Modeling the human classification of acute decompensated heart failure using abstracted medical record with machine learning

Heart failure (HF) is a clinical syndrome in which the heart is not able to properly pump blood because of structural and functional defects, resulting in the body not getting enough blood. It affects millions of people in the United States and often leads to hospitalization or mortality. Acute decompensated heart failure (ADHF) is a sudden worsening of HF symptoms and is a powerful predictor of readmission for HF and death of patients with chronic HF post-discharge. In this thesis, we used data from the Atherosclerosis Risk in Communities (ARIC) study’s community surveillance of heart failure hospitalizations to develop machine learning classification models to accurately classify if a patient did have or did not have ADHF. We used abstracted hospital records and ADHF diagnosis, done by clinician review, to train the classifiers to identify ADHF cases. After data preparation through imputation and handling collinearity in the data, we had 2,925 records in our training set and 116 records in our test dataset. Data preparation, cross-validation, model creation, and data analysis were all done in R, and we created a decision tree using the rpart package and a boosted decision tree using the adabag package. Using these models, we observed classification accuracy rates of approximately 75% in the decision tree model and 79% in the boosted decision tree model. These rates were fairly consistent with those found in literature of machine learning models that were used to classify general HF and general heart disease cases. The success of our models, relative to those in literature, demonstrate the potential for machine learning to help identify ADHF cases among HF-related hospitalizations in the clinical setting.Bachelor of Science in Public Healt