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

AI summary

Overview

• Research area: Automated medical signal analysis — specifically multilabel classification of cardiac abnormalities from 12-lead electrocardiogram (ECG) recordings, positioned by arXiv under Computer Vision (cs.CV), paper ID arXiv:2510.13046v2, CC BY 4.0.
• Technical level: Intermediate (requires familiarity with deep learning sequence models; the paper does not provide introductory background on ECG physiology or state space models beyond its own preliminaries).
• Scope: The paper proposes and benchmarks a single hybrid architecture, 1DCNN-ECG-Mamba, against previously published challenge-winning methods on the PhysioNet/Computing in Cardiology (CinC) Challenge 2020 and 2021 datasets.

What This Paper Is About

Detecting heart abnormalities from ECG signals is important clinically, but existing deep learning models — residual networks and Transformers — struggle when they must process long sequential recordings, because Transformer attention scales quadratically with sequence length. The authors ask whether Mamba, a selective state space model that processes sequences in linear time, can be adapted to 12-lead ECG data to beat the best published results. Their answer is a hybrid pipeline that uses a one-dimensional convolutional layer to turn the raw 12-lead signal into patch-like tokens, then feeds those tokens through a bidirectional Vision Mamba (Vim) encoder for multilabel classification.

Key Contributions

1. A new architecture, 1DCNN-ECG-Mamba, which adapts the bidirectional Vision Mamba (Vim) encoder to 12-lead ECG signals, using a single 1D convolutional layer as the front end. The authors describe this as establishing a strong baseline for heart abnormality detection.
2. Specific architectural modifications to Vim for ECG data: stochastic dropout is disabled; forward and backward outputs are preserved as-is rather than divided by two; and the fused residual-addition-plus-normalization kernel (fused_add_norm=False) is replaced with the standard PyTorch sequential formulation for reproducibility across hardware.
3. Empirical demonstration that the model outperforms state-of-the-art methods reported in the PhysioNet/CinC Challenges of 2020 and 2021 in terms of macro AUPRC and macro AUROC.
4. Ablation-style comparisons against the original Mamba, Audio Mamba (AuM), ECG-Mamba, a 2D-CNN, and the challenge-winning ISIBrno and Prna entries, isolating the effect of the bidirectional blocks and the number of convolutional layers (one, versus the two used by ECG-Mamba).

Main Findings

• PhysioNet/CinC 2021 results: 1DCNN-ECG-Mamba reached an AUPRC of 0.6410 and an AUROC of 0.9695.
• Comparison to the 2021 challenge winner: Team ISIBrno (ResNet-based) is listed at 0.5230 AUPRC and 0.8960 AUROC with a mixture loss function; the random-lead variant is listed at 0.5041 and 0.9036. Restricting to 12 leads produced what the authors report as an approximately 3.7% increase in AUPRC and a 0.84% decrease in AUROC.
• Attention removal: ISIBrno without attention reached 0.5452 AUPRC and 0.8863 AUROC — a reported 4.2% AUPRC gain and 1.3% AUROC loss relative to the 12-lead ISIBrno configuration.
• Versus 2D-CNN: The 2D-CNN scored 0.5439 AUPRC and 0.9348 AUROC. The proposed model is reported to increase AUPRC by 17.8% and AUROC by 3.7% over it.
• Versus other Mamba variants on 2021: The original Mamba reached 0.6312 AUPRC / 0.9678 AUROC, Audio Mamba (AuM) 0.6301 / 0.9671, and ECG-Mamba 0.6100 / 0.9643. The proposed model exceeded all three.
• Bidirectional blocks matter: The authors attribute the gain over Audio Mamba to the inclusion of bidirectional blocks, noting that Audio Mamba removes the backward convolution from Vim.
• PhysioNet/CinC 2020 results: The top 2020 method, team Prna (Transformer-based), scored 0.5115 AUPRC and 0.9318 AUROC. 1DCNN-ECG-Mamba scored 0.5787 and 0.9571; ECG-Mamba scored 0.5452 and 0.9497 on the same dataset.
• Note on internal consistency: The paper's prose rounds the ISIBrno baseline to 0.5040/0.9036 while Table III lists 0.5041/0.9036 for "ISIBrno (Random-lead)" and 0.5230/0.8960 for "ISIBrno." The related-work section also cites a 2% AUPRC gain from removing attention, while the results section reports 4.2%. The dataset totals are likewise inconsistent: Table II and the abstract state 88,253 combined recordings, while the dataset-description text states the 2021 dataset comprises 43,101 recordings. These discrepancies are not addressed in the paper.

Methodology in Plain English

Data. The experiments draw on the PhysioNet/CinC Challenge 2020 and 2021 datasets. The paper states these comprise 88,253 ECG recordings pooled from seven institutions across four countries and three continents; elsewhere it states the 2021 dataset contains 43,101 recordings and that the 2020 dataset is contained within it. The 2021 dataset defines 26 categories (25 cardiac abnormalities plus sinus rhythm); the 2020 dataset defines 27 (26 abnormalities plus sinus rhythm). Of the 2021 recordings, 87,663 are at 500 Hz, 516 at 1000 Hz, and 74 at 257 Hz. Table II lists the contributing sources: CPSC and CPSC-Extra (10,330 files, 6–144 sec, 500 Hz), St Petersburg INCART 12-lead (74 files, 1800 sec, 257 Hz), PTB and PTB-XL (22,353 files, 10–120 sec, 500 or 1000 Hz), Georgia (10,344 files, 5–10 sec, 500 Hz), Ningbo First Hospital (34,905 files, 10 sec, 500 Hz), and Chapman University / Shaoxing People's Hospital (10,247 files, 10 sec, 500 Hz).

