Using a machine learning model to risk stratify for the presence of significant liver disease in a primary care population

Background: Current strategies for detecting significant chronic liver disease (CLD) in the community are based on the extrapolation of diagnostic tests used in secondary care settings. Whilst this approach provides clinical utility, it has limitations related to diagnostic accuracy being predicated on disease prevalence and spectrum bias, which will differ in the community. Machine learning (ML) techniques provide a novel way of identifying significant variables without preconceived bias. As a proof-of-concept study, we wanted to examine the performance of nine different ML models based on both risk factors and abnormal liver enzyme tests in a large community cohort. Methods: Routine demographic and laboratory data was collected on 1,453 patients with risk factors for CLD, including high alcohol consumption, diabetes and obesity, in a community setting in Nottingham (UK) as part of the Scarred Liver project. A total of 87 variables were extracted. Transient elastography (TE) was used to define clinically significant liver fibrosis. The data was split into a training and hold out set. The median age of the cohort was 59, mean body mass index (BMI) 29.7 kg/m², median TE 5.5 kPa, 49.2% had type 2 diabetes and 20.3% had a TE >8 kPa. Results: The nine different ML models, which included Random Forrest classifier, Support Vector classification and Gradient Boosting classifier, had a range of area under the curve (AUC) statistics of 0.5 to 0.75. Ensemble Stacker model showed the best performance, and this was replicated in the testing dataset (AUC 0.72). Recursive feature elimination found eight variables had a significant impact on model output. The model had superior sensitivity (74%) compared to specificity (60%). Conclusions: ML shows encouraging performance and highlights variables that may have bespoke value for diagnosing community liver disease. Optimising how ML algorithms are integrated into clinical pathways of care and exploring new biomarkers will further enhance diagnostic utility.