Machine learning approach for the binary classification of biomedical literature

Abstract Background: We have applied machine learning techniques to automate the screening of biomedical literature prior to the manual curation of clinical databases such as performed by the Human Gene Mutation Database (HGMD). Methods: We have developed two machine learning models, one based on title and abstract data only, the other on the full text of the article. The models were built using a Natural Language Processing (NLP) pipeline and a logistic regression classifier. Our pipelines are implemented in Python and can be run using Docker. They are made available to the wider community via GitHub (https://github.com/annacprice/nlp-bio-tools) and Docker Hub. Results: During testing, both models performed well, correctly predicting HGMD relevant articles more than 93% of the time and correctly discarding irrelevant articles more than 96% of the time, with Matthews Correlation Coefficients (MCC's) of over 0.89. Evaluation of the finalised model using an unseen validation dataset demonstrated that the full text model correctly predicted HGMD-relevant articles more than 97% of the time, an accuracy 9.5% higher than that obtained with the title/abstract model. Conclusions: Through this work we have demonstrated that machine learning models can act as an effective pre-screen of biomedical literature, with the results indicating that a full text approach to screening biomedical literature is preferable to using just the title/abstract data.