Baseline Drift Tolerant Signal Encoding for ECG Classification with Deep Learning.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Automated electrocardiogram (ECG) analysis is crucial for diagnosing cardiac conditions.
• Common artefacts such as baseline drift, rescaling, and noise degrade the performance of machine learning models.
• Existing methods struggle to maintain accuracy in the presence of these signal imperfections.

Purpose of the Study:

• To introduce Derived Peak (DP) encoding, a novel non-parametric method for ECG signal preprocessing.
• To evaluate the robustness and performance of DP encoding against common ECG artefacts.
• To assess the impact of DP encoding on machine learning model accuracy for cardiac condition identification.

Main Methods:

• Developed Derived Peak (DP) encoding, a non-parametric technique generating signed spikes from signal derivatives.
• Applied DP encoding to the PTB-XL dataset (n=18,869) for 12-lead ECG data.
• Trained 1D-ResNet-18 models using DP-encoded data for myocardial infarction, conductive deficits, and ST-segment abnormality detection.
• Assessed robustness by corrupting ECG data with baseline drift, shift, rescaling, and noise.

Main Results:

• DP encoding demonstrated invariance to shift and scaling artefacts, requiring no user-defined parameters.
• While other methods showed significant accuracy drops, DP encoding maintained a baseline AUC of 0.88 under drift, shift, and rescaling.
• DP encoding outperformed unencoded inputs in the presence of shift (AUC 0.91 vs 0.62) and rescaling artefacts (AUC 0.91 vs 0.79).

Conclusions:

• Derived Peak (DP) encoding offers a simple yet effective approach to enhance robustness in automated ECG analysis.
• This method significantly improves machine learning model performance by mitigating the impact of common ECG artefacts.
• DP encoding represents a valuable advancement for reliable and accurate automated ECG interpretation.