💻 computer science

One Dimensional CNN ECG Mamba for Multilabel Abnormality Classification in 12 Lead ECG

This paper introduces a hybrid One Dimensional CNN ECG Mamba framework that combines convolutional feature extraction with a bidirectional Mamba state space model to achieve superior multilabel abnormality classification performance on 12-lead ECGs compared to existing deep learning methods.

Original authors: Huawei Jiang, Husna Mutahira, Gan Huang, Mannan Saeed Muhammad

Published 2026-03-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Huawei Jiang, Husna Mutahira, Gan Huang, Mannan Saeed Muhammad

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your heart is a busy orchestra, and an ECG (Electrocardiogram) is the sheet music recording how all the instruments (the heart's electrical signals) are playing together. Doctors need to read this sheet music to spot if a musician is playing out of tune (a heart abnormality).

For a long time, computers tried to read this music using two main tools:

The "Spotter" (CNNs): Good at looking at small, local details, like a single note.
The "Context Reader" (Transformers): Good at understanding the whole song, but they get overwhelmed and slow if the song is too long.

The problem? Heart recordings are long, complex songs. The old tools either missed the big picture or got too slow to be useful in a real hospital.

Enter the New Hero: 1DCNN-ECG-Mamba

This paper introduces a new AI model called 1DCNN-ECG-Mamba. Think of it as a super-intelligent, super-fast music conductor who has a special trick up their sleeve.

Here is how it works, broken down into simple parts:

1. The "Warm-Up" (The 1D CNN)

Before the conductor starts listening to the whole symphony, they need to tune the instruments. The model starts with a 1D Convolutional Neural Network (CNN).

Analogy: Imagine a quick scan of the sheet music to highlight the most important notes and ignore the background noise. It's like a spotlight that instantly finds the "soloists" (key features) in the messy data.

2. The "Mamba" (The Selective State Space)

This is the magic part. The model uses something called Mamba, which is a new type of AI architecture.

The Old Way (Transformers): Imagine trying to remember a 10-minute story by reading every single word and comparing it to every other word in the story. It's like trying to hold a conversation with everyone in a stadium at once. It's powerful but exhausting and slow.
The Mamba Way: Imagine a smart librarian who doesn't read every book on the shelf. Instead, they have a magical memory that knows exactly which books are relevant to your current question and only looks at those. They can process a 10-minute story in the time it takes to read a paragraph, without forgetting the beginning.
Why it matters: Mamba is designed to handle long sequences (like long heartbeats) efficiently. It remembers the start of the heartbeat while listening to the end, without getting tired.

3. The "Two-Way Street" (Bidirectional Vision Mamba)

The researchers took a model originally designed for images (Vision Mamba) and adapted it for heart signals.

The Twist: Most models read a story from left to right (forward). But heart signals have hidden clues that make more sense if you look at them from right to left (backward) too.
Analogy: It's like reading a mystery novel. If you only read from page 1 to 100, you might miss a clue hidden on page 90 that explains page 10. This model reads the story forward and backward simultaneously, ensuring it catches every single clue, no matter where it is in the timeline.

The Big Race: How Did It Do?

The authors tested this new "Conductor" against the best previous champions in two major international competitions (PhysioNet Challenges 2020 and 2021).

The Opponents: They faced off against heavy hitters like ResNet (the old reliable), Transformers (the big memory), and even an earlier version of Mamba called ECG-Mamba.
The Result: The new 1DCNN-ECG-Mamba didn't just win; it crushed the competition.
- It found more heart abnormalities (higher AUPRC score).
- It was more accurate at saying "Yes, this is a problem" or "No, this is fine" (higher AUROC score).

Think of it this way: If the old models were like a detective who finds 50% of the clues, this new model finds 64% of the clues and does it faster.

Why Should You Care?

This isn't just about beating a scoreboard.

Faster Diagnosis: Because Mamba is so efficient, it can analyze heart signals quickly, even on smaller computers.
Telemedicine: Imagine a wearable device in a remote village that can instantly analyze a patient's heart and tell a doctor in a city hospital exactly what's wrong, without needing a supercomputer.
Saving Lives: By catching subtle abnormalities that older models miss, this technology could help diagnose heart issues earlier, leading to better treatment and more lives saved.

In a nutshell: The researchers built a new AI that combines a quick "spotlight" for details with a "smart librarian" memory for long stories, reading heart signals in both directions to find problems faster and more accurately than ever before.

1. Problem Statement

Accurate detection of cardiac abnormalities from 12-lead Electrocardiogram (ECG) recordings is critical for clinical diagnostics. However, existing deep learning approaches face significant challenges:

Long Sequential Dependencies: ECG signals are long time-series data. Traditional Convolutional Neural Networks (CNNs) struggle to capture long-range dependencies, while Transformer architectures suffer from quadratic computational complexity ( $O(n^2)$ ) when processing these long sequences.
Class Imbalance: Medical datasets often suffer from severe class imbalance, making it difficult for models to detect rare but critical abnormalities.
Performance Limits: State-of-the-art methods from the PhysioNet/CinC Challenges (2020 and 2021), such as ResNet-based models and Transformers, have reached performance plateaus in terms of Area Under the Precision-Recall Curve (AUPRC) and Area Under the Receiver Operating Characteristic Curve (AUROC).

