Keisuke Ogawa

Method of Screening the Health of Persons with High Risk for Potential Lifestyle-related Diseases using LDA

Recently, the number of patients with lifestyle-related diseases, such as diabetes mellitus, has increased dramatically. Lifestyle-related diseases are responsible for 60% of deaths in Japan. In order to screen persons at potentially high risk for these diseases, medical checkups for metabolic syndrome are used throughout Japan. Prediction and prevention of lifestyle-related diseases would yield a direct reduction in medical costs. However, many cases cannot be screened with a metabolic syndrome checkup. In this paper, we propose a new machine-learning-based screening method using medical checkup data and medical billings. By processing the medical data into a bag-of-words representation and classifying the health factors using latent Dirichlet allocation (LDA), the screening method achieves high accuracy. We evaluate the method by comparing the accuracy of predictions of the future incidence of the diseases. The results show that F-measure increases 0.17 compared with the conventional method. In addition, we confirmed that the proposed method classified persons with different health risk factors, such as a combination of metabolic disorders, hypertensive disorders, and mental disorders (stress).

Editing Training Sets from Imbalanced Data Using Fuzzy-Rough Sets

In this research, we study several instance selection methods based on rough set theory and propose an approach able to deal with inconsistency caused by noise and imbalanced data. Recent attention has focused on the significant results obtained in selecting instances from noisy data using fuzzy-rough sets. For imbalanced data, fuzzy-rough sets approach is also applied before and after using balancing methods in order to improve classification performance. In this study, we propose an approach that uses different criteria for minority and majority classes in fuzzy-rough instance selection. It thus eliminates the step of using balancing techniques employed in controversial approach. We also carry out some experiments, measure classification performance and make comparisons with other methods.