Concept extraction for disability insurance payment evaluation

Abstract

Automated evaluation of claims for medical and disability insurance benefits poses a difficult challenge that will take years to be solved. The precise wording of insurance rules and the terse language in medical history files make it difficult for humans, let alone computers, to assess insurance payment qualification accurately. In this thesis, we work towards building a tool that will aid, but not replace, human evaluators. We automate the extraction of relevant parts of medical history files; if sufficiently accurate, this would eliminate the need for human evaluators to comb through hundreds of pages of medical history files. We first create a list of medical concepts, mainly disease and procedure names, from the cardiovascular section of the "Blue Book" for Disability Evaluation under Social Security. Then, using a variation of the longest common substring algorithm, we characterize each medical file line using its substring overlaps with the list of medical concepts. Finally, with human annotations of whether each medical file line is relevant or not, we build machine learning classifiers predicting each line's relevance using its overlap characterization. The classifiers we use are Naive Bayes and Support Vector Machines.