Automated disambiguation of acronyms and abbreviations in clinical texts: window and training size considerations

Acronyms and abbreviations within electronic clinical texts are widespread and often associated with multiple senses. Automated acronym sense disambiguation (WSD), a task of assigning the context-appropriate sense to ambiguous clinical acronyms and abbreviations, represents an active problem for medical natural language processing (NLP) systems. In this paper, fifty clinical acronyms and abbreviations with 500 samples each were studied using supervised machine-learning techniques (Support Vector Machines (SVM), Naïve Bayes (NB), and Decision Trees (DT)) to optimize the window size and orientation and determine the minimum training sample size needed for optimal performance. Our analysis of window size and orientation showed best performance using a larger left-sided and smaller right-sided window. To achieve an accuracy of over 90%, the minimum required training sample size was approximately 125 samples for SVM classifiers with inverted cross-validation. These findings support future work in clinical acronym and abbreviation WSD and require validation with other clinical texts.