Experimental Validation of Knowledge-Augmented Graph Networks for Clinical Acronym and Abbreviation Disambiguation in Electronic Health Records

The problem of medical abbreviation disambiguation is one of the most complex tasks in the domain of medical discrete text processing. The same abbreviation may be used to denote a number of different medical terms (concepts) depending on the clinical note in which it appears. Although large transformer models appear to have improved the understanding of the contextual meaning of an abbreviation, they usually do not make use of any information available in structured relationships in biomedical knowledge bases. Consequently, they can make errors in abbreviation disambiguation. In this paper, we present an experimental assessment of an Explainable Knowledge-Augmented Graph Attention Network (EKAGAT). The proposed framework is grounded on transformer-based contextual embeddings, and ontology-guided graph attention, which together allows the model to reason jointly over the textual context and a structured representation of biomedical knowledge. The proposed framework involves the construction of a heterogeneous medical knowledge graph using readily available biomedical ontologies. The model then uses attention-based message propagation to encode relationships among abbreviations, clinical entities and their candidate concepts. In addition, the model incorporates a semantic consistency regularizer to encourage the learned representations to align with the ontology and to enhance the interpretability of the model predictions. The proposed framework is empirically evaluated on three of the most common electronic health record (EHR) benchmark datasets for medical abbreviation disambiguation, namely, CASI, i2b2 and MIMIC-III, and compared against a number of state-of-the-art transformer-based and graph-based approaches. This will influence the design of future reliable and trustworthy clinical NLP. We also provide insights into the practical aspects of designing and deploying such models for real-world applications.