Compression of Multilead Electrocardiogram Using Principal Component Analysis and Machine Learning Approach

摘要

In this work, a multi-lead Electrocardiogram (ECG) data compression using principal component analysis (PCA) combined with a machine learning technique is proposed to achieve a high compression ratio (CR) with low reconstruction error (within 2% percentage root mean squared difference, or, PRD). The beat detection procedure was inspired by the Pan-Tompkins algorithm with some necessary modifications. A lead-wise PCA decomposition was performed for dimensionality reduction with a single beat from each lead at a time using a fixed energy reconstruction criteria. The optimal quantization levels of the principal components were allocated using multi-layer perceptron neural network (MLP-NN) using lead clinical features as the input. This MLP-NN was tuned offline by a particle swarm optimization (PSO) generated data for quantization level of coefficients of PC as the reference. The proposed technique was evaluated using 8 types of cardiac abnormalities record from multi-lead ECG data from the PTB Diagnostic ECG database, with an average CR, PRD and PRDN of 16.2, 1.47% and 1.84% respectively. The reconstructed records were clinically acceptable. The proposed technique provides superior performance than few recent published works on multilead ECG compression.