Hospital readmissions occur frequently, they are expensive, and a high number of them decrease the institution's perceived quality. In machine learning, the readmission task aims to predict a patient's risk of readmission. Several solutions have been proposed using Electronic Health Records (EHR). EHRs have all information related to an admission: lab tests, free notes, demographic data, and International Classification of Diseases (ICD) codes. ICD is an international standard that provides codes for diagnoses and procedures. Initial solutions to the readmission problem used ICD codes as features via categorized representations or representations learned from their local context. Recent solutions in-gest all EHR data, adding unnecessary complexity. In this research, we explore new representations for ICD codes. We leverage their text descriptions using Natural Language Processing techniques and their ontological representation through graph embedding al- gorithms. We provide benchmarks for the readmission task using a novel dataset from a large Chilean hospital, with a clear evaluation framework, and achieve results comparable with the state of the art. Generated ICD mappings and representations are publicly available.
How to join?
The defense will be in spanish. If you want to attend, please send an email to Tamara (email@example.com) at least one day before the defense.