Predicting baby feeding method from unstructured electronic health record data

Abstract

Obesity is one of the most important health concerns in United States and is playing an important role in rising rates of chronic health conditions and health care costs. The percentage of the US population affected with childhood obesity and adult obesity has been on a constant upward linear trend for past few decades. According to Center for Disease control and prevention 35.7% of US adults are obese and 17% of children aged 2-19 years are obese. Researchers and health care providers in the US and the rest of world studying obesity are interested in factors affecting obesity. One such interesting factor potentially related to development of obesity is type of feeding provided to babies. In this work we describe an electronic health record (EHR) data set of babies with feeding method contained in the narrative portion of the record. We compare five supervised machine learning algorithms for predicting feeding method as a discrete value based on text in the field. We also compare these algorithms in terms of the classification error and prediction probability estimates generated by them.