Finding patient-oriented evidence in PubMed abstracts
Robinson, David Alexander
MetadataShow full item record
This project develops a computational method to improve searches of the medical literature by selecting the studies that report reliable evidence of patient-oriented outcomes. These outcomes include morbidity, mortality, symptom severity, and quality of life. Four machine learning methods, support vector machines, naive bayes, naive bayes multinomial and Logistic Regression, achieve over 70% accuracy on the identification of such studies in PubMed abstracts. The accuracy attainable by hand in this task is about 95%. The best machine learning results, just over 80% accurate, were obtained with naive bayes multinomial on a combination of single words and contiguous pairs of words using WEKA 3.7.5.