Data mining information from electronic health records produced high yield and accuracy for current smoking status
T. Katrien J. Groenhof, Laurien R. Koers, Enja Blasse, Mark de Groot, Diederick E. Grobbee, Michiel L. Bots, Folkert Wouter Asselbergs, A. Titia Lely, Saskia Haitjema, Wouter W. van Solinge, Imo E. Hoefer, Saskia Haitjema, Mark de Groot, Enja Blasse, Folkert Wouter Asselbergs, Hendrik M. Nathoe, Gert‐Jan de Borst, Michiel L. Bots, Mirjam I. Geerlings, M H Emmelot · Journal of Clinical Epidemiology · 2019
OBJECTIVES: Researchers are increasingly using routine clinical data for care evaluations and feedback to patients and clinicians. The quality of these evaluations depends on the quality and completeness of the input data. STUDY DESIGN AND SETTING: We assessed the performance of an electronic health record (EHR)-based data mining algorithm, using the example of the smoking status in a cardiovascular population. As a reference standard, we used the questionnaire from the Utrecht Cardiovascular Cohort (UCC). To assess diagnostic accuracy, we calculated sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV). RESULTS: We analyzed 1,661 patients included in the UCC to January 18, 2019. Of those, 14% (n = 238) had missing information on smoking status in the UCC questionnaire. Data mining provided information on smoking status in 99% of the 1,661 participants. Diagnostic accuracy for current smoking was sensitivity 88%, specificity 92%, NPV 98%, and PPV 63%. From false positives, 85% reported they had quit smoking at the time of the UCC. CONCLUSION: Data mining showed great potential in retrieving information on smoking (a near complete yield). Its diagnostic performance is good for negative smoking statuses. The implications of misclassification with data mining are dependent on the application of the data.