We explore the application of large language models to clinical documentation and medical knowledge management. Our interest centres on models adapted to medical terminology and the structure of clinical language — models capable of extracting information from unstructured medical text, summarising patient histories, and answering natural language queries across complex health information systems.

Adapting language models to medical data raises requirements that do not arise in other domains: the confidentiality of training data, the reliability of generated responses, and the need to substantiate every claim. Our research focuses on how these requirements can be met without sacrificing usability, with applications that reduce the administrative burden on health professionals and improve documentation quality.

Retrieval Augmented Generation (RAG)

RAG systems combine the linguistic capability of large models with dynamic access to verified sources of medical knowledge. Rather than answering solely from what the model retained during training, it first retrieves the relevant literature, clinical guidelines or patient-specific data and grounds its response in them. This approach substantially reduces the risk of generating incorrect information and makes it possible to trace any answer back to its source — an essential property for any clinical decision support application.

Knowledge Graphs

We build medical knowledge graphs that map the relationships between conditions, symptoms, treatments, medications and clinical outcomes, forming a semantic network of medical information. These structured representations allow our systems to grasp clinical context, identify connections that are not immediately apparent, and support reasoning for differential diagnosis and treatment planning. The graphs interface with electronic health records and medical databases, serving as a foundation for intelligent clinical applications.

We develop biosignal processing algorithms to extract clinically useful information from physiological signals and biosensor data. Our work covers signal pre-processing and cleaning, feature extraction, and the application of machine learning methods to identify patterns and anomalies.

Real-time processing raises particular challenges: biosignals carry noise, motion artefacts distort measurements, and the boundary between normal variation and a clinically significant finding is not always clear-cut. Our research focuses on developing methods that remain reliable under these conditions, so that continuous monitoring produces actionable alerts rather than alarm noise.

We develop computer vision systems for medical image analysis, drawing on deep learning architectures and image processing techniques. Our applications address the detection of abnormalities, measurement of anatomical structures, tracking of disease progression, and identification of patterns that are difficult to discern through visual inspection alone.

The clinical value of such systems depends less on model accuracy than on how well they fit into everyday workflow. We therefore prioritise explainable AI approaches that provide visual evidence for their findings, so the clinician retains full oversight of the decision, and we focus on integration with existing PACS systems and clinical processes.

We research continuous monitoring systems that draw on data from wearable devices, sensors and medical equipment, with the aim of recognising changes in a person’s condition early. Our work covers the analysis of physiological trends over time, the mechanisms that trigger alerts when measurements fall outside expected ranges, and the presentation of that information in a form both the individual and the health professional following them can act on.

The critical question in such systems is not detection itself but the balance between sensitivity and usability. A system that alerts too often becomes tiring and is eventually ignored, while an overly conservative one risks missing significant changes. Our research focuses on that balance, through personalised alert thresholds and predictive methods that shift intervention from reaction towards prevention.

We research federated learning approaches, which allow AI models to be trained collaboratively across multiple organisations without centralising sensitive data. Rather than moving the data, the model moves: each institution trains locally on its own data and shares only the resulting parameters, while patient information stays within its own infrastructure.

This approach is particularly valuable where a single institution’s data is not enough — in rare disease research, population-level studies, and the development of models that must perform reliably across different patient groups and clinical practices. It also brings challenges of its own: data heterogeneity between sites, the computational and communication cost of coordination, and ensuring that the model parameters themselves do not leak information about the underlying data. Our research works around these questions.

We research running AI models directly on the device, rather than sending data to centralised cloud infrastructure. This calls for lightweight models, optimised to operate within constrained computing resources and power budgets while retaining sufficient accuracy for the task at hand.

The appeal of the approach is twofold. Practically, it enables real-time analysis where connectivity is limited or unreliable, without the latency introduced by a round trip to the cloud. In privacy terms, processing locally means raw data need never leave the device — only the result is transmitted. Our research focuses on the balance between model size, computational efficiency and the reliability of the output.

Security and data protection run through the whole of our research activity. We design systems that protect health information across its entire lifecycle — from collection and processing through to storage and transmission — in compliance with GDPR, HIPAA and the national regulatory framework.

Beyond technical safeguards, we apply the privacy-by-design principle: data protection is built into the architecture from the earliest design stage rather than added afterwards. Our research covers anonymisation and pseudonymisation techniques, differential privacy methods, and approaches that allow health data to be put to use without exposing the individuals behind it. We also follow technologies still at the research stage, such as homomorphic encryption, which permits computation directly on encrypted data.

In every case the aim is balance: systems open enough for data to produce knowledge, and rigorous enough to maintain the trust of patients and of the institutions that hold their data.