Poster L11 - Multimodal Data Model for Early Diagnosis and Management of Non-Alcoholic Fatty Liver Disease (NAFLD)

Introduction: Non-alcoholic fatty liver disease (NAFLD) is a condition primarily characterized by the abnormal accumulation of liver fat, the disease affects an estimated 25% of the population. When left unchecked, it progresses into non-alcoholic steatohepatitis (NASH), a stage where fibrosis starts to develop in the liver alongside a higher accumulation of liver fat. Both conditions are reversible when diagnosed before the disease progresses into liver cirrhosis. This study investigates the impact of multi-modal data to create a model that can help with early diagnosis and management of NAFLD.

Materials and

Methods: We utilized the data from Dr. Reboussin’s 2021 study [1], which focused on analyzing the effects of a habitual diet (n=881) versus an avocado diet (n=874) on multimodal data from a patient. The variables that were acquired were categorized into health eating index (HEI) variables, Social determinants of health (SDoH), Metabolic, Electronic Health Records (EHR), and Biomarkers. The statistical methods employed to develop a model include logistic regression, variable importance analysis, and ROC curve analysis to evaluate the accuracy and significance of variables across the demographic groups and gender.

Results, Conclusions, and Discussions: Our findings rank the importance of these variables subsets as follows: Biomarker Variables (Accuracy ~0.72, AUC ~0.80), Metabolic Variables (Accuracy ~0.66, AUC ~0.71), EHR Variables (Accuracy ~0.65, AUC ~0.71), and both SDoH and HEI Variables (Accuracy ~0.55 Accuracy, AUC ~0.55). By focusing on a specific set of variables, we aim to improve understanding of the disease and the interaction between the available and unavailable variables that are affected by NAFLD.

The initial findings indicate that the biomarker variables possess the highest predictive accuracy for NAFLD progression, with a model AUC of 0.80±0.03, and an accuracy of 0.72±0.03. Another important finding is that even when reducing the number of features utilized in the model to the top 10% of features which is 11 different variables in the dataset the accuracy over the testing dataset improved, with an accuracy of 0.74, and an AUC of 0.81. This methodology also allowed for a bootstrapped approach to determine key variables influencing NAFLD progression.

When comparing the results to similar models like the one reported by Pan et al. (2023) [2], which exhibited and UAC from 0.777 to 0.826 depending on their dataset. The model demonstrated an advantage since it utilized a smaller set of variables to predict for the NAFLD outcome. From the

The study highlights the significance of using specific biomarkers and metabolic variables for accurate prediction of NAFLD and demonstrating the potential for a targeted approach to variable collection. To improve sensitivity and specificity of NAFLD prediction, as well as the need for further testing in patients that are found to have high chances of the disease.

Acknowledgements (Optional):