Automated labeling of terms in medical reports in Serbian

Abstract

Nowadays, many electronic health reports (EHRs) are stored daily. They consist of the structured part and of an unstructured section written in natural language. Due to the limited time for medical examination, EHRs are short reports which often contain errors and abbreviations. Therefore it is a challenge to process an EHR and extract knowledge from this part of the text for different purposes. This paper compares the results of three proposed methods for automatic labeling of medical terms in unstructured parts of EHRs. All words are categorized as words within the medical domain (symptoms, diagnoses, therapies, anatomy, specialties etc.) and those beyond the medical domain (numbers, places, stop words etc.). The first method is based on dictionaries of medical terms, the second on the training set, and the third on the training set and rules. The results of application of different methodologies to reduce a word to its basic form (pure, prefix, stem) are given for each of the methods. The paper shows that in labeling medical terms, the methods based on medical dictionaries (diagnosis, symptoms, medications etc.) do not produce best results, therefore it is better to use manually annotated part of the data set as a model. A significant number of words (17.36%) in medical reports are abbreviations and errors, so for better results, we should focus on creating rules to solve this problem. Better results are obtained for supervised methods compared to the dictionary-based method (with relative improvement of 42.82%). The inclusion of the algorithm for processing errors and abbreviations increased the results (with a relative improvement of 4.21%) and gave the largest F1-measure (0.9082). The advantage of the proposed method is that the use of rules for processing errors and abbreviations provides good results regardless of how the word is reduced to its basic form.

Creating a stop word dictionary in Serbian

“…Medical and non-medical terms are manually marked in the medical reports. For each medical report is assigned a diagnosis code (Avdić et al, 2020). A corpus (DMRC, Table 1) of 100 discharge lists and 50 reports from doctors from the Faculty of Dentistry at the University of Belgrade was used to evaluate the system's effectiveness in automatically analyzing temporal expressions of medical narrative texts.…”

“…In a separate chapter on sentiment analysis below, sentiment word lexicons and other lexicons used in sentiment analysis are described more. Avdić et al (2020) The Serbian stop word dictionary (SSW dictionary) contains 1241 different stop words for the Serbian language. It was created based on the grammar of the Serbian as well as by comparing with available sets of stop words for the Serbian language and a set of stop words for the Croatian language.…”

“…They have manually checked 100 documents, reporting accuracy ranging from 64% to 90%. A mathematical model for medical term recognition was proposed by (Avdić et al, 2020). They have proposed three methods.…”