Investigating Comorbidities from Clinical Texts: A Propensity Score Approach

Abstract

Understanding comorbidity patterns is crucial for improving patient outcomes and optimising healthcare strategies. In this study, we propose an approach to detect comorbidities of two diseases from clin-ical discharge notes. To account for the complexities of textual data, we summarise the information through propensity scores, which repre-sent the probability of receiving a certain diagnosis conditional on the extracted text. These scores are then used as covariates in a logistic regression model to explore the association between diseases. Specifi-cally, we compare models trained on TF-IDF weighted document-term matrices and text embeddings, employing LASSO regression, XGBoost, and multilayer perceptrons (MLP). Our results, obtained by applying this method to study the association between diabetes and Chronic Kid-ney Disease, demonstrate the potential of Natural Language Process-ing (NLP) and machine learning techniques in advancing observational healthcare research.