Bayesian Classification of Triage Diagnoses for the Early Detection of Epidemics

May 2003

Proceedings of the Sixteenth International Florida Artificial Intelligence Research Society Conference (FLAIRS 2003)

Volume

Proceedings of the Sixteenth International Florida Artificial Intelligence Research Society Conference (FLAIRS 2003)

All Papers

The distribution of illnesses reported by emergency departments from hospitals in a region under surveillance is particularly informative for the early detection of epidemics. The most direct source of data for construction of such a distribution is the final diagnoses of patients being seen in the emergency departments, but the delay in their availability impinges on the requirement that detection be timely. Free-text descriptions of patients’ symptoms, called triage diagnoses, and ICD-9 values that encode the symptoms are entered when patients are admitted and, consequently, are timelier sources of data. An experiment to evaluate the accuracy of Bayesian classification of triage diagnoses into syndromes (i.e., illness categories) was performed, resulting in areas under the ROC curve (AUC) between .80 and .97 for the various syndromes. The classification accuracies using triage diagnoses surpass the classification accuracies using ICD-9 codes reported by previous studies. Triage diagnoses, therefore, are a more accurate source of data than ICD-9 codes for the early detection of epidemics.