Hepatitis B Virus Infection in Nigeria: Disease Prevalence and Machine Learning Intervention for Early Detection

Abstract

In Nigeria and globally, hepatitis B virus (HBV) is associated with substantial morbidity and mortality, posing a significant threat to public health. An estimated 296 million people live with this virus, and 90% of infected people are unaware of their infection status and risk infecting others. Immunoassay is the current gold standard for hepatitis B surface antigen detection, involving assays that are prohibitively expensive and require specialised facilities. However, access to this specialised test in resource-constrained settings like Nigeria is limited, particularly for rural and isolated laboratories.

In this thesis, I contribute to understanding the epidemiology of HBV in Nigeria and provide evidence on how machine learning (ML) could enhance clinical decision making for early detection of the virus.

The thesis first delves into a comprehensive meta-analysis of HBV data in Nigeria to provide epidemiological estimates, and critical insights into disease burden and regional variations. The outcome reveals a high prevalence of HBV infection in Nigeria (9.5%), with significant sub-national variations in regions (>12%). These findings corroborate the need for an appropriate intervention that enhances disease detection, optimises patient care and contributes to the World Health Organization's efforts to eliminate HBV as a public health threat by 2030.

To explore the potential of ML as a public health intervention for HBV problem in Nigeria, I conducted two other studies. First, I synthesised evidence on the development and quality of existing ML models for predicting clinical outcomes associated with blood-borne viral infections, including HBV. Whilst promising approaches were identified for the ML prediction models, their lack of robust validation, interpretability, and reproducibility, hampered clinical relevance. To address these limitations, I develop a checklist for the development of ML innovations to guide safe implementation and translation into routine clinical workflow.

The exploration of ML for clinical-decision making provides valuable insights for the third study, which focussed on the development of a ML decision support system for early detection of HBV in Nigeria via routine blood test data. Through the ML interrogation of routine pathology data, I identified aspartate aminotransferase, white blood count, alanine aminotransferase, albumin, and patient age as the most powerful predictive markers of HBV infection, with established decision thresholds. These markers were then used to develop HepB LiveTest, a ML-enabled diagnostic model (translated into a web-accessible app) that predicts a patient's HBV infection status, with a suite of real-time decision support.

Recognising the need for external validation of clinical prediction models, I conducted an additional study to validate the robustness and clinical portability of HepB LiveTest in two external patient cohorts, including an Australian cohort, to inform evidence for its cross-site transportability. From the results, HepB LiveTest proves to be highly accurate for predicting HBV infection, with optimal validity in an external validation cohort of hospital-based Nigerian patients. However, it shows limited clinical validity in an Australian patient cohort, a population with a low incidence of HBV infection.

This thesis highlights the high prevalence of HBV in Nigeria, and provides a platform to offer timely decision support to clinicians and drive significant reductions in HBV prevalence through improved access to early screening. The findings provide valuable evidence to support public health efforts in tackling HBV and optimise the quality of life for millions of infected patients, particularly in underserved populations. Show more