Morphology-Aware ECG Reconstruction Framework for Missing Precordial Leads in 12-Lead ECGs

Official implementation of "Morphology-Aware ECG Reconstruction Framework for Missing Precordial Leads in 12-Lead ECGs".

---

Table of Contents

• Environments
  • Requirements
• Installation
  • Using Docker
  • Using Conda
• Data Preparation
• How to Run
• Experiments
  • Hyperparameters
  • Results
• Citation
• Contact

---

Environments

• Docker: 27.0.1
• CPU: 128-core Intel(R) Xeon(R) Gold 6430
• GPU: NVIDIA GeForce RTX 4090 (24GB)
• NVIDIA Driver: 560.35.03

Requirements

• Python == 3.11
• PyTorch == 2.6.0

For a complete list of dependencies, please refer to requirements.txt.

---

Installation

Using Docker

We provide a Dockerfile to build the image.
Simply run the following command to run the container:

bash build.sh

Using Conda

conda create -n morphrec python=3.11
conda activate morphrec
conda install pytorch==2.6.0 cudatoolkit=12.6 -c pytorch
git clone <repo-url>
cd MorphRec
pip install -r requirements.txt

How to Run

python main.py configs/config.yaml

---

Data Preparation

Dataset Download Instructions

We provide links to publicly available ECG datasets used in our experiments. Please download the raw ECG signal files from the following sources:

• PTB-XL
  https://physionet.org/content/ptb-xl/1.0.3/

• Chapman
  https://physionet.org/content/challenge-2021/1.0.3/training/chapman_shaoxing/#files-panel

Data Preprocessing

Each ECG signal must be saved as a .pkl file with the following dictionary format:

{'I': array([-0.02 , -0.03 , -0.035, ..., -0.02 , -0.035, -0.04 ]),
 'II': array([-0.39 , -0.4  , -0.405, ..., -0.06 , -0.065, -0.07 ]),
 'III': array([-0.37, -0.37, -0.37, ..., -0.04, -0.03, -0.03]),
 'AVR': array([0.205, 0.215, 0.22 , ..., 0.04 , 0.05 , 0.055]),
 'AVF': array([-0.38 , -0.385, -0.385, ..., -0.05 , -0.045, -0.05 ]),
 'AVL': array([ 0.17 ,  0.165,  0.165, ...,  0.005, -0.005, -0.01 ]),
 'V1': array([0.14 , 0.14 , 0.14 , ..., 0.04 , 0.04 , 0.045]),
 'V2': array([0.185, 0.18 , 0.165, ..., 0.055, 0.045, 0.045]),
 'V3': array([ 0.13 ,  0.14 ,  0.13 , ..., -0.06 , -0.065, -0.07 ]),
 'V4': array([ 0.09 ,  0.08 ,  0.065, ..., -0.06 , -0.065, -0.07 ]),
 'V5': array([ 0.17 ,  0.16 ,  0.14 , ..., -0.005, -0.01 , -0.01 ]),
 'V6': array([ 0.155,  0.14 ,  0.135, ..., -0.065, -0.065, -0.065])}

⚠️ Important Note

Please ensure that all signals are converted to millivolt (mV) scale, not microvolt (µV) scale, before processing.

Data Organization

Please download index of dataset from here, and place files as follows:

assets/
configs
src/
dataset/
├── 📁 index/                          # Pickled index files per dataset
│   ├── 📁 chapman/
│   │   ├── 📁 0/                      # Split 0
│   │   │   ├── 📄 train_df.pkl
│   │   │   ├── 📄 valid_df.pkl
│   │   │   └── 📄 test_df.pkl
│   │   ├── 📁 1/                      # Split 1
│   │   │   ├── 📄 train_df.pkl
│   │   │   ├── 📄 valid_df.pkl
│   │   │   └── 📄 test_df.pkl
│   │   ├── 📁 2/                      # Split 2
│   │   │   ├── 📄 train_df.pkl
│   │   │   ├── 📄 valid_df.pkl
│   │   │   └── 📄 test_df.pkl
│   │   ├── 📁 3/                      # Split 3
│   │   │   ├── 📄 train_df.pkl
│   │   │   ├── 📄 valid_df.pkl
│   │   │   └── 📄 test_df.pkl
│   │   └── 📁 4/                      # Split 4
│   │       ├── 📄 train_df.pkl
│   │       ├── 📄 valid_df.pkl
│   │       └── 📄 test_df.pkl
│   │
│   └── 📁 ptb-xl/
│       ├── 📁 0/                      # Split 0
│       │   ├── 📄 train_df.pkl
│       │   ├── 📄 valid_df.pkl
│       │   └── 📄 test_df.pkl
│       ├── 📁 1/                      # Split 1
│       │   ├── 📄 train_df.pkl
│       │   ├── 📄 valid_df.pkl
│       │   └── 📄 test_df.pkl
│       ├── 📁 2/                      # Split 2
│       │   ├── 📄 train_df.pkl
│       │   ├── 📄 valid_df.pkl
│       │   └── 📄 test_df.pkl
│       ├── 📁 3/                      # Split 3
│       │   ├── 📄 train_df.pkl
│       │   ├── 📄 valid_df.pkl
│       │   └── 📄 test_df.pkl
│       └── 📁 4/                      # Split 4
│           ├── 📄 train_df.pkl
│           ├── 📄 valid_df.pkl
│           └── 📄 test_df.pkl
│
├── 📁 chapman/                        # Chapman waveform data
│   └── 📄 *.pkl                       # Waveform pickle files (e.g., MUSE_20180116_120827_04000.pkl)
│
└── 📁 ptb-xl/                         # PTB-XL waveform data
    └── 📄 *.pkl                       # Waveform pickle files (e.g., 00001.pkl)

---

Experiments

Hyperparameters

The following hyperparameters are explored in the configuration. You can modify their values in the configuration file at configs/config.yaml to experiment with different settings:

Name	Values
LAMBDA_GUID	{1, 0.8, 0.4, 0.2, 0.1}
LAMBDA_PATCH	{1, 0.8, 0.4, 0.2, 0.1}
PATCH_SIZE	{25, 50, 75}
LR	{1e-4, 2e-4, 4e-4}

---

Results

---