Categorisation of Healthcare spend data using Machine Learning and Transfer Learning

Abstract

New Zealand is divided into 20 independent District Health Boards (DHBs) for the delivery of healthcare services. Most DHBs have their own independent software system to manage their purchases and suppliers, and items in each of these systems have their own naming conventions. In order to ascertain and analyse the spread of spending in healthcare, there is a drive to standardise and centralise historic information automatically. The aim of this research is to consolidate the spend data across all DHBs by classifying the data into categories . We propose the use of classification algorithms in order to achieve highest accuracy when categorising DHB spend data. Given the large number of response values which exceeds 1,800 unique values, accuracy is of paramount importance. The random forest model was the best performing model and given abundant labeled data gave us an F1-score of 0.98. Since our labeled data for certain DHBs is larger than other DHBs we also explore transfer learning and domain adaptation techniques to improve the prediction accuracy of the DHB with scarce amounts of labeled data. We use the data from the DHB with abundant labeled data by either adapting the data to fit the target domain or transferring parts of an actual model built using the abundant data to the target model. All methods have positive outcomes and increase the accuracy of our prediction model. However, the transfer learning method where parts of the actual model (knowledge) are transferred improved the F1-score from 0.47 to 0.72. This method showed the greatest accuracy improvement of any of the techniques used in our thesis.