Preprocessing. All recordings are standardized to 500 Hz. Signals at 1000 Hz are downsampled with polyphase filtering to limit aliasing; signals at 257 Hz are upsampled via the Fast Fourier Transform. Each signal is then right zero-padded or truncated to a fixed 8192 samples, with random 8192-sample segments taken from longer records.

Architecture. A raw 12 × N ECG tensor enters a single Conv1d layer that maps 12 input channels to 384 output channels using kernel size 16, stride 8, and no padding. The output is split into patches of shape (1023, 384); a class token is appended, giving a (1024, 384) sequence, and positional embeddings are added. This sequence passes through 16 Vim encoder blocks, each containing selective state space layers sandwiched between linear projections (one expanding the feature dimension, one reducing it), plus position-wise MLP layers. The class token output serves the classification head. Best performance came with 16 blocks.

Training. Everything runs in PyTorch on an NVIDIA GeForce RTX 3090 Ti with 24 GB of memory, NVIDIA driver 535, CUDA 11.8, Ubuntu 22.04. The optimizer is Adam with initial learning rate 6 × 10⁻⁴, β₁ = 0.9, β₂ = 0.98, ε = 10⁻⁹. The schedule warms up linearly over 5 epochs from 1 × 10⁻⁵ to the peak, then decays along a cosine curve over 13 epochs to a minimum of 1 × 10⁻⁶. Most experiments use binary cross entropy loss; the ISIBrno baselines use a mixture loss.

Evaluation protocol. The 2021 data is split 80/20 into training and testing, stratified by the 26 classes with random shuffling, and evaluated with five-fold cross-validation. Each fold uses 70,602 training and 17,651 testing records. The 2020 data is split 80/20 as well — 34,481 training and 8,620 test records — again with stratified five-fold cross-validation averaging. Performance is measured with macro AUPRC and macro AUROC, chosen because they treat all classes equally and are suited to imbalanced multilabel problems.

Why This Matters

Research impact. The paper argues that linear-complexity selective state space models are a viable alternative to Transformers and CNNs for long physiological time series, and it sets reference numbers on two widely used public benchmarks. It reproduces the structure of its baselines closely (aligning hyperparameters with team Prna for 2020 and team ISIBrno for 2021) so comparisons are as direct as the authors can make them.

Real-world applications

• Remote and telemedicine ECG interpretation, where the paper explicitly claims enhanced accessibility.
• Wearable ECG devices and continuous remote monitoring, which the paper cites as generators of the large cardiac datasets motivating automated analysis.
• Resource-constrained healthcare settings where access to cardiology specialists is limited; automated analysis is described as reducing clinician workload and inter-observer variability.
• Early diagnosis and personalized treatment pathways, which the abstract names as the downstream benefit of reliable ECG classification.

Industry relevance. A model that classifies 12-lead ECG across 25+ abnormality categories at linear sequence cost is directly relevant to device makers, remote-monitoring vendors, and clinical decision-support software. The paper's explicit choice to sacrifice Vim's GPU fused-kernel optimization for standard PyTorch operations is framed as a portability and reproducibility tradeoff — relevant for deployment across heterogeneous clinical hardware, though no inference speed or latency measurements are reported.

Future Directions

• Multi-lead generalization. The 2021 challenge covers 12-, 6-, 4-, 3-, and 2-lead configurations, and random-lead training was present in the dataset, but this study deliberately evaluated only 12-lead data. Extending to reduced-lead formats is the most obvious next step and is not reported here.
• Resolving the internal inconsistencies. Reconciling the 88,253 versus 43,101 recording counts, the 2% versus 4.2% attention-removal AUPRC figures, and the differing ISIBrno numbers would strengthen the reproducibility claim the authors make for their architectural choices.
• External and prospective validation. All results come from PhysioNet/CinC datasets. Whether the AUPRC and AUROC figures hold on other populations, acquisition hardware, or prospective clinical cohorts is not reported.
• Efficiency and interpretability. The paper claims linear-time processing but reports no throughput, parameter-count, or memory comparisons against the Transformer and ResNet baselines, and no attention or saliency analysis showing which ECG regions drive each class prediction. Both are open questions.

Target Audience

Researchers and engineers working on deep learning for biomedical time series — particularly those already familiar with Transformers, state space models, or the Mamba family — will get the most from this paper. It is also useful for clinicians and clinical informatics teams evaluating whether Mamba-based architectures are mature enough for ECG decision support, and for benchmark-focused researchers who need the specific AUPRC and AUROC reference numbers for the PhysioNet/CinC 2020 and 2021 tasks. Readers without prior exposure to state space models or ECG lead conventions will need supporting material, since the paper's preliminaries are brief.

Authors’ abstract

Accurate detection of cardiac abnormalities from electrocardiogram recordings is regarded as essential for clinical diagnostics and decision support. Traditional deep learning models such as residual networks and transformer architectures have been applied successfully to this task, but their performance has been limited when long sequential signals are processed. Recently, state space models have been introduced as an efficient alternative. In this study, a hybrid framework named One Dimensional Convolutional Neural Network Electrocardiogram Mamba is introduced, in which convolutional feature extraction is combined with Mamba, a selective state space model designed for effective sequence modeling. The model is built upon Vision Mamba, a bidirectional variant through which the representation of temporal dependencies in electrocardiogram data is enhanced. Comprehensive experiments on the PhysioNet Computing in Cardiology Challenges of 2020 and 2021 were conducted, and superior performance compared with existing methods was achieved. Specifically, the proposed model achieved substantially higher AUPRC and AUROC scores than those reported by the best previously published algorithms on twelve lead electrocardiograms. These results demonstrate the potential of Mamba-based architectures to advance reliable ECG classification. This capability supports early diagnosis and personalized treatment, while enhancing accessibility in telemedicine and resource-constrained healthcare systems.