Ensemble Learning Stratification of Liver Histologic Fibrosis using Multi-Modal MRI and EHR Data in Pediatric and Adult Patients

Motivation: Chronic liver diseases can lead to variable liver fibrosis. Percutaneous liver biopsy is the clinical standard; however, it has limitations. Goal(s): Develop an ensemble learning model to stratify liver fibrosis using non-invasive clinical MRI and EHR data. Approach: In this retrospective, multi-site study, we retrieved liver tissue specimens, multi-modal MRI, and EHR data. Using biopsy-derived liver fibrosis stages as reference, a stacking ensemble learning model was trained to classify a patient into the no/mild or advanced liver fibrosis. Results: Using multi-modal MRI and EHR data, our model achieved an AUROCs of 0.69 - 0.75 on liver fibrosis stratification in internal and external cohorts. Impact: Our study demonstrated that an ensemble learning model had a moderate performance in stratifying liver fibrosis using clinical multi-modal MRI and EHR data. With further tuning, it provides a potential non-invasive means for monitoring and screening of liver fibrosis.