Exploring Electronic Medical Records for Public Health Surveillance

Introduction: Early syndromic surveillance systems focused on the use of traditional data sources, such as emergency department visit counts, chief complaint reports, and over-the-counter medications sales. However, these sources can be limited by data quality and lack of links between data sources. The objective of our research is to explore methods to effectively integrate both clinical and syndromic surveillance data and develop decision support algorithms that provide information to both public health and clinical users proactively. Our hypothesis is that information fusion models based on Bayesian Networks that utilize both traditional syndromic data and new linked data sources from Electronic Medical Record (EMR) will enhance outbreak recognition performance and increase public health user’s situational awareness.

 Method: The Intelligent Severity Score Estimation Model (ISSEM) was built to calculate a score that estimates a patient’s disease-severity. The model is structured so that the inference process reflects the decision-making logic of public health experts. ISSEM calculates the severity score based on numbers from respiratory ICD9 encounters; laboratory, radiologic, and prescription-therapeutic orders; chronic disease evidence; and the provider’s general practice-behavior patterns from the EMR. Adjusted counts for daily visits are calculated based on the ISSEM severity scores for patients visiting medical facilities. The Population Health Bayesian network fuses the anomalous number of daily visits with syndromic and EMR data to estimate the probability of an outbreak.

 Results: ISSEM sensitivity was determined from 200 randomly selected patients with upper- and lower-respiratory tract ILI; specificity was determined from 300 randomly selected patients with URI only. ISSEM performance demonstrated 93.5% sensitivity and 77.3% specificity across all age groups.

 Conclusion: Our preliminary results show that the severity estimation model can potentially identify patients at the early stage of developing lower respiratory complications when the number of cases is relatively small.