Learning and recommending treatments using electronic medical records

Treatment directly affects patient health status. In recent years, the fast development electronic medical records (EMRs) has provided valuable resources for solving healthcare issues, especially learning and recommending treatments. However, most of the related studies are limited in exploiting various patient information, handling varying-length treatment records, capturing different kinds of treatment patterns and interpreting the recommendation mechanism. This research proposes three methods, one for learning and two for recommending treatments, to overcome the above drawbacks. All methods adopt a mixed-variate restrict Boltzmann machine to represent different kinds of patient records. The treatment learning method captures significant changes in prescription indication to split varying-length records flexibly and organizes sequences of prescription drugs into regimen trees to reveal many more different kinds of treatment patterns. The two treatment recommendation methods illustrate different ideas that can improve the treatment recommendation mechanism by combining the treatments derived from patient groups and neighbor patients. Our experimental evaluation was conducted on three acute disease cohorts extracted from the MIMIC III database. The obtained results show that the proposed methods are able to provide different kinds of treatment patterns and yield competitive efficacy with better interpretability as compared to relevant studies. (C) 2019 Published by Elsevier B.V.