Rajapura Sreedhar, V (2024) Development of Artificial Intelligence Algorithms in Cardiovascular Research: Case Studies in Atrial Fibrillation and Myocardial Infarction. Doctoral thesis, Liverpool John Moores University.

Text 2024vikramphd.pdf - Published Version Restricted to Repository staff only until 11 May 2025. Available under License Creative Commons Attribution Non-commercial. Download (5MB)

Abstract

This thesis explores the application of artificial intelligence techniques in the early detection and management of adverse cardiovascular events. The research is grounded in the quest to identify novel biomarkers using multi-modal data, including electrocardiograms, cardiac magnetic resonance imaging, and clinical records. A comprehensive literature review reveals the evolution of segmentation approaches in cardiac health, highlighting the transition from traditional methods to advanced artificial intelligence driven techniques. The study then focuses on the development and validation of U-Net a neural network in segmenting cardiac magnetic resonance data of left atrium for Atrial Fibrillation fibrosis, incorporating explainability features like Gradient-weighted Class Activation Mapping and saliency maps to enhance transparency and interpretability. Further, the research extends to analysing electrocardiogram data through machine learning and deep learning techniques, identifying critical gaps and limitations in existing approaches and suggesting areas for innovation. The integration of demographic and clinical records with electrocardiogram modalities is explored, emphasizing the potential for discovering multi-modal biomarkers that could revolutionize cardiac health management. Generative Adversarial Networks are explored for their potential to enhance model robustness by augmenting datasets, addressing data scarcity and class imbalances prevalent in cardiac research. The implications of these findings are discussed in terms of their clinical relevance and potential applications. Generative Adversarial Networks are investigated for their potential to enhance model robustness by augmenting datasets, addressing data scarcity and class imbalances prevalent in cardiac research. In the realm of electrocardiogram data, Earth Mover's Distance scores reveal the congruence between real and synthetic data distributions. Results indicate that Generative Adversarial Networks effectively capture the underlying pathology of myocardial infarction, with a mean Earth Mover's Distance score of 3.278 in principal component analysis and 5.544 in t-distributed stochastic neighbour embedding, suggesting a closer alignment in principal component analysis space for myocardial infarction signals. However, Generative Adversarial Networks posed challenges in recognizing healthy control instances, with a mean Earth Mover's Distance score of 4.576 in principal component analysis and 7.838 in t-distributed stochastic neighbour embedding. These findings underscore the need for careful consideration of the trade-offs introduced by synthetic data augmentation in medical diagnostics. In further analysis, this study employs machine learning techniques, including Logistic Regression, Random Forest, and Extreme Gradient Boosting, to advance the classification of cardiovascular disorders. Interbeat intervals were extracted from electrocardiogram data and combined with clinical factors to enrich the dataset for binary and multi-class classifications of disorders such as myocardial infarction, bundle branch block, cardiomyopathy, dysrhythmia, and healthy heart conditions. The augmentation of this data with synthetic samples generated by Generative Adversarial Networks significantly enhanced model performance. Notably, in binary classification tasks, Random Forest model without data augmentation, demonstrated the highest accuracy and Area Under the Curve at 0.89 and 0.93 respectively, suggesting its effectiveness in accurately classifying and differentiating conditions. Meanwhile, models augmented with synthetic data, showed remarkable precision for the myocardial infarction class, with Extreme Gradient Boosting achieving the highest Area Under the Curve of 0.96. Precision for healthy controls detection increased from 0.69 to 0.99, while the 5 F1 score improvises from 0.64 to 0.80. These results underscore the balanced performance across various classes, significantly enhanced by the integration of synthetic data, leading to improved diagnostic capabilities. Furthermore, the study delves into multi-class classification of cardiac disorders, where data augmented by Generative Adversarial Networks significantly boosts model performance. Overall accuracy increased from 0.68 to 0.75, and precision and recall were notably improved across various classes, including bundle branch block (precision from 0.76 to 0.82, recall from 0.50 to 0.82), cardiomyopathy (precision from 0.67 to 0.75, recall from 0.50 to 0.70), and dysrhythmia (precision from 0.71 to 0.80, recall from 0.50 to 0.70). These results emphasize the importance of data quality in synthetic data generation, especially when faced with class imbalances. In conclusion, this study showcases the efficacy of Generative Adversarial Networks -based data augmentation in enhancing the accuracy and reliability of predictive analytics in cardiovascular health. It provides valuable insights into the potential and challenges of synthetic data in medical diagnostics, with implications for improved patient care and disease detection. Conclusively, this research contributes significantly to the field of cardiology, offering novel artificial intelligence driven solutions for early detection and management of adverse cardiovascular conditions. It underscores the transformative potential of artificial intelligence in healthcare, advocating for the development of ethically responsible artificial intelligence applications that align with clinical needs and enhance patient health outcomes.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	Cardiac Conditions: Arrhythmias, Myocardial Infarction, Cardiomyopathies, Bundle Branch Block, Atrial Fibrillation, Ventricular Dysfunction, Dysrhythmias. Artificial Intelligence in Cardiology: Machine Learning, Deep Learning, Computational Cardiology, AI in Healthcare. Data Analysis Techniques: Support Vector Machines, Decision Trees, Neural Networks, Recurrent Neural Networks, Generative Adversarial Networks (GANs), Wasserstein GAN (WGAN), Ensemble Model, Data Fusion, Multimodal Analysis. Medical Imaging: Cardiac MRI, Cardiac Magnetic Resonance Imaging, Electrocardiography (ECG), Electroencephalography (EEG). Feature Engineering: Electrocardiogram Analysis, Interbeat Intervals (IBIs), Peak Detection, Image Segmentation. Statistical and Analytical Techniques: Logistic Regression, Random Forest, Extreme Gradient Boosting, Survival Analysis, Causal Inference, t-SNE (t-distributed Stochastic Neighbor Embedding). Explainability Techniques: Gradient-weighted Class Activation Mapping (Grad-CAM), Saliency Maps, Explainable AI, SHAP (SHapley Additive exPlanations). Distance and Loss Metrics: Wasserstein Distance, Earth Mover’s Distance, Wasserstein Loss. Data Augmentation: Synthetic Data, Data Imbalance, Generative Adversarial Networks (GANs), WGAN, CLAHE (Contrast Limited Adaptive Histogram Equalization), Histogram Equalization, Contrast Stretching. Healthcare Data Integration: Clinical Records, Multimodal Biomarkers, Cardiac Health Diagnostics, PTB Database. Clinical Applications: Prediction and Classification of Cardiac Events, Health Outcome Improvement, Personalized Medicine.
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QH Natural history > QH301 Biology R Medicine > RC Internal medicine
Divisions:	Computer Science and Mathematics
SWORD Depositor:	A Symplectic
Date Deposited:	11 Nov 2024 14:20
Last Modified:	11 Nov 2024 14:20
DOI or ID number:	10.24377/LJMU.t.00024724
Supervisors:	Olier, I, Ortega-Martorell, S and Lip, G
URI:	https://researchonline.ljmu.ac.uk/id/eprint/24724

View Item