Classification of patients by severity grades during triage in the emergency department using data mining methods

Objective To test the feasibility of classifying emergency department patients into severity grades using data mining methods. Design Emergency department records of 402 patients were classified into five severity grades by two expert physicians. The Naïve Bayes and C4.5 algorithms were applied to produce classifiers from patient data into severity grades. The classifiers' results over several subsets of the data were compared with the physicians' assessments, with a random classifier, and with a classifier that selects the maximal-prevalence class. Measurements Positive predictive value, multiple-class extensions of sensitivity and specificity combinations, and entropy change. Results The mean accuracy of the data mining classifiers was 52.94 ± 5.89%, significantly better (P < 0.05) than the mean accuracy of a random classifier (34.60 ± 2.40%). The entropy of the input data sets was reduced through classification by a mean of 10.1%. Allowing for classification deviations of one severity grade led to mean accuracy of 85.42 ± 1.42%. The classifiers' accuracy in that case was similar to the physicians' consensus rate. Learning from consensus records led to better performance. Reducing the number of severity grades improved results in certain cases. The performance of the Naïve Bayes and C4.5 algorithms was similar; in unbalanced data sets, Naïve Bayes performed better. Conclusions It is possible to produce a computerized classification model for the severity grade of triage patients, using data mining methods. Learning from patient records regarding which there is a consensus of several physicians is preferable to learning from each physician's patients. Either Naïve Bayes or C4.5 can be used; Naïve Bayes is preferable for unbalanced data sets. An ambiguity in the intermediate severity grades seems to hamper both the physicians' agreement and the classifiers' accuracy.