Clinical Bert fine tuning for question answering in ICD coding

Abstract

In this paper we will propose a method for clinical Bert model fine tuning for questions answering task which utilized in the international classification of disease (ICD) coding process, the fine-tuning process enhance the accuracy of  any model since the models are trained on generic data for generic tasks, we used medical clinical Bert model for fine-tuning it on customized medical dataset, the preparation of the training data set is done in many steps to be  converted into question/answering suitable format, that the training data set converted to have positive answers and negative answers mapped to the ICD codes, the original data set was contains records of patient claims textual data with the mapped ICD codes, it was converted in a suitable format for questions answering based on the ICD catalog, the data conversion done by utilizing biomedical-ner-all model as medical named entity recognition model to detect the medical entities in training textual data, a training parameters was set to the fine tuning process, before model fine-tuning the accuracy in questions/answering in ICD coding was about 40% , after evaluation our fine-tuned model the accuracy reached to 90%.