Abstract
The current state of healthcare is prime for the development of new methodologies utilizing the wealth of electronic data regarding patient history to measure and track quality of care automatically. According to the Institute of Medicine, 30% of US health care services are inappropriate or wasteful, only 55% of needed health care services are delivered, and care can vary significantly by geographic location. In light of these facts, the importance of developing tools to assess whether patients are receiving consistent and quality care is paramount. This study addresses an approach to measuring treatment variation between congestive heart failure patients based on their clinical orders and diagnoses. The data used for this research consists of procedure and medication order records collected using the Electronic Medical Record System (EMR) in place at the University of Virginia Health Center. Through the use of a vector space model, we provide a visual and quantitative method for measuring treatment similarity between patients based on treatments received. For this study, we investigated the variation of care provided to a selection of congestive heart failure patient groups including inpatient visits, outpatient visits, and combined inpatient and outpatient visits. Additionally we investigated patients who were diagnosed with one of the following comorbidities: chronic kidney disease, diabetes mellitus, or chronic obstructive pulmonary disease. Calculating similarity distributions for the general inpatient, outpatient, and combined IP/OP groups revealed that patients receive less varied care as an inpatient than as an outpatient. Similarity distributions calculated for comorbidity groups suggest that treatment variation decreases for these comorbidity groups as compared to the general CHF population. Additionally, our results show that CHF patients with COPD or chronic kidney disease are divided in such a way that a distinct portion of patients receive care to that similar of CHF patients with diabetes and a distinct portion of patients receive care which is dissimilar from the treatment provided to CHF patients with diabetes. Finally, as a first step toward validating our methods, two subject matter experts were asked to provide their own assessment of the treatment similarity between a selection of patients and chosen reference patients. Taking into account a Kappa statistic of 0.284 obtained for inter-rater agreement, our results suggest that our approach to measuring care variation does provide some insight on patients which are being treated with lesser or greater similarity. Thus, the techniques discussed in this study may support a framework for monitoring treatment variation automatically in electronic health record systems.
