Machine Learning Analysis of Longitudinal Laboratory and Imaging Data to Personalize the Prediction of Severe Complications in Veterans with Crohn s Disease
Abstract
Crohn s disease is a condition affecting close to 1 million patients in the United States where inflammation of intestines and colon cause symptoms of diarrhea, bleeding, and abdominal pain. Furthermore, despite many patients not having symptoms, ongoing chronic inflammation results in irreversible damage and severe complications. These complications often result in hospitalization, with 60% of patients needing surgical removal of unsalvageable intestine. The best way to avoid severe complications is to suppress intestinal inflammation, thereby preventing chronic damage. However, (1) effective therapies suppress the immune system, increasing the risk of infections and cancers; (2) they need to be taken indefinitely; (3) they are expensive ($15,000-$50,000 annually); (4) they don t help everyone; and (5) upwards of 40% of patients will never experience complications. Therefore, it is important to determine (a) who will develop future complications, (b) when will the complications occur, and (c) how much will medical therapy reduce the risk of complications for the individual patient. Colonoscopy, CT (computed tomography) scans, blood and stool studies are unable to sufficiently and consistently describe Crohn s disease accurately enough to predict the risk of future complications. The inability to reliably predict future complications and the strength of treatment needed to avoid complications are critical problems in Crohn s disease management. Analytic morphomics is a novel approach developed by our group where existing CT scans undergo image analysis for precise measurement of bowel thickness, enhancement, dilation, length of disease, and even changes outside the intestine in Crohn s disease. While morphomics is unable to replace a radiologist, it can accurately quantify features of disease, rather than simply grade by mild, moderate, or severe criteria. Morphomic measurements are also likely to be more reproducible than measures between radiologists. Finally, the time needed for radiologists to make multiple bowel measurements in inflammatory bowel disease (IBD) is impractical for clinical use; morphomic analysis is automated and relatively fast. Another problem is that using high volume complex data, which changes over time, to predict future events poses challenges when using traditional statistics. Other industries like marketing, meteorology, and agriculture have used machine learning to sift through different classes of transaction, weather, and crop yield data to find patterns that inform decisions. We believe that modern machine learning analysis techniques are best suited to find relevant, perhaps unexpected, relationships between disease activity measurements and the occurrence of severe Crohn s disease complications. We hypothesize that machine learning analysis of clinical, laboratory, and analytic morphomic data will predict if and when future complications will occur in Veterans with Crohn s disease. In preparation for developing predictive models, we created the Veteran Affairs IBD cohort, which includes all electronic health record data from 30,000 de-identified Veterans with IBD. The first aim of this project will be to refine the bowel-specific analytic morphomics tools, perform quality assurance testing, and compare morphomic measurement to those of a board-certified radiologist. This work will utilize a test set of CT scans from 300 Veterans with Crohn s disease and leverage the dedicated resources of the University of Michigan Morphomic Analysis Group. In Aim 2, we will develop machine learning algorithms using all demographic, clinical, medication, laboratory, and morphomic data in the Veteran Affairs IBD cohort. We will process approximately 1500 CT scans from the VISN 11 Network using optimized analytic morphomics algorithms. Datasets will be reformatted into several formats to explore multiple machine learning strategies in the process of developing the best model to predict
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 31, 2017
- Source ID
- W81XWH1610682
Entities
People
- Akbar K Waljee
Organizations
- United States Army
- Veterans Education and Research Association of Michigan