Automated non-alphanumeric symbol resolution in clinical texts

Area of Science:

• Medical Natural Language Processing (NLP)
• Computational Linguistics
• Bioinformatics

Background:

• Clinical texts contain numerous symbols requiring accurate interpretation.
• Symbol resolution is an under-addressed area within medical NLP research.
• Interpreting symbols is analogous to Word Sense Disambiguation (WSD).

Purpose of the Study:

• To evaluate the accuracy of supervised techniques for resolving common clinical symbols.
• To determine the effectiveness of symbol context and other features in disambiguation.
• To assess the feasibility of integrating symbol disambiguation into medical NLP systems.

Main Methods:

• Extracted 1000 instances of four symbols (+, -, /, #) from clinical documents.
• Annotated symbols and their context by domain experts.
• Evaluated features including symbol context, Bag-of-Words (BoW), and heuristic rules.
• Utilized Naïve Bayes, Support Vector Machine, and Decision Tree classifiers with 10-fold cross-validation.

Main Results:

• Achieved high accuracies: 80.11% for '+', 80.22% for '-', 90.44% for '/', and 95.00% for '#', using Naïve Bayes.
• Symbol context was the most significant feature for disambiguation.
• Bag-of-Words (BoW) also contributed to the disambiguation of certain symbols.

Conclusions:

• Supervised techniques can achieve reasonable accuracy for clinical symbol disambiguation.
• Symbol disambiguation can be effectively implemented as a module within medical NLP systems.
• This research enhances the potential for more sophisticated analysis of clinical text data.