Skip to content

EXPERIMENTS.md

Latest commit

 

History

History
163 lines (109 loc) · 3.9 KB

EXPERIMENTS.md

File metadata and controls

163 lines (109 loc) · 3.9 KB

Reproducing Experiments

This document provides exact commands to reproduce the results reported in our paper.

Hardware Requirements

Component Minimum Recommended
GPU NVIDIA RTX 3080 (10GB VRAM) NVIDIA RTX 3090 / A100 (24GB+)
RAM 32 GB 64 GB
Storage 50 GB (for datasets + cache) 100 GB
CUDA 11.8 12.x

Note: Training with batch_size=192 requires ~20GB VRAM. Reduce to 64-128 for smaller GPUs.

Dataset Preparation

  1. Download datasets from PhysioNet (see data/README.md)
  2. Extract to the following structure: 
    data/
├── chapman/          # ~45k records
├── cpsc2021/         # For zero-shot AF
└── ptbxl/            # For zero-shot multi-class

Interactive Experiments (Web App)

For real-time inference and clinical visualization, we provide a full-stack web application.

# Start the Clinical Dashboard
cd web_app
run_app.bat

See web_app/README.md for detailed usage.

Table 1: Chapman-Shaoxing 5-Fold Cross-Validation

Expected Results: Macro F1 ≈ 0.31, ROC-AUC ≈ 0.85

Training

# Full training (5 folds, 20 epochs each)
python scripts/train.py

Output:

  • Checkpoints: models/fold{1-5}_best.pt
  • Logs: reports/logs/training_log_epochs.csv

Evaluation (OOF)

python scripts/eval_oof.py

Expected Output:

f1_macro           0.3115 ± 0.0132
precision_macro    0.3421 ± 0.0198
recall_macro       0.3012 ± 0.0156
auprc_macro        0.2845 ± 0.0187

Table 2: Zero-Shot Transfer (CPSC-2021)

Expected Results: PR-AUC (AF) ≈ 0.708

Prerequisites

  1. Complete Table 1 training (need fold*_best.pt)
  2. Prepare CPSC-2021 dataset in data/cpsc2021/

Evaluation

python scripts/eval_zeroshot.py

Table 3: Ablation Study

Ablations are controlled via the ablation dict passed to ECGRambaV7Advanced:

from src.model import ECGRambaV7Advanced
from configs.config import CONFIG

# Full model (default)
model_full = ECGRambaV7Advanced(CONFIG, ablation={})

# Without MiniRocket
model_no_rocket = ECGRambaV7Advanced(CONFIG, ablation={"no_rocket": True})

# Without HRV
model_no_hrv = ECGRambaV7Advanced(CONFIG, ablation={"no_hrv": True})

# Without Multi-Scale Tokenizer
model_no_multiscale = ECGRambaV7Advanced(CONFIG, ablation={"no_multiscale": True})

# Without Cross-Attention Fusion
model_no_fusion = ECGRambaV7Advanced(CONFIG, ablation={"no_fusion": True})

To run ablation experiments, modify scripts/train.py to pass the desired ablation config.

Inference-Time Ablation

For quick ablation during inference (without retraining):

# Disable MiniRocket features at inference
logits = model(x, xh, xhr, use_rocket=False)

# Disable HRV features at inference
logits = model(x, xh, xhr, use_hrv=False)

# Disable Cross-Attention Fusion (use simple concat)
logits = model(x, xh, xhr, use_fusion=False)

Troubleshooting

Issue Solution
CUDA out of memory Reduce batch_size in configs/config.py
mamba-ssm installation fails Use pre-built wheels or install from source
Different results Ensure numpy==1.26.4 and same CUDA version

Citation

If you use these experiments, please cite:

@article{nguyen2025ecg,
  title={ECG-RAMBA: Zero-Shot ECG Generalization...},
  author={Nguyen, Hai Duong and Tran, Xuan-The},
  journal={arXiv preprint arXiv:2512.23347},
  year={2025}
}