Recommendation System for Clinical Concept Mapping

In the past decade, the healthcare industry has shifted from paper-based document storage to Electronic Health Records (EHR), enabling quick, safe access to patient data. A key role is played by Semantic Interoperability (SI) which enables seamless data exchange between diverse care settings and clinical software. SI necessitates the linking of bio-medical data (aka. clinical events) with shared, standardized, and controlled vocabularies like SNOMED, LOINC, etc. However, the healthcare data across various client domains are filled with ambiguous textual representations of clinical events that may be present in the form of synonyms, acronyms, and abbreviations. To make interoperability work, Healthcare IT service providers must map related clinical events with the appropriate standard concepts, which requires additional time and resources. Natural Language Processing (NLP) plays a vital role in addressing the challenges of SI by learning effective representations of text words in the bio-medical domain thereby capturing their semantic meaning. Our method utilizes various pre-trained word embeddings trained on the bio-medical corpus like BioWordVec fastText and SapBERT that captures fine-grained semantic relationships. In this study, we have developed a recommendation system that provides recommendations of Top ‘N’ clinical events for mapping to a standard concept. The recommendation system showed good performance with a sensitivity of above 99 % using both the pre-trained word embedding. Further, this product can be integrated into the mapping workflow to help make accurate automated suggestions that minimize manual effort.