C-C3-04: Using Electronic Health Record Data to Predict Heart Failure Diagnosis

Abstract

Background: Heart Failure (HF) is one of the most common and serious progressive illnesses among elderly patients. It is usually detected at a relatively advanced stage, after irreversible damage has occurred. Early detection offers the potential to substantially reduce patient disability and health care costs. Objective: To develop a novel detection strategy, making use of longitudinal electronic health record (EHR) data to create an HF early detection prediction model. Methods: All data for this study were obtained from Geisinger Clinic’s EHR among patients who had a primary care provider. A prediction model was developed using a nested case control study design, where HF cases diagnosed between 2003–2006 were identified and controls were randomly selected matched on sex, age, and clinic. We used conditional logistic regression to model the relation between EHR data and detection of HF 6+ months and 18+ months before the actual date of diagnosis. Variables for the model included diagnoses, the most recent lab and clinical (e.g., SBP, DBP, pulse pressure) measures, medication orders and ambulatory care use in the previous two years, and smoking status. Data were only used if they occurred in the record either 6+ months or 18+ months before the diagnosis date, depending on the specific model. Model results were validated by combining a bootstrap re-sampling approach with a backwards elimination selection method. Results: We identified at least one matching control for 2,239 of the 2,764 cases; 9 or 10 controls were identified for 81% of the cases. A total of 24,249 controls were selected. The model for detecting HF 6+ months before usual diagnosis had an AUC of 0.80; the parallel model for detecting HF 18+ months before usual diagnosis had an AUC of 0.75 (95% CI: 0.73, 0.79). The AUC findings were similar for separate models completed on systolic HF and diastolic HF. Conclusions: In practice, clinicians do not have the time or ability to process seemingly disparate data points over a series of visits that might suggest a preclinical HF for a given patient. Our analysis of EHR data indicate that HF can be detected 6 or more months before usual diagnosis with good AUC and high specificity. These findings suggest that routine evaluation of EHR data may be useful in screening for patients at high risk of HF, creating numerous opportunities for early and aggressive intervention and potentially altering the natural history of heart failure for many patients.