University of Thessaly (UTH)
Development of deep learning methods on big data with a focus on natural language processing
Abstract
dc:descriptionThe rapid proliferation of large-scale biomedical and natural language data has created both unprecedented opportunities and persistent challenges for Artificial Intelligence (AI). Deep Learning (DL) methods are particularly powerful for modeling sequential and heterogeneous data, yet their deployment in practice is limited by missing data, computational overhead, and the demand for domain-specific adaptation. This dissertation investigates these challenges and proposes novel DL methodologies that extend the state of the art in both biomedical prediction and Natural Language Processing (NLP). In the biomedical domain, a framework for predictive precision medicine has been developed using longitudinal Electronic Health Records (EHRs). The framework incorporates Recurrent Neural Networks (RNNs), Ensemble Learning strategies, and robust imputation techniques to improve the stability and accuracy of clinical risk prediction. This work further examines the impact of medication data and patient adherence, demonstrating how these factors influence predictive performance in real-world healthcare settings. In the NLP domain, three distinct contributions are made. First, transformer-based models are applied to large-scale social media data, and a sentiment change detection pipeline is introduced to track public opinion shifts during the COVID-19 pandemic. Second, the thesis addresses the emerging problem of distinguishing human-authored text from AI-generated text, presenting a transfer learning approach that achieves high accuracy and generalizes across domains ranging from scientific abstracts to book reviews. Third, a scalable and energy-efficient framework for topic modeling is introduced, integrating hierarchical clustering with targeted Large Language Model (LLM) prompting to deliver coherent topics while substantially reducing computational cost. Collectively, these contributions demonstrate how DL can be adapted to diverse forms of large-scale sequential data while emphasizing robustness, interpretability, and efficiency. The thesis advances methodological innovation and establishes new applications across healthcare, social media, and scientific text mining. Ultimately, this work underscores the imperative of developing DL systems that are not only accurate but also efficient and grounded in real-world constraints.
Degree
thesis:*- Grantor dc:publisher
- University of Thessaly (UTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Theocharopoulos, Panagiotis
- Θεοχαρόπουλος, Παναγιώτης
Subjects
dc:subject × 32- Μηχανική μάθηση
- Μεγάλα δεδομένα (BigData)
- Βαθιά μάθηση
- Επεξεργασία φυσικής γλώσσας
- Μεγάλα γλωσσικά μοντέλα (LLMs)
- Ηλεκτρονικός Φάκελος Υγείας (ΗΦΥ)
- Θεματική μοντελοποίηση
- Αλγόριθμοι ομαδοποίησης
- Τεχνητή νοημοσύνη (ΤΝ)
- Μέθοδος πρόβλεψης
- Αλγόριθμοι ταξινόμησης
- Υγειονομική περίθαλψη
- Ανάλυση δεδομένων
- Προγνωστική Ιατρική Ακριβείας
- Machine learning methods
- Big data
- Natural Language Processing (NLP)
- Deep learning algorithms
- Large Language Models (LLMs)
- Electronic health record ( EHR )
- Topic modeling
- Artificial intelligence
- Clustering algorithms
- Predictive modelling
- Classification methods
- Data analysis
- Predictive Precision Medicine
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Φυσικές Επιστήμες
- Τεχνητή νοημοσύνη
- Computer and Information Sciences
- Natural Sciences
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/61294