Supervised, Semi-supervised and Unsupervised Training Frameworks for Biomedical Signal Analysis

Biomedical signal analysis is critical in modern healthcare, providing essential information for monitoring, diagnosis, and treatment. However, the complexity and variability of these signals pose significant challenges. This research addresses these challenges by proposing novel methods to extract meaningful features from biomedical signals and efficiently deploy these solutions in real-world scenarios, particularly in embedded systems with limited resources. This thesis presents a comprehensive exploration of the development of advanced training frameworks for analysing biomedical signals, focusing on innovative approaches to feature extraction and the efficient deployment of these models in embedded systems. A set of advanced algorithms for feature extraction is introduced. These algorithms are designed to process biomedical signals efficiently. Using techniques such as deep learning, signal processing, and statistical analysis, the proposed methods demonstrate superior capability in identifying critical features that improve the accuracy and reliability of subsequent studies. The adaptability of these algorithms to various conditions has been extensively evaluated, demonstrating their robustness and effectiveness in different biomedical applications. Furthermore, this thesis focuses on deploying these algorithms in embedded systems, a crucial aspect of real-time biomedical signal analysis. Considering the limited resources of embedded devices, this research introduces optimized models that balance computational efficiency and analytical performance. Implementing these models includes hardware acceleration techniques, efficient memory management, and real-time processing strategies. The thesis provides an in-depth analysis of the trade-offs in embedding these complex algorithms, offering guidelines and best practices for their deployment.