Data compression of the exercise ECG using a Kohonen neural network

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Data compression is crucial for electrocardiographic (ECG) data, balancing compression ratios with signal distortion.
• Vector quantization (VQ) is a common technique in image and speech compression, but its application to ECG data is novel.
• Minimizing distortion is paramount for measurement-sensitive ECG data.

Purpose of the Study:

• To apply vector quantization (VQ) using a Kohonen neural network for ECG data compression.
• To create an optimal codebook of vectors that represent the low-frequency ECG data space.
• To evaluate the effectiveness of this method in maximizing compression while minimizing signal distortion.

Main Methods:

• Utilized a Kohonen neural network, an unsupervised learning model, to generate a codebook of vectors.
• Employed VQ by replacing original ECG vectors with addresses of best-matching codebook vectors.
• Implemented a distortion threshold to retain original vectors if compression introduces unacceptable signal alteration, preserving key segments like QRS and T waves.

Main Results:

• The Kohonen network adapted codebook vectors based on distance measurements and temporal control.
• Compression was achieved by substituting vectors with codebook addresses, provided distortion remained within limits.
• Critical ECG segments (QRS, T waves) were typically preserved during the compression process.

Conclusions:

• The Kohonen neural network paradigm offers a promising approach for ECG data compression.
• This method effectively balances compression efficiency with the preservation of diagnostically important ECG signal components.
• Further compression can be achieved using lossless techniques on the generated compressed file.