Deep Learning Techniques for Structured Data Analysis in Biomedical Research

Biomedical research increasingly relies on the analysis of complex, structured data, necessitating advanced methodologies to model the intricate relationships that characterize them. This thesis explores the application of Machine Learning and Deep Learning techniques across diverse biomedical domains, including genomic analysis, medical signal processing, medical imaging, and protein–protein interaction prediction. Contributions of this thesis include the development of a medical signal processing framework to reconstruct 12–lead ECG signals from single–lead input while simultaneously classifying their medical significance, ensuring both accuracy and efficiency through Deep Learning and model quantization. In the medical imaging setting, the work introduces a liver tumor segmentation method that integrates DeepLabV3+ with Hidden Markov Models, achieving improved performance in delineating complex tumor morphology. For protein–protein interaction prediction, a hybrid BiLSTM and Graph Neural Network model is proposed, leveraging enriched graph representations of secondary structure elements to effectively identify interacting regions. Finally, in the context of genomic analysis, a Machine Learning–based tool, miXer, was developed to detect Copy Number Variants from exome sequencing data with high accuracy, combining Support Vector Machines and Hidden Markov Models.