Mixing Domain Rules with Machine Learning for Radiology Text Classification

In this work we compare and contrast proposed techniques and experiments in the subdomain of radiology text classification. In the past, systems have relied on two main approaches: rule-based or natural language processing plus machine learning (NLP/ML). Simple rule-based approaches have demonstrated great efficacy, at the high cost of requiring medical professionals to build and maintain. Complex NLP pipeline and ML approaches have also demonstrated efficacy, but rely heavily on computing professionals to build and maintain, in addition to requiring the involvement of medical domain experts. We propose a hybrid classification mechanism for radiology text combining rules and NLP/ML that in our trials surpasses existing rule-based and NLP/ML techniques for the task of classifying incidental findings. We evaluate our approach against three approaches from related work using our own gold standard data set of 661 records. Our hybrid approach achieves a 13% F-measure gain over our prior rulebased approach[16] and a 4% F-measure improvement vs. a manual classification process in a hospital setting.