2. Methodology: 1DCNN-ECG-Mamba

The authors propose a hybrid framework named 1DCNN-ECG-Mamba, which integrates 1D Convolutional Neural Networks with Mamba, a selective State Space Model (SSM). The architecture is an adaptation of Vision Mamba (Vim), originally designed for visual data, tailored specifically for 12-lead ECG signals.

Key Architectural Components:

1D CNN Feature Extraction:
- A single 1D Convolutional layer is applied at the input to process the raw 12-lead ECG signals (12 channels $\times$ $N$ time samples).
- Configuration: Input size 12, Output size 384, Kernel size 16, Stride 8.
- This layer transforms the multichannel ECG data into a feature map compatible with the Mamba encoder.
Patch Embedding & Tokenization:
- The output of the CNN is divided into patches of shape $(1023, 384)$ .
- A Class Token is appended, resulting in a sequence length of 1024.
- Positional embeddings are added to preserve temporal order.
Bidirectional Vision Mamba (Vim) Encoder:
- The core of the model uses 16 stacked Vim blocks.
- Bidirectional Processing: Unlike standard unidirectional SSMs, this model processes the sequence in both forward and backward directions to capture global temporal dependencies effectively.
- Selective State Space: Utilizes the Mamba mechanism to maintain linear computational complexity ( $O(n)$ ) while retaining a global receptive field, addressing the efficiency issues of Transformers.
Specific Modifications for ECG:
- Stochastic Dropout: Disabled to improve stability.
- Residual Addition: The outputs of forward and backward passes are preserved as-is rather than being divided by two.
- Normalization: The fused residual-addition and normalization operation is disabled; standard PyTorch sequential application (Residual Add $\rightarrow$ Normalization) is used to ensure reproducibility across hardware.
- Learning Rate Schedule: Uses a warm-up phase (5 epochs) followed by cosine annealing over 13 epochs.

Data Preprocessing:

Resampling: All signals are normalized to 500 Hz (downsampling 1000 Hz via polyphase filtering; upsampling 257 Hz via FFT).
Length Normalization: Signals are truncated or zero-padded to a fixed length of 8192 samples.

3. Key Contributions

Novel Architecture: Introduction of 1DCNN-ECG-Mamba, the first model to adapt the bidirectional Vision Mamba architecture specifically for 12-lead ECG multilabel classification.
Performance Benchmarking: The model establishes a new state-of-the-art baseline, significantly outperforming previous winning solutions from the PhysioNet/CinC Challenges (2020 and 2021).
Architectural Validation: Demonstrates that modifying the Vim architecture (e.g., using bidirectional blocks and specific normalization strategies) yields superior results for ECG data compared to the original ECG-Mamba and Audio Mamba variants.

4. Experimental Results

Experiments were conducted on the PhysioNet/CinC Challenge 2021 dataset (88,253 recordings, 26 classes) and the 2020 dataset. The model was evaluated using Macro AUPRC and Macro AUROC.

Performance Comparison (PhysioNet/CinC 2021):

Method	AUPRC	AUROC
ISIBrno (ResNet)	0.5041	0.9036
ISIBrno (w/o Attention)	0.5452	0.8863
2D-CNN	0.5439	0.9348
ECG-Mamba	0.6100	0.9643
1DCNN-ECG-Mamba (Ours)	0.6410	0.9695

Key Findings:

vs. 2021 Best (ISIBrno): 1DCNN-ECG-Mamba improved AUPRC by ~17.8% and AUROC by ~3.7% compared to the 2D-CNN baseline, and significantly outperformed the ResNet-based ISIBrno.
vs. 2020 Best (Prna): On the 2020 dataset, it achieved an AUPRC of 0.5787 and AUROC of 0.9571, surpassing the Prna Transformer model (0.5115 / 0.9318).
vs. Mamba Variants: It outperformed the original Mamba and Audio Mamba, validating the importance of the bidirectional architecture for capturing ECG temporal context.

5. Significance and Impact

Clinical Utility: The high AUPRC scores indicate the model's superior ability to detect rare cardiac abnormalities, which is crucial for early diagnosis and reducing false negatives in clinical settings.
Efficiency: By leveraging Mamba's linear complexity, the model offers a computationally efficient alternative to Transformers for long ECG sequences, making it suitable for real-time applications and resource-constrained environments (e.g., telemedicine).
Generalizability: The framework demonstrates that SSMs can be successfully adapted from visual and audio domains to biomedical signal processing, opening new avenues for time-series analysis in healthcare.

In conclusion, the paper presents a robust, high-performance solution for automated ECG diagnosis, proving that hybrid CNN-Mamba architectures can effectively overcome the limitations of current deep learning models in handling long, complex medical time-series data.