[2G3-GS-2e] 機械学習：予測

[2G3-GS-2e-03] Patient Disease Prediction and Medical Feature Extraction using Matrix Factorization

〇Yuki Sumiya¹, Atsuyoshi Matsuda², Kenji Araki³, Kazuhide Nakata¹ (1. Tokyo Institute of Technology, 2. Logbii, Inc., 3. University of Miyazaki Hospital)

Keywords:Medical Data, Matrix Factorization, Prediction, Feature Learning, Data Mining

It is expected to extract and apply useful information from the big EMR data. In particular, "prediction / prevention of diseases that patients may develop" and "analysis of features and relationships of medical events" will help solve problems caused by the shortage of doctors. In order to achieve these goals, we used Matrix factorization-based methods to "A: calculate the risk of developing each disease for each patient" and "B: obtain and analyze the feature representations of patients, diseases, and patient characteristics". However, it was difficult to apply the existing methods. In this study, we developed a new method called PCMF to solve concerns about these methods. Then, by applying PCMF to the EMR data, we aimed to achieve the above A and B simultaneously. Our experiments showed PCMF predicts future diseases more accurately than other methods. We also analyzed the obtained feature representations, and showed PCMF can extract useful information.