Development of machine learning methods and algorithms for predicting cardiovascular pathologies based on echocardiography and electrocardiography

To develop methods of automatic annotation and interpretation of echocardiography and ECG based on the collected labeled data in order to identify the largest possible set of image anomalies and cardiac conditions using a CNN algorithm, and to develop a quality assessment of echocardiography images and ECG data with automatic correction of image artifacts using a CNN algorithm.

• Existing artificial intelligence methods and algorithms for the visualization and diagnosis of cardiovascular diseases for cardiology specialists were studied.
• The applicability of the artificial intelligence concept for the visualization and diagnosis of cardiovascular diseases was analyzed.
• An analytical review and grouping of unification and classification methods for the visualization and diagnosis of cardiovascular diseases were carried out.
• A system of key indicators for the visualization and diagnosis of cardiovascular diseases using artificial intelligence technologies was developed.
• Computational and statistical algorithms were applied to the machine learning model for diagnosing cardiovascular diseases.
• Echocardiogram and electrocardiogram data were collected and processed.
• Mathematical decision-making models were developed and visualized. Standard two-dimensional (2D) transthoracic echocardiography produces several video clips from different viewpoints using two-dimensional cross-sectional images. The clips do not always include labeled views, so view identification is an essential process in artificial intelligence applications. This step is required before training a deep neural network and performing automated measurements and image interpretation. Correct view selection is necessary for an accurate echocardiographic assessment. Moreover, since training deep learning algorithms requires a significant amount of annotated data, a view identification algorithm capable of handling this time-consuming task is needed. The pre-processing stage is expected to accelerate the development of artificial intelligence applications in echocardiography.

Figure 1. Data pre-processing and labeling (a — echocardiography video files; b — .png files; c — data labeling tool; d — labeling process)

Data pre-processing consists of the following steps: 1. Echocardiography video files (.avi format). 2. Fragmentation of an echo video file into individual .png files (used for labeling and further model training).

These files are then ready for the labeling process.

Data labeling for the segmentation task consists of the following steps: 1. https://segments.ai/ — a data labeling tool. The images from the previous step are uploaded to segments.ai for further labeling. 2. At the end of the labeling process, the labeled data must be released so that it can be used for model training. Segments.ai provides an easy-to-use interface for releasing new datasets. In echocardiography with normal anatomy, each standard view has a clear structure of the heart chambers. Therefore, annotations should be made for all standard views, including their unique chambers. The standard apical four-chamber view includes the following labels: aorta (Ao), left atrium (LA), left ventricle (LV), right atrium (RA) and right ventricle (RV), atrial septal defect (ASD), ventricular septal defect (VSD), tricuspid valve (TV), mitral valve (MV) and aortic stenosis (AS).

The annotated images provide a visual representation of the location of the defect in the atrium, the ventricle or both. Ultimately, the entire set of 401 tagged images is defined as a database with labeled regions and saved together with the corresponding statistical information. In total, physicians labeled 401 images.

Expert professionals were selected to take part in the image labeling procedure: a top-category physician of the Mediterra private medical center with experience in cardiac surgery and a top-category cardiologist from the Research Institute of Cardiac Surgery. This guarantees the highest level of professionalism and accuracy of medical image annotations and is crucial for achieving optimal results. The project also obtained approval from the Ethics Committee for research involving human subjects (Extract from Minutes No. 3 of October 30, 2023 of the Local Ethics Commission of RSE "IGP").

1. Mukasheva, A., Rakhmetulayeva, S., Astaubayeva, G., & Gnatyuk, S. (2022). Developing a system for diagnosing diabetes mellitus using big data. Eastern-European Journal of Enterprise Technologies, 5(2), 75–85. https://doi.org/10.15587/1729-4061.2022.266185
2. Rakhmetulayeva, S., & Syrymbet, Z. (2022). Implementation of convolutional neural network for predicting glaucoma from fundus images. Eastern-European Journal of Enterprise Technologies, 6(2), 70–77. https://doi.org/10.15587/1729-4061.2022.269229
3. Rakhmetulayeva, S. B., Kulbayeva, A. K., & Mukasheva, A. K. (2022). Development of electronic medical records system with data security algorithm in healthcare environment of Kazakhstan. Bulletin of Almaty University of Power Engineering and Telecommunications. ISSN 2790-0886.
4. Mukasheva, A. (2022). Data processing and analysis (in Kazakh). Bulletin of Almaty University of Power Engineering and Telecommunications (Monograph), Almaty, Kazakhstan.
5. Rakhmetulayeva, S. B., Ukibassov, B. M., Zhanabekov, Zh. O., & Mukasheva, A. K. (2023). Research on machine learning methods for recognition and classification of cardiovascular pathologies. Bulletin of Almaty University of Power Engineering and Telecommunications, DOI: 10.51775/2790-0886_2023_61_2_106
6. Rakhmetulayeva, S., Ukibassov, B., Zhanabekov, Z., & Mukasheva, A. (2023). Machine learning methods and algorithms for predicting congenital heart pathologies. In Proceedings of the 17th IEEE International Conference Application of Information and Communication Technologies (AICT). IEEE. https://doi.org/10.1109/AICT59525.2023.10313184
7. Ukibassov, B. M., Rakhmetulayeva, S. B., Zhanabekov, Zh. O., Bolshibayeva, A. K., & Yasar, A.-U.-H. (2024). Implementation of anatomy constrained contrastive learning for heart chamber segmentation. Procedia Computer Science, 238, 536–543. https://doi.org/10.1016/j.procs.2024.06.057
8. Ukibassov, B. M., Rakhmetulayeva, S. B., & Bolshibayeva, A. K. (2024). Semantic segmentation deep learning models in echocardiography: Custom dataset-based fine-tuning. Bulletin of Abai KazNPU, Physical and Mathematical Sciences Series, 1(85), 149–155. https://doi.org/10.51889/2959-5894.2024.85.1.014
9. Rakhmetulayeva, S., Ukibassov, B., Zhanabekov, Z., & Bolshibayeva, A. (2024). Development of data-efficient training techniques for detection and segmentation models in atrial septum defect analysis. Eastern-European Journal of Enterprise Technologies, 5(131). In press.
10. Bolshibayeva, A., Rakhmetulayeva, S., Ukibassov, B., & Zhanabekov, Z. (2024). Advancing real-time ultrasound heart imaging with a novel hybrid deep learning architecture for enhanced pattern recognition. Eastern-European Journal of Enterprise Technologies, 6(132).
11. Rakhmetulayeva, S. B. (2024). Development of methods and algorithms for machine learning for predicting pathologies of the cardiovascular system based on echocardiography and electrocardiography. BALAUSA print. (Monograph) ISBN 978-601-08-4538-1.
12. Sabina Rakhmetulayeva, Zhandos Zhanabekov, Aigerim Bolshibayeva, Evaluation of Modern Generative Networks for EchoCG Image Generation, Computers, Materials and Continua, Volume 81, Issue 3, 2024, Pages 4503-4523, ISSN 1546-2218, https://doi.org/10.32604/cmc.2024.057974