Practical implementation of an existing smoking detection pipeline and reduced support vector machine training corpus requirements

摘要

This study aimed to reduce reliance on large training datasets in support vector machine (SVM)-based clinical text analysis by categorizing keyword features. An enhanced Mayo smoking status detection pipeline was deployed. We used a corpus of 709 annotated patient narratives. The pipeline was optimized for local data entry practice and lexicon. SVM classifier retraining used a grouped keyword approach for better efficiency. Accuracy, precision, and F-measure of the unaltered and optimized pipelines were evaluated using k-fold cross-validation. Initial accuracy of the clinical Text Analysis and Knowledge Extraction System (cTAKES) package was 0.69. Localization and keyword grouping improved system accuracy to 0.9 and 0.92, respectively. F-measures for current and past smoker classes improved from 0.43 to 0.81 and 0.71 to 0.91, respectively. Non-smoker and unknown-class F-measures were 0.96 and 0.98, respectively. Keyword grouping had no negative effect on performance, and decreased training time. Grouping keywords is a practical method to reduce training corpus size.