A framework for measuring and mitigating bias in electronic health record data: a case study in childhood obesity incidence prediction

Elizabeth Campbell

doi:10.17918/00001094

Electronic Health Records (EHRs) are valuable sources of healthcare data with tremendous potential to improve patient care and reduce financial costs. Advanced data mining and analytical methods are necessary to obtain, transform, and analyze EHR data for secondary uses such as clinical and health policy research, and machine learning methods are key to addressing challenges in secondary EHR uses. However, developing machine learning models equitably is a significant challenge. As machine learning tools become more ubiquitous in healthcare research and clinical practice, it is critical that predictive models are developed that work to alleviate, rather than exacerbate, existing disparities in the healthcare system. Bias may be introduced into a machine learning model through the data used to train the model, algorithm selection, or in how a model is implemented in an applied setting. In this work, we present a framework for assessing and alleviating bias in dataset development for a classification model derived from EHR data. Classification, which refers to inputting observations from data into categories, is a common machine learning task and a method that is increasingly used in medical research. Classification may be used to predict future health outcomes, identify patients who are at risk for developing a condition, and to make clinical diagnoses. In this work, we present a case study in childhood obesity incidence prediction where a classification model is developed that aims to predict future obesity among pediatric patients based on demographic data and the presence of clinical diagnostic patterns in their medical history. We use easily accessible EHR data that may be obtained to predict other outcomes of interest, and present a method to evaluate data to develop classifiers that may be used for prediction of other chronic diseases (such as heart disease, diabetes, and asthma). First, temporal pattern mining is applied to EHR data to identify common diagnostic patterns that surround pediatric obesity incidence. Next, a dataset is developed for the study cohort, that indicates whether or not patients had the temporal condition patterns that were previously identified and demographic information (such as race/ethnicity, gender, etc.) obtained from their EHR. Then, classification models are trained and tested using this dataset. We develop a framework for assessing bias in EHR-derived datasets, and present its application in our pediatric obesity incidence case study. We conclude by discussing implications of our work in healthcare research and outlining directions for future research, including assessing and alleviating bias in algorithm selection and model development.

A framework for measuring and mitigating bias in electronic health record data: a case study in childhood obesity incidence prediction

Files and links (1)

Abstract

Metrics

Details

A framework for measuring and mitigating bias in electronic health record data: a case study in childhood obesity incidence prediction

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media