Scalogram-CNN fusion with interactive 3D parameter exploration for inter-patient ECG classification.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence in Medicine

Background:

• Electrocardiogram (ECG) signals are crucial for diagnosing heart conditions but are challenging to analyze due to their non-stationary nature.
• Classifying ECG signals from unseen patients (inter-patient generalization) remains a significant hurdle in clinical applications.

Purpose of the Study:

• To investigate ECG classification performance using 2D scalogram images derived from ECG signals under unseen-patient conditions.
• To develop and evaluate an interactive 3D visualization framework for optimizing ECG analysis parameters.

Main Methods:

• ECG signals were transformed into 2D scalogram images.
• A Convolutional Neural Network (CNN) was employed for classifying the scalogram images.
• An interactive 3D visualization framework using Plotly Dash was developed to explore parameters like wavelet types, feature extraction levels, and scale ranges.
• Handcrafted statistical and wavelet-derived features were also evaluated on unseen patient data.

Main Results:

• Classification accuracy varied significantly based on the choice of wavelet and feature extraction strategy.
• Up to 89.78% accuracy was achieved using Complex Morlet wavelets in Experiment II, and 83.04% with Mexican Hat wavelets in Experiment I.
• The interactive 3D dashboard facilitated efficient parameter selection, result exploration, and analysis, demonstrating consistent performance across unseen-patient test settings.

Conclusions:

• The proposed framework effectively converts ECG signals into analyzable 2D images for classification.
• The interactive visualization tool enhances parameter exploration for ECG analysis, offering potential benefits for telemedicine and remote monitoring.
• This approach provides a valuable framework for small-data ECG analysis and improving inter-patient generalization.