Supervised Clinical Abbreviations Detection and Normalisation Approach

The ambiguous acronyms and abbreviations in clinical reports can be quite confusing for patients and doctors to understand, which will potentially lead to medical malpractice [15]. To solve this problem, we proposed a supervised approach to detect abbreviations in given clinical reports and normalise these abbreviations to medical concepts. In the step of detection, a seq2seq model with the attention mechanism was built and achieved the micro-average F1 score of 83.85% among 99 test reports. In the step of normalisation, we used both internal and external senses inventories to build one disambiguation classifier for each abbreviation. Finally, the proposed normalisation method achieved a micro-average accuracy of 74.7%, beating the first ranked team in the ShARe/CLEF eHealth 2013 competition, Task 2. This work provided a complete pipeline to handle ambiguous abbreviations in clinical documents, which is essential for healthcare providers and researchers to understand and subsequently leverage the clinical reports.