ECG-XPLAIM: Deep learning tool for accurate and explainable arrhythmia detection from 12-lead electrocardiogram (ECG) signals

ECG-XPLAIM is a deep learning-based model designed for explainable ECG classification, optimized for multi-label classification of 12-lead electrocardiogram (ECG) signals. The model integrates a custom Inception-style one-dimensional convolutional neural network (CNN), specialized for time-series analysis, capturing both local waveform features (waves, intervals, QRS morphology) and global rhythm patterns (RR variability, conduction disturbances).

1. Multiple Versions for Different Classification Tasks

This repository provides different versions of ECG-XPLAIM, each trained and optimized for a specific arrhythmia classification task. Each model was trained separately using task-specific balanced subsets, allowing for more fine-tuned, task-specific performance. The versions include:

• TACHY (ecg_xplaim_TACHY.keras): Differentiates Atrial Fibrillation (AFib) from Sinus Tachycardia (STach) and non-AFib/STach samples.
• CD (ecg_xplaim_CD.keras): Detects Conduction Disturbances, including Right Bundle Branch Block (RBBB), Left Bundle Branch Block (LBBB), and Left Anterior Fascicular Block (LAFB).
• LQT (ecg_xplaim_LQT.keras): Identifies Long QT interval (LQT).
• WPW (ecg_xplaim_WPW.keras): Detects Wolff-Parkinson-White syndrome (WPW).
• PACE (ecg_xplaim_PACE.keras): Identifies paced ECG rhythms (PACE).

2. Input and Output Shape for Each Task

Input Shape (all versions): (batch_size, 5000, 12) [5000 time points, 10sec at 500Hz sampling, × 12 ECG leads, corresponding to: I, II, III, aVR, aVL, aVF, V1-V6]

Output Shape (multi-label classification): (batch_size, N_classes) [binary multi-label output for arrhythmia presence/absence]

Each output is a vector of probabilities (sigmoid activation; 0 to 1 probability), where each index corresponds to a specific arrhythmia class. The model does not assume mutual exclusivity—i.e., an ECG can have multiple conditions simultaneously.

3. How to Load and Use the Model in Python

ECG-XPLAIM is implemented in TensorFlow/Keras.

# Load TensorFlow/Keras
import tensorflow as tf

# Load model (example for the TACHY version)
model_path = "path/to/ecg_xplaim_TACHY.keras"
model = tf.keras.models.load_model(model_path)

# Simulated ECG data (100 samples, of 12 leads, each with 5,000 data points)
input_ecg_data = np.random.randn(100, 5000, 12).astype(np.float32)

# Get predicted labels
output_labels = model.predict(ecg_sample)

4. Additional Resources

• Code Repository: GitHub Repository (https://github.com/ppantele/ECG-XPLAIM)
• Published Paper: (link to be added soon)