Abstract/Summary

Rheumatic and musculoskeletal disease (RMDs) is a chronic disease, which affects over 20 million people in the UK, and approximately 1.71 billion people in the world. Particularly, there are two prominent subdivisions of RMDs, inflammatory arthritis (IA) and non-inflammatory conditions (NIC), which have clearly different treatment and management pathways (e.g., disease-modifying drugs for IA; surgeries such as joint replacements for NIC e.g., osteoarthritis). Therefore, accurately differentiate IA and NIC is essential for patients to be referred to the right specialists and receive the right treatment rapidly. Early detection of RMDs is challenging because it often features vague symptoms, and there is currently no diagnostically definitive single biomarker for the detection. Moreover, the prevailing manual review process is hampered by low efficiency, primarily attributed to the intricacy of data modalities originating from general practitioners (GP), which encompasses unstructured GP referral letters, structured blood test results, and semi-structured clinical information summary (CIS). In this thesis, extracting features from various above-mentioned data modalities acts as the main challenge to develop a machine learning based system for the early RMDs identification. To address these data challenges, distinct machine learning-based methods have been proposed, including large language model (LLM) based methods to address unstructured data challenge, general machine learning model-based methods to address structured data challenge, graph neural network (GNN) based methods to deal with semistructured data challenge, and multimodal machine learning based methods to address multimodal data challenge that happened during the referral triage process. Experimental findings demonstrated relatively strong performance of the proposed methods on various datasets that have been collected from the Royal Berkshire Foundation Trust (RBFT), Reading, UK. Additionally, a small pilot trial has been conducted at the Rheumatology Department of RBFT. The experimental results indicated that the proposed models outperformed clinicians across various measurement metrics. Overall, this thesis is the first study to research early rheumatic disease diagnosis by using machine learning-based decision support methods, aimed at improving the hospital referral triage with interpretable and trustworthy results provided for clinicians. This study revealed promising performance of machine learning-based risk stratification methods in early RMDs differentiation, demonstrating great potential for future practical applications.