Department of Computer Engineering

This thesis presents a modular, privacy‑preserving clinical decision‑support architecture that fuses large‑language‑model (LLM) reasoning with federated learning (FL) and event‑driven streaming analytics. Structured vitals, laboratory data, and unstructured triage notes are processed in near real‑time, while differential privacy, secure aggregation, and encrypted model serving ensure GDPR compliance and FAIR data stewardship. A two‑stage evaluation shows both knowledge‑level and bedside effectiveness. (i) On a 750‑question USMLE benchmark, a fine‑tuned Llama 3.1‑70 B model reaches 74% overall accuracy and 83% on Step 2, outperforming the untuned baseline by 22 percentage points and the medical‑student mean by > 15 points. (ii) In a real‑world emergency‑department cohort of 132 infectious‑disease encounters, the pipeline achieves 75.8% first‑pass agreement with specialist review for diagnosis, justification, and next‑step recommendations. Parallel inference tests show that 4‑bit quantisation delivers the highest single‑query throughput, whereas 8‑bit models sustain steadier performance under four‑thread loads. These results demonstrate that the proposed CDSS combines state‑of‑the‑art reasoning depth with audit‑compliant and low‑latency deployment.