ECG-Greedy: Multi-Modal ECG Classification with CWT Scalograms & Phasograms

SOTA ensemble AUC 0.9238 on PTB-XL Super-Diagnostic Task using early fusion of scalograms + phasograms + raw signals.

This repository implements the multi-representation ECG classification framework from combining:

• Time-domain: XResNet1d101 (raw ECG)
• Time-frequency: CWT → Scalograms
• Phase-domain: CWT → Phasograms
• Fusion: Early & Late
• Backbones: Custom 2DCNN, ResNet50, EfficientNet-B0, Swin, Hybrid Swin

---

Highlights

• Ensemble AUC: 0.9238 (beats XResNet1d101: 0.9224)
• Hybrid Swin is the best single backbone (0.9023, early fusion)

Performance of Top Models and Ensemble on the PTB-XL Superdiagnostic Task

Domain	Model	Loss Function	AUC	F1
Raw Signals	XResNet1d101	BCE	0.9224	0.7265
Early Fusion	Hybrid Swin	Focal (W1)	0.9023	0.6970
Early Fusion	ResNet50	Focal (W1)	0.8915	0.6850
—	Ensemble	—	0.9238	0.7402

• Early > Late > Single modality
• Full preprocessing, training, and evaluation pipeline
• Reproducible on PTB-XL with patient-wise splits

Note: All the commands are based on a Unix based system. For a different system look for similar commands for it.

Setup

We are using Python version 3.11.9

$ python --version
Python 3.11.9

Requirements

# Clone and install
git clone https://github.com/kagozi/MultiModal-ECG.git
cd MultiModal-ECG

pip install -r requirements.txt

Python virtual environment

Create a virtual environment:

python3 -m venv .venv

.venv is the name of the folder that would contain the virtual environment.

Activate the virtual environment:

source .venv/bin/activate

Windows

source .venv/Scripts/activate

Quick Start

Download the PTBXL dataset to a preferred directory:

wget https://physionet.org/content/ptb-xl/get-zip/1.0.3/

Run full pipeline: preprocess → train → evaluate → ensemble

python3 load_and_standardize.py # creates standardized raw signal representations
python3 generate_cwt.py # creates Wavelet representations (Phasogram and Scalograms)
python3 benchmark.py # Trains XResNet1d101 on raw signals
python3 train_models.py # Trains CWT based models on Phasogram and Scalograms
python3 tests_ensemble.py # Run tests, generate confusion matrices, and evaluate different ensembling strategies