BACKGROUND: Nebraska conducts influenza-like illness (ILI) surveillance by using Electronic Surveillance System for Early Notification of Community-Based Epidemics (ESSENCE), which applies a text-processing algorithm to chief complaints from emergency department (ED) electronic health records (EHRs). Available data differ by what can be entered into an EHR and what clinicians enter. We assessed whether using different EHRs influences ILI detection because of chief complaint field length.
METHODS: We randomly selected 622 ED visits during January 1–March 31, 2015, from 2 hospitals with different EHR types; 1 allowed shorter character entry and 1 allowed longer character entry. An ILI standard was established if patients’ EHR records upon chart review documented fever (>100ºF) and cough or sore throat with no alternative diagnosis. We applied the ESSENCE ILI algorithm and assessed agreement. We performed logistic regression to assess whether agreement was dependent on chief complaint word count, and if word count was a reliable indicator of source EHR. We compared the predictive accuracy of ESSENCE by using receiver operator characteristic (ROC) curves. Youden’s index was calculated to identify optimal cutoff points.
RESULTS: Short and long chief complaint EHR medians were 3 words (range: 1–9, n = 322) and 17 words (range: 1–45, n = 300), respectively. Predictive accuracy of ESSENCE varied by word count (likelihood ratio of 15.4, P <0.0001). Word count was demonstrated to be a valid indicator of EHR (Wald chi-square of 345.3, P<0.001). Separate ROC curves were compared and demonstrated similar concordance; 0.597 for the short EHR and 0.591 for the long EHR. For the short EHR, Youden’s index identified an optimal probability of 0.68, corresponding to a word count of 7 and identifying 73% of events as correct. For the long EHR, Youden’s index identified an optimal probability of 0.88, corresponding to a word count of 15 and identifying 78% of events as correct.
CONCLUSIONS: Word count differs by EHR and affects predictive accuracy of ESSENCE for ILI detection. Longer chief complaints EHR performed better than short chief complaints. EHR vendors can consider providing data-entry fields to capture longer chief complaints. Additionally, providers can record chief complaints of 7–15 words to enhance ESSENCE ILI detection accuracy from EHRs.