Development of a Machine Learning Model to Integrate Clinical, Laboratory, Sonographic, and Elastographic Data for Noninvasive Liver Tissue Characterization in NAFLD

点击翻译按钮获取中文摘要

Abstract Non-alcoholic fatty liver disease (NAFLD) is exceptionally common, with an estimated one hundred million afflicted people in the United States. Detection and risk stratification of this very common disease remains a major challenge. Despite recent advances, including development of numerous therapeutic agents presently in phase 2 and 3 trials, NAFLD remains a silent disease in which the vast majority of patients accumulate progressive liver damage without signs or symptoms and, undiagnosed, receive no medical care. The NAFLD patients at highest risk of cirrhosis are those with moderate or greater liver fibrosis at the time of diagnosis, a group of patients who are described as having high risk non-alcoholic steatohepatitis (hrNASH). The current reference standard for identifying people with hrNASH is liver biopsy, which is expensive, invasive, and limited by interobserver variability. The focus of this project is to develop and validate low cost non-invasive diagnostic technology to diagnose hrNASH. We propose to accomplish this in three Specific Aims. First, we will expand and annotate an existing database of patients with chronic liver disease from 328 subjects to 1,000 subjects, ~40% of whom will have NAFLD. The database will contain ~20,000 images (~10,000 ultrasound elastography images and ~ 10,000 conventional ultrasound images) and multiple demographic and clinical data points for each subject (a total of ~30,000 clinical, laboratory, and demographic data points). We have previously developed advanced image processing techniques to make ultrasound elastography more accurate and less variable. We will use this large database to develop, customize and refine our image processing techniques for NAFLD evaluation (Aim 1), with the goal of improving ultrasound elastography diagnosis of hrNASH. Second, we will combine conventional ultrasound elastography imaging, conventional ultrasound imaging, our advanced image analysis techniques, and the demographic, clinical, and laboratory data in a machine learning model to predict hrNASH and will compare the performance of our predictive model with the FIB4, a widely-used blood test-based prediction rule (Aim 2). Third, we will validate our predictive model in an independent prospective cohort of NAFLD subjects undergoing biopsy for NAFLD risk stratification (Aim 3). We hypothesize that the combination of image processing-enhanced elastography and conventional ultrasound imagery combined with demographic, clinical, and laboratory data will have greater predictive power for hrNASH than clinical or sonographic data alone. The proposed predictive models have the potential to (1) reduce the number of liver biopsies performed for hrNASH detection, (2) facilitate recruitment for clinical trials of NAFLD therapeutics, and (3) improve care quality for the most common liver disease in the United States.