Technical Summary: HexagonalWarriorMamba (HWMamba)

Problem Statement

The accurate automated diagnosis of cardiac abnormalities from 12-lead electrocardiograms (ECGs) is critical for managing cardiovascular disease. However, detecting concurrent conditions (multi-label classification) remains a significant challenge for traditional deep learning models. Conventional approaches, such as 1D Convolutional Neural Networks (CNNs) and Transformers, often struggle to model the long-range dependencies inherent in ECG signals. Furthermore, while recent State Space Models (SSMs) like Mamba have shown promise in handling long sequences, their application to 12-lead ECGs has primarily focused on 1D time-series inputs. A key limitation in current research is the disparity between threshold-independent metrics (e.g., AUROC) and threshold-dependent metrics (e.g., Challenge Score, Subset Accuracy), which are more reflective of real-world clinical utility where specific decision thresholds determine diagnosis.

Methodology

The paper proposes HexagonalWarriorMamba (HWMamba), a novel framework that reimagines 12-lead ECG processing by treating the signal not as a 1D time series, but as a single-channel 2D image. This transformation allows the model to leverage established 2D image analysis architectures to capture both inter-lead (spatial) and intra-lead (temporal) relationships.

Core Architecture

• Input Representation: The 12-lead ECG is structured as a rectangular 2D image with dimensions corresponding to 12 leads (spatial) and 8192 samples (temporal). Notably, the authors intentionally omit standard normalization (e.g., Z-score) to preserve amplitude-related diagnostic information.
• Hierarchical Design: Inspired by Swin Transformer and VMamba, HWMamba employs a hierarchical structure with progressive downsampling. The architecture transitions from high-resolution, shallow feature maps to lower-resolution, deeper feature maps.
• Specialized Components:
  • Initial Convolution: A rectangular kernel (3×16) with a stride of (1, 16) is used to accommodate the asymmetric nature of the ECG data (short spatial dimension, long temporal dimension).
  • SS2D Mechanism: The core processing unit utilizes a 2D Selective Scan (SS2D) mechanism derived from VMamba. This scans the 2D feature map in four directions (left-to-right, right-to-left, top-to-bottom, bottom-to-top) to model global context and spatial dependencies efficiently.
  • Gated MLP: The architecture incorporates a Gated Multi-Layer Perceptron (MLP) with a 4x expansion factor. This gating mechanism facilitates adaptive feature selection and contributes to more stable training compared to standard MLPs.
  • Downsampling: Empirically determined to use a stride of (1, 2) to effectively reduce the temporal dimension while preserving critical inter-lead spatial information.

Experimental Setup

• Dataset: The model was evaluated on the PhysioNet/Computing in Cardiology (CinC) Challenge 2021 dataset, comprising 88,253 multi-lead ECG recordings from seven international institutions. The dataset includes 26 diagnostic labels (25 abnormalities plus sinus rhythm).
• Training Strategy: The study utilized a 5-fold cross-validation strategy with an 80/20 train-test split. Thresholds for evaluation were optimized exclusively on the training set to prevent data leakage.
• Optimization: The model was trained using the Adam optimizer with a cosine annealing scheduler and a warmup phase.

Key Contributions

1. Benchmarking and Performance: The study establishes a new reference point for multi-label ECG classification. HWMamba outperforms current State-of-the-Art (SOTA) methods across five threshold-dependent metrics, including Challenge Score and Subset Accuracy. The authors term this consistent performance across multiple dimensions "Hexagonal Warrior" performance.
2. Generalization and Threshold Stability: The paper highlights that strong performance on threshold-independent metrics (like AUROC) does not guarantee practical effectiveness. HWMamba demonstrates a small performance gap between thresholds derived from training data versus testing data, indicating robust generalization and stability in real-world settings.
3. Architectural Innovation: By converting 12-lead ECGs into 2D images and applying the SS2D mechanism, the model successfully models complex spatial relationships between leads, addressing limitations found in pure 1D sequence models.

Results

Evaluated against SOTA methods (including ISIBrno, 2D-CNN, 2DRU+LC, ECG-Mamba, and 1D-Mamba), HWMamba achieved the following:

• Threshold-Dependent Metrics: HWMamba achieved the highest scores in Subset Accuracy (59.51%), Challenge Score (0.7294), Macro F1-score (0.6156), and Weighted F1-score (0.7730). It also achieved the lowest Hamming Loss (0.0250).
• Threshold-Independent Metrics: HWMamba maintained near-SOTA performance in Macro AUROC (0.9698), trailing only slightly behind the 2DRU+LC method (0.9729).
• Per-Class Performance: The model consistently outperformed baselines across most cardiac abnormality classes, showing particular resilience in challenging categories where other models struggled.
• Ablation Studies: The results indicated that:
  • Normalization: Applying Z-score normalization degraded performance, confirming that raw amplitude contains vital diagnostic information.
  • Architecture: The Gated MLP configuration was essential for achieving the highest Challenge Score and F1 scores, despite standard MLPs yielding slightly higher subset accuracy in isolation.
  • Stride: A (1, 2) downsampling stride was found optimal compared to symmetric or other asymmetric strides.

Significance and Claims

The paper claims that HWMamba represents a robust and versatile approach for multi-label ECG classification. Its significance lies in bridging the gap between high discriminative capability and effective threshold selection derived from training data. Unlike some existing methods (e.g., 2DRU+LC) that show large performance gaps when moving from training-optimized thresholds to testing-optimized thresholds, HWMamba demonstrates strong threshold stability.

The authors argue that this "Hexagonal Warrior" performance profile—balancing strong results across all six evaluation dimensions—positions the model as a reliable tool for clinical deployment. The study suggests that the combination of 2D spatial modeling via SS2D and the preservation of raw signal amplitude allows the model to capture complex cardiac dependencies that 1D models miss, while avoiding the generalization pitfalls of architectures that rely heavily on threshold-independent metrics. The paper concludes that future work could explore combining the structural advantages of U-Net architectures with Mamba-based sequence modeling to further improve threshold robustness.