Advancing Hepatitis B Functional Cure and Preventing Cirrhosis and Liver Cancer: Large-Scale Machine Learning from Ethnically Diverse Multicentre Longitudinal Healthcare Records

Background Chronic hepatitis B (CHB) is a leading global cause of cirrhosis and hepatocellular carcinoma (HCC). The NHS has recently identified more CHB cases in the UK, necessitating dedicated resources for patient care. As treatment guidelines expand to include more patients, understanding treatment responses for optimal outcomes (functional cure) is crucial. Current risk scores, developed from single-ethnicity populations and lacking consideration of comorbidities, are inadequate for the UK's diverse CHB patients. Extensive patient records from the NHS can be used to develop better prediction models, and AI and machine learning have great potential to analyse complex large datasets. Aims & Objectives This research aims to: Convert CHB medical reports into structured data and understandable text. Understand HBsAg dynamics and predict functional cure. Assess the impact of comorbidities and predict cirrhosis and HCC early, considering comorbidities. Create early versions (prototypes) of software tools based on developed models. Methods Work Package (WP)1: Develop scalable natural language processing (NLP) approaches using large language models to convert medical reports into structured data incorporating domain knowledge and understandable summaries. WP2: Use latent class mixed modelling to identify distinct HBsAg trajectory classes. Develop functional cure prediction models using machine learning based on the transformers and other deep learning techniques. WP3: Identify important comorbidities using polytomous latent class analysis mixture modelling and investigate their impact. Explore various machine learning paradigms for predicting cirrhosis and HCC, including prediction using data at presentation, time to cirrhosis and HCC prediction using time-varying data, and multimodal learning to integrate structured data, free-text reports, and knowledge graphs (incorporating domain knowledge). WP4: Develop tool prototypes using Shiny, preparing for development of practical tools/applications in patient care. Incorporate feedback from patients, carers, clinicians, and healthcare software developers. Timelines for delivery WP1 (Years 1-2): Deliver NLP approaches for processing CHB medical reports. WP2 (Years 2-3): Deliver HBsAg trajectory classes and classification model, and functional cure prediction model. WP3 (Years 2-4): Deliver comorbidity clusters and their impact, and prediction models for cirrhosis and HCC. WP4 (Years 4-5): Deliver three software prototypes. Anticipated impact & dissemination Advancing an analytical framework for multi-site healthcare records to answer research questions. Tackling the burden of CHB in the UK. Identifying patient subsets most at risk for cirrhosis and liver cancer will enable implement targeted treatment interventions or screening programs. Conversely, identifying patients who are likely to achieve functional cure without additional interventions- with significant cost savings to the NHS. Understand co-morbidities in CHB will enable holistic treatment approaches for the patient (the current focus on HBV infection in isolation). Streamline CHB medical records analysis. Foster new collaborations with Pharma industry, as evidenced by ongoing interest, so contributing to economic growth. Publications in academic journals, supporting scientific progress. Inform public health policies for HBV in the UK. Software prototypes can be developed for NHS use to support clinical decision-making, with a long-term vision to standardise care and improve care service efficiency. Disseminate findings to patients, carers, clinicians/NHS, academics, and the public through demonstrations, conferences, charities, publications, and social media.