Andrew Oriti
Research Mentor: Matthew Patrick
Mentor Department: Department of Dermatology, Medicine
Author(s): Andrew Oriti, Matthew Patrick
Session: Session 4 (1:00 PM – 1:50 PM)
Presentation Type: Poster 139
Abstract
Electronic medical data contains a wide variety of predictive information that can be utilized for disease prediction in patients. By building a disease network we can better identify which diseases may be related and often develop alongside one another. In this work, we use python scripting and libraries to build a disease network using publicly available word embeddings from a combined source of 100,000+ medical records, insurance claims, and journal articles. Links are created between disease CUIs (Concept Unique Identifier) if the calculated vector cosine similarity value between their respective embeddings matrix meets the provided threshold. After testing and graphing the threshold values ranging from 0.1 to 1.0, we calculated the optimal threshold to use by finding where the maximum curvature occurs, which we decided was a value of 0.35. At this threshold weight of 0.35, a total of 420,830 connections exist giving an average of 67.5 connections per node. Using this disease network, we examined the 6200+ diseases found within the word embedding data using different network analysis methodologies (such as random walkers and clustering) to calculate which diseases are most similar to a given disease node. In order to further test these results, we wrote an algorithm to run the random walker simulation multiple times to graph the variance between trials with a box plot. Then by removing all diseases with the same icd10 group as the disease we are testing, we were able to focus on diseases that are most likely comorbidities rather than other relationships. Our results aim to provide this code as a publicly available tool, in the form of a python library, that improves upon current medical network analysis to allow for better patient diagnoses.
