagnosing various heart conditions; however, their manual interpretation is time-consuming and error-prone. In this paper, we propose xLSTM-ECG, a novel approach that leverages an extended Long Short-Term Memory (xLSTM) network for multi-label classification of ECG signals, using the PTB-XL dataset. To the best of our knowledge, this work represents the first design and application of xLSTM modules specifically adapted for multi-label ECG classification. Our method employs a Short-Time Fourier Transform (STFT) to convert time-series ECG waveforms into the frequency domain, thereby enhancing feature extraction. The xLSTM architecture is specifically tailored to address the complexities of 12-lead ECG recordings by capturing both local and global signal features. Comprehensive experiments on the PTB-XL dataset reveal that our model achieves strong multi-label classification performance, while additional tests on the Georgia 12-Lead dataset underscore its robustness and efficiency. This approach significantly improves ECG classification accuracy, thereby advancing clinical diagnostics and patient care. The code will be publicly available upon acceptance.

Figures (7)

• Figure 1: The architecture of the proposed xLSTM-ECG model begins with the input of 12-lead ECG waveforms $x_i$. After pre-processing, these waveforms are transformed into frequency domain features $\hat{x}_i$. These features are then fed into the layer fusion xLSTM blocks $\mathcal{M}$, which consist of the sLSTM (teal color) and mLSTM (lavender color) modules. The outputs from both modules $\hat{x}_{i}^{s}$ and $\hat{x}_{i}^{m}$ are merged to create a fusion feature $\hat{x}_{i}^{f}$. After passing through two blocks as shown in dashed rectangles, the final fusion feature undergoes average pooling to obtain a fixed-size vector $v_{i}$. This vector is then projected into probabilities for five different superclasses using five independent linear layers. Positive labels are indicated in Green, while negative labels are indicated in Red.
• Figure 2: Sequential fusion strategy: features are integrated at the final layer of both sLSTM and mLSTM modules.
• Figure 3: Layer fusion strategy: features are integrated during each layer of both sLSTM and mLSTM modules.
• Figure 4: Confusion matrices corresponding to the five categories: NORM, CD, HYP, MI, and STTC.
• Figure 5: Label co-occurrence matrices for multi-class, multi-label classification: (a) shows the ground-truth co-occurrence matrix, and (b) shows the predicted co-occurrence matrix.