Design and performance comparison of DNN architectures for multi-label classification in clinical opinions generation of comprehensive blood test

Abstract

As interest in healthy life increases, the number of people who take the health check-up increases. It has clinical experts analyze lots of test results. The clinical experts suffer from heavy workloads. In order to handle this problem, in this thesis, we propose the novel method to generate clinical opinions from comprehensive blood tests. We define the clinical opinions generation as a multi-label classification problem, and select clinical opinions for a patient's test results in a clinical opinion set. Based on the problem definition, we propose seven deep neural network (DNN) models for clinical opinions generation. The baseline DNN model shows low performance for low frequency clinical opinions, and has an inappropriate architecture for sparse datasets. It also, generate inconsistent clinical opinions. In order to improve this limitations, we propose six DNN models. We compare and analyze the performance of the proposed models. We confirm the best model for clinical opinions generation from comprehensive blood tests. Additionally, we evaluate the performance of each proposed model for normal and abnormal clinical opinions. The normal clinical opinions indicate that a patient is healthy, while the abnormal clinical opinions indicate that a patient has diseases. The conflict avoidance disease-dependent multi-label DNN model shows the best performance, and it shows the highest performance for normal and abnormal clinical opinions.