Optimization of generative-adversarial networks for the synthesis of electrocardiograms by adaptive quantization

Objective. This work is devoted to overcoming high requirements for computational resources and RAM volume through the development of effective optimization methods for generative adversarial networks, adapted for generating electrocardiogram signals.
Method. The work employed methods of deep learning generative adversarial networks, electrocardiogram signal processing techniques, and implementations of adaptive quantization of weight coefficients for searching the optimal quantization level to reduce memory consumption while maintaining data synthesis quality.
Result. Based on experiments, a proprietary architecture of a generative adversarial network for synthesizing electrocardiogram signals was developed and trained. Comparative analysis of the Fréchet Distance metric confirmed high similarity between generated data and real ECG recordings. The developed adaptive quantization method AGWR achieved a 75% reduction in RAM consumption while maintaining synthesis quality, outperforming the traditional Min-Max approach even at extreme weight compression to 4 bits.
Conclusion. The work overcomes the computational resource limitations of GANs, opening opportunities for practical implementation in medical systems with constrained hardware. The technique helps fill the gap in clinical data, accelerating the development of intelligent diagnostic tools and personalized medicine in cardiovascular pathology.

1. Hannun A. Y., Rajpurkar P., Haghpanahi M., et al. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature Medicine. 2019;25:65–69.

2. Wagner P., Strodthoff N., Bousseljot R., et al. PTB-XL, a large publicly available electrocardiography dataset. Scientific Data. 2020;7:154.

3. IanJ. Goodfellow. Generative Adversarial Nets. // arXiv preprint arXiv: 1406.2661. – 2014.

4. Jacob B., Kligys S., Chen B., et al. Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2018; 2704–2713.

5. Moody G. B., Mark R. G. The impact of the MIT-BIH Arrhythmia Database. IEEE Engineering in Medicine and Biology Magazine. 2001; 20(3):45–50.

6. O.P. Timofeeva, M.M. Gordeev, D.A. Koblyakov. ECG Image Processing and Generation. Proceedings of the R.E. Alekseev NSTU 2021;1(132):24.(In Russ)

7. Hochreiter S., Schmidhuber J. Long Short-Term Memory. Neural Computation. 1997;9(8):1735–1780.

8. Heusel M., Ramsauer H., Unterthiner T., et al. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. Advances in Neural Information Processing Systems (NIPS). 2017:. 6626–6637.

9. Fel Zhu, Ye Fei.Electrocardiogram generation with a bidirectional LSTM-CNN Generative adversarial network. Translational Biomedical Informatics. 2019:1-9.

10. A. Zhou et al. Incremental network quantization: Towards lossless CNNs with low-precision weights. Proceedings of the International Conference on Learning Representations (ICLR). 2017.

11. Z. Wang et al. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing. 2004; 13(4, Apr.):600–612.

12. Field, A. Discovering Statistics Using IBM SPSS Statistics. – London: Sage Publications Ltd., 2018.

13. H. Li et al. Pruning Filters for Efficient ConvNets. Proceedings of the International Conference on Learning Representations (ICLR). 2017.

14. Lloyd, S. P. Least Squares Quantization in PCM. IEEE Transactions on Information Theory. 1982; 28(2): 129–137.