Aristotle University Of Thessaloniki (AUTH)
Predictive modelling for the prevention of complications of heart failure patients using artificial intelligence and deep learning techniques
Abstract
dc:descriptionHeart failure (HF) remains a major global health challenge, contributing to high morbidity and mortality rates. Early detection and risk stratification are crucial for preventing complications, yet traditional clinical methods often lack accuracy and reliability. Recent advancements in artificial intelligence (AI) and deep learning (DL) offer promising solutions for improving predictive modeling in HF management. This dissertation begins with a systematic review of the role of AI in HF evaluation. A comprehensive literature search across four major databases – PubMed, Scopus, Web of Science and IEEE Xplore – identified 137 relevant studies, categorized into five primary application areas: cardiovascular disease classification, HF detection, image analysis, risk assessment and other clinical analyses. The findings highlight the increasing adoption of AI techniques in cardiovascular medicine, demonstrating their potential for improving diagnostic accuracy, identifying underlying cardiomyopathies, refining risk assessment models, and personalizing treatment strategies. Building on these insights, this dissertation develops AI-driven predictive models for HF-related complications, focusing on three key aspects. First, a novel hybrid feature selection methodology is introduced, integrating Extremely Randomized Trees with non-linear correlation measures to enhance one-year all-cause mortality prediction in HF patients using classic echocardiographic and key demographic data. Unlike conventional feature selection methods, this approach is model-independent, ensuring robustness and generalizability. The optimal feature subset as identified through loss graph inspection maintains high predictive performance while significantly reducing dimensionality. Additionally, SHapley Additive exPlanations is employed to enhance model interpretability, offering insights into how individual features influence predictions. Second, a DL framework is developed for the automated detection of HF with reduced ejection fraction using echocardiographic video data. The study leverages the TimeSformer architecture, a Transformer-based model optimized for spatiotemporal feature extraction, in conjunction with transfer learning. Pretraining on a large open-access dataset followed by fine-tuning on a specialized clinical dataset significantly enhances classification accuracy. Furthermore, left ventricle masking is incorporated to improve model focus on diagnostically relevant structures, further boosting performance. Finally, this research addresses the challenge of uncertainty in medical imaging predictions by introducing an uncertainty-aware convolutional neural network. The model utilizes test-time augmentation to quantify aleatoric uncertainty in cardiac MRI data, enabling the generation of empirical distributions for test samples. This methodology not only provides reliable uncertainty estimates but also enhances overall classification performance, offering a valuable tool for improving clinical decision-making. The results of this dissertation underscore the transformative potential of AI-based predictive modelling in HF assessment. The proposed methodologies enhance diagnostic accuracy, interpretability and reliability, paving the way for their integration into clinical workflows. Future research should focus on refining these models with larger, more diverse datasets and exploring their real-world applicability in personalized patient care. Ultimately, this dissertation contributes to the advancement of precision medicine in HF care, paving the way for more accurate, individualized and clinically actionable AI solutions.
Degree
thesis:*- Grantor dc:publisher
- Aristotle University Of Thessaloniki (AUTH)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Petmezas, Georgios
- Πετμεζάς, Γεώργιος
Subjects
dc:subject × 26- Καρδιακή ανεπάρκεια
- Τεχνητή νοημοσύνη
- Βαθιά μάθηση
- Μηχανική μάθηση
- Ερμηνεύσιμη τεχνητή νοημοσύνη
- Πρόβλεψη θνησιμότητας
- Ανίχνευση καρδιακής ανεπάρκειας με μειωμένο κλάσμα εξώθησης
- Ποσοτικοποίηση αβεβαιότητας
- Υπερηχοκαρδιογραφία
- Μαγνητική τομογραφία καρδιάς
- Heart failure
- Artificial intelligence
- Deep learning
- Machine learning
- Explainable AI
- Mortality prediction
- HFrEF detection
- Uncertainty quantification
- Echocardiography
- Cardiac MRI
- Επιστήμες Υγείας
- Ιατρική και Επιστήμες Υγείας
- Ιατρική Πληροφορική
- Health Sciences
- Medical and Health Sciences
- Medical Informatics
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/59127