Democritus University of Thrace (DUTH)
Emerging machine learning methods and their applications to life sciences
Abstract
dc:descriptionEmerging Machine Learning (ML) methods are increasingly embedded in the Life Sciences, particularly in health and social care, where reliable predictive models must be developed from heterogeneous data under strict privacy, ethical, and regulatory constraints. Federated Learning (FL) has emerged as a practical paradigm for collaborative model training that avoids centralizing sensitive records, thereby preserving privacy while retaining predictive utility. This thesis investigates FL for real-world healthcare and social-care applications along two complementary axes: (i) designing effective and sustainable predictive models under statistical heterogeneity and class imbalance, and (ii) assessing privacy risks and strengthening resilience against membership inference threats. First, we study early dropout prediction for healthy ageing applications using a real dataset characterized by non-IID client distributions and severe class imbalance. We propose and evaluate principled client-selection strategies and sampling schemes that improve representativeness, stabilize convergence, and enhance minority-class performance while substantially reducing training time and environmental footprint. The results demonstrate that carefully designed participation policies enable FL models to match or surpass centralized baselines in realistic deployments. Second, we present a systematic analysis of Membership Inference Attacks (MIAs) in an FL healthcare setting. We formalize both record-level and client-level MIAs, evaluate the impact of FL configuration choices---including client selection, sampling strategies, and differential privacy---and examine the role of synthetic data. We further introduce a threshold-tuning mechanism that strengthens the evaluation of client-level attacks and clarifies privacy--utility trade-offs across threat models. The analysis provides actionable guidance for configuring FL systems that balance accuracy with reduced susceptibility to inference attacks. Finally, we address the automated classification of bicuspid versus tricuspid aortic valves (BAV vs. TAV) from routine echocardiographic cine loops. We develop a robust and interpretable stacked video-ensemble that combines multiple pretrained backbones through leakage-free, strictly out-of-fold stacking, applies probability calibration for reliable risk estimation, and delivers dual-level interpretability via frame-level Grad-CAM and meta-level SHAP. The results demonstrate accurate and transparent BAV versus TAV discrimination under real-world variability, highlighting the applicability of modern, trustworthy ML methods to clinical decision support. Taken together, this thesis demonstrates how principled design of emerging ML methods, particularly Federated Learning, can deliver accurate, privacy-aware, and resource-conscious models in sensitive Life Sciences domains, and provides practical insights for practitioners navigating trade-offs between performance, sustainability, and privacy.
Degree
thesis:*- Grantor dc:publisher
- Democritus University of Thrace (DUTH)
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Nikolaidis, Christos-Chrysanthos
- Νικολαΐδης, Χρήστος-Χρύσανθος
Subjects
dc:subject × 26- Ομοσπονδιακή μάθηση
- Μηχανική μάθηση
- Βιοεπιστήμες
- Υγεία
- Ιδιωτικότητα
- Επιλογή συμμετεχόντων
- Ανισορροπία Κλάσεων
- Βιωσιμότητα
- Επιθέσεις συμπερασμού συμμετοχής
- Καρδιακή απεικόνιση
- Federated learning
- Machine learning
- Life sciences
- Healthcare
- Privacy
- Client selection
- Class imbalance
- Sustainability
- Membership inference attacks
- Cardiac imaging
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Τεχνολογία μέσων
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Media Technology
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/61732