Knowledge-Augmented Graph Framework for Clinical Acronym Disambiguation

Electronic Health Records (EHRs) are rapidly becoming the norm in clinical environments all over the world being used as the main source of medical history of patients, clinical records and notes, diagnoses, and treatment reports. The prominence of abbreviations and acronyms is a particular characteristic of the clinical documentation that clinicians are forced to employ to make the process more efficient [1]. However, these acronyms can be significantly expanded by clinical situations in most cases. An example of this would be the word BP may refer to Blood Pressure or the Bell’s Palsy and the word MS may be used to refer to Multiple Sclerosis, Mitral Stenosis or Mental Status among others. This unavoidable ambiguity is a major issue of the automated NLP systems which are expected to analyze and divide clinical text [2]. Manual systems of abbreviation disambiguation like rule-based lookup table & dictionary matching have their weaknesses in regard to generalization to other clinical sub-domains and the resources that must be manually maintained [3]. Contextual understanding in the clinical NLP has improved significantly with the emergence of deep learning, and particularly transformer-based models, such as BERT and BioBERT [4]. The models are however applicable to raw text and do not explicitly incorporate structured medical knowledge in biomedical ontology such as the Unified Medical Language System (UMLS), SNOMED CT and the Medical Subject Headings (MeSH). One potential such gap is the desire to develop Knowledge- Augmented Graph Networks (KAGNs) that combine representational power of contextual embeddings with semantic power of biomedical knowledge graphs to empower better and explainable acronym disambiguation in EHRs. The remainder of this paper will be arranged in the following manner. Section 2 summarizes previous research in the areas of clinical acronym disambiguation and biomedical NLP. Section 3 provides the most important contributions of the proposed framework. The methodology and experimental framework is outlined in Section 4. Section 5 addresses the results of the experiment, and Section 6 brings to an end the paper with a research direction.