Enhancing aspects of Thai chief complaint classification Performance

Abstract

In this paper, we describe the aspects affecting in our experimental results of classifying Thai chief complaint (ThCC) into ICD-10 code. By merging our proposed Thai word separator to machine learning-based classifiers, ThCC have been converted into ICD-10 code which stands for International Classification of Diseases, Tenth Revision, and is a standard code used by physicians and other healthcare professionals to identify all diagnoses, signs and symptoms. At the beginning of experiments, the dataset from the sign and symptom description ranged in group R00 to R69 of ICD-10 have been used for training the classifiers. Subsequently the classifiers have been applied to the test dataset represented by 150 chief complaint cases in order to assign the related ICD-10 codes, and to evaluate classification accuracy. The experiment achieves 85% precision, 76% F1-measure, and 71% recall using our proposed Thai word separator with Classification and Regression Trees (CART) technique. However, we need to increase the precision which is strong enough to support our proposed separator. The additional experiment has been done by adding 50 chief complaint cases to the test dataset. We also have applied our proposed techniques including conflict element finding and classification criteria setting to improve the precision. Consequently, the later experimental results get higher classification accuracy by decreasing the false positives to mitigate the low recall problem.