P82 An interim report on the development of a novel algorithm for the prediction of fatty liver disease in healthy and patient trial volunteers

BackgroundNon-alcoholic fatty liver disease (NAFLD) affects approximately one in four of the global adult population, and ranges in severity from benign fatty liver infiltration, to hepatitis, cirrhosis, hepatocellular carcinoma, and death. NAFLD has important implications for clinical trial volunteers as an occult co-morbid condition – there is evidence that NAFLD modulates drug metabolism, with studies1 suggesting that Grade 3–4 liver reactions are four times more common in healthy volunteers with probable NAFLD than without.This research sought to develop a non-invasive, low-cost tool, utilising supervised machine learning techniques, to predict NAFLD in a clinical trial population.MethodsThis is an ongoing, observational cross-sectional study with a total of 1500 clinical trial volunteers attending the unit for a single day of biomarker assessment, including bioimpedance vector analysis (visceral fat%, total body fat% and skeletal muscle%), anthropometric measurement (BMI, waist circumference), and laboratory bloods (including HBA1c, liver enzymes and WCC). A FibroScan is performed as a pragmatic gold standard ‘outcome’ for NAFLD.In this interim analysis of 570 volunteers, a logistic regression model was trained for ability to predict the outcome of the fibroscan from a data subset containing: Age, Sex, Ethnicity/Race, Height, Weight, BMI, Waist Circumference, Body Fat Percentage. The programming language R was used to build the model.The data was divided into a training set of 399 volunteers (70%) used to build the model, and 171 volunteers (30%) used to validate its predictive accuracy.ResultsDemographics of the included volunteers were assessed as representative of a typical early-phase population. An initial fitting of individual predictive variables into the model showed that Age, Weight, BMI, Waist Circumference and Body Fat Percentage were highly statistically significant.On fitting a full model, Age, Ethnicity/Race, and Waist Circumference were ultimately included. Weight, Height, and BMI were removed due to a high degree of collinearity, problematic to model performance.When validating the model against the volunteer test dataset, it achieved a 82.4% predictive accuracy in identifying FibroScan result.Abstract P82 Figure 1ConclusionThis preliminary model shows promising results as an early-stage clinical tool to accurately predict steatosis amongst trial volunteers.This is a group where the effective detection of concomitant, occult fatty live