ECG-Mamba-V2

ECG-Mamba-V2: Architectural Refinements to a Bidirectional State Space Model for Multi-Label 12-Lead ECG Classification

This paper has been accepted as a letter in Frontiers of Computer Science (FCS).

This repository contains the official code for the paper.

---

Overview

ECG-Mamba-V2 is a bidirectional state space model for multi-label classification of 12-lead ECG recordings. It refines the ECG-Mamba architecture with:

• bidirectional depthwise convolution in addition to the bidirectional SSM scan,
• the class token placed at the end of the token sequence,
• removal of the output scaling factor,
• uniform dropout (rate 0.1 in every block),
• a cosine annealing schedule with linear warm-up.

Results

Macro-averaged scores, mean over 15 runs (3 seeds x 5 folds), patient-grouped 5-fold cross-validation.

Dataset	Model	AUPRC	AUROC
PhysioNet/CinC 2021	ECG-Mamba	0.6083	0.9643
PhysioNet/CinC 2021	ECG-Mamba-V2	0.6494	0.9716
PhysioNet/CinC 2020	ECG-Mamba	0.5554	0.9524
PhysioNet/CinC 2020	ECG-Mamba-V2	0.5681	0.9561

Complexity: 17.17 M parameters, 17.49 GMac, 252 samples/s (batch size 20, RTX 3090 Ti).

Requirements

Needs a CUDA GPU. mamba-ssm and causal-conv1d must be installed for the fast path.

Data

Download the PhysioNet/CinC Challenge 2021 (and 2020) training data from https://physionet.org/content/challenge-2021/, then build the cross-validation folds:

Recordings that share the proxy patient key (source, age, sex, label set) are kept in the same fold to avoid leakage.

Training

bash ECG_scenario2021.sh (from scripts)

Repeat for folds 0-4 and seeds 0, 1, 2, then average.

Complexity profiling

python compute_flops.py

Citation

@article{ecgmambav2,
  title   = {ECG-Mamba-V2: Architectural Refinements to a Bidirectional State Space Model for Multi-Label 12-Lead ECG Classification},
  author  = {<authors>},
  journal = {Frontiers of Computer Science},
  note    = {Letter, accepted},
  year    = {2026}
}

Acknowledgement

The implementation builds on Vision Mamba (Vim) and Mamba.

License

<MIT / Apache-2.0 — choose one>