Supporting Information Retrieval From Clinical Narrative Texts Using Text Classification and Visualization Techniques

Abstract

Electronic health records (EHRs) are increasingly becoming common in healthcare delivery settings, enabling physicians to continuously document patients’ care episodes in EHRs and therefore creating digital health records of individual patients. This along with general advances in software and computer hardware has made it increasingly easy to capture and store patient clinical data electronically which has, in turn, yielded an increasing supply of readily available patients’ clinical information in electronic health records. However, retrieving clinical information from long clinical narrative texts is a challenging problem. Clinical narrative texts provide significant difficulties for conventional natural language processing techniques when it comes to retrieving information from long clinical text documents. Additionally, a detailed assessment of past studies revealed a lack of theoretical support for the computational techniques employed, and a scarcity of research on information retrieval techniques applicable to clinical narrative texts. Most clinical information is in narrative text form which limits users from quickly finding desired information given the large volume of texts that needs to be read. A physician's capacity to read and get information from text for clinical overview is severely affected when records get longer. Thus, medical practitioners are increasingly confronted with information flood which provides more information than practitioners can process especially in time constraint healthcare delivery settings. Physicians still use a standard linear text layout in computer screens which one must read through the same way one would their paper counterparts. Thus, the challenges experienced by physicians in navigating, retrieving, and synthesizing paper-based patient records remain unsolved with current electronic health systems. This makes the process of identifying the most critical and significant nuggets of clinical information in a given patient clinical record challenging but worthwhile task. Due to the continuous growth of clinical data, automated retrieval of important pieces of information from clinical narrative text is becoming an important research problem. In this thesis, we proposed text classification and visualization model to support information retrieval from clinical narrative texts in electronic health records. The objective of this study is to overcome the challenges encountered in retrieving information from clinical narrative texts. In our approach clinical narrative texts are classified into different class orientations and then visualized as a cluster map. A set of five information facets were identified to characterize relevant facets of information. We developed a deep learning algorithm to classify clinical narrative texts into a predefined set of classes. From the state of art, it was

found out that training deep learning models is difficult, and getting them to converge in a reasonable amount of time is still challenging. To overcome this problem, we proposed a novel layer normalization technique called range normalization which outperformed conventional techniques. An artefact was developed and user study was conducted to collect feedback from healthcare practitioners on the usefulness of such a tool in supporting information retrieval from clinical texts. The results showed that such a model would be useful to physicians in reviewing clinical narrative texts especially for patients with lengthy medical histories. The results from the study show that the artefact is useful and effective in supporting physicians during care episodes. The results suggest that integrating visualization into an electronic health record would be beneficial to physicians in satisfying their information needs. Modelling texts into meaningful information facets provides organized clinical narrative text documents.