Innovations in Optimization and Applications in Healthcare

While AutoML promises to democratise AI, its adoption in high-stakes domains like healthcare is impeded by prohibitive computational cost, fragile adaptation to real-world data (shift, missingness, device heterogeneity), and limited transparency. Spanning nine chapters, this thesis proposes an end-to-end methodology for practical, resource-aware, and clinically-viable AutoML. The methodology composes four capabilities: (i) a constraint-aware optimisation engine that searches efficiently while preserving diversity; (ii) knowledge-aware initialisation with foundation models grounded in metadata to avoid cold starts; (iii) resource-faithful benchmarking with short, repeatable budgets; and (iv) clinically grounded validation through end-to-end pipelines embedded in real workflows. From a computer science perspective, the thesis contributes new algorithms and evaluation protocols for constraint-aware evolutionary AutoML and LLM-guided initialisation under strict budgets. From a clinical engineering perspective, it delivers ECG analysis pipelines that remain interpretable and aligned with cardiology practice. We introduce GEISHA, an asynchronous generalised island model that accelerates pipeline discovery under explicit time/compute budgets while maintaining population diversity. A systematic study of ten migration topologies shows that balanced connectivity (e.g., grids, small-world/hypercube) mitigates premature convergence and supports robust exploration in high-dimensional spaces. Multi-fidelity evaluation is presented as one effective mechanism among budgeted strategies to prune unpromising configurations early.We integrate large language models (LLMs) with retrieval over prior metadata and dataset similarity (including optimal transport) to provide knowledge-aware initialisation of AutoML search. This LLM-guided prompting engine generates executable pipeline drafts and sensible hyperparameter ranges, reducing time-to-first-viable solutions for both classification and regression. We also identify limits: prompt sensitivity, motivating validator layers and iterative planning-verification protocols rather than single-shot generation. We critique the high cost of standard evaluations and show that short budgets (5–30 minutes) can preserve framework rank consistency relative to multi-hour runs, enabling greener, more continuous assessment and broader accessibility. Early termination policies further reduce cost without sacrificing comparative validity. The methodology is validated in biomedical applications. First, we analyse long-term ECG recorded via a self-adhesive dry-electrode patch over five days in cardiac patients, revealing diurnal quality patterns and quantifying fitness-for-use with morphology- and HRV-based Signal Quality Indicators (SQIs). Second, we deliver an interpretable, resource-efficient ST-segment deviation detector that classifies elevated/depressed/normal segments and quantifies microvolt-level shifts, producing lead- and time-specific outputs that reflect the reasoning of cardiologists. In the context of the INNO4HEALTH project, these pipelines make long-term monitoring more robust (fewer unusable segments, better handling of noise) and more actionable (automatic, explainable ST alerts), illustrating how constraint-aware search, knowledge-aware initialisation, and realistic evaluation yield tools that are concretely more helpful in clinical workflows.By treating efficiency, knowledge, evaluation, and deployment as a single methodology, this thesis advances the state of the art in AutoML and demonstrates its feasibility in a real healthcare project. For the AutoML and broader computer science community, it provides a general recipe and open-source implementations for constraint-aware optimisation, LLM-guided initialisation, and resource-faithful benchmarking. For healthcare and INNO4HEALTH, it delivers and validates ECG-based decision-support components that meet clinically meaningful explanations and can be integrated into end-to-end monitoring pathways, moving AutoML from promise to operational readiness for resource-aware, clinically-viable AI.