Probabilistic model-based time-series forecasting neural networks and related applications to biosignal forecasting

2020 Fiscal Year Final Research Report

• Project/Area Number: 17K12752
• Research Category: Grant-in-Aid for Young Scientists (B)
• Allocation Type: Multi-year Fund
• Research Field: Soft computing
• Research Institution: Kyushu University
• Principal Investigator: Hayashi Hideaki 九州大学, システム情報科学研究院, 助教 (00790015)
• Project Period (FY): 2017-04-01 – 2021-03-31
• Keywords: ニューラルネットワーク / 深層学習 / 時系列予測 / 確率モデル / 生体信号

Outline of Final Research Achievements

We proposed mathematical models that can represent biological signals, which are electrical signals that can be measured from the human body, and applied them to the analysis of actual data. For example, we modeled myoelectric signals, which are electrical signals of muscles, and applied them to the analysis of muscle strength. We also proposed models for electrocardiogram and electroencephalogram and applied them to signal classification. Furthermore, we developed neural networks based on probabilistic models, and applied them to data classification and time series forecasting. In addition, we constructed large medical datasets such as biological signals of a pregnant woman called cardiotocography and endoscopic images, and applied them to fetal condition prediction and organ classification.

Free Research Field

機械学習

Academic Significance and Societal Importance of the Research Achievements

学術的意義として最大の点は，本研究においてニューラルネットワーク（NN）へのドメイン知識埋め込み法を提案している点である．提案法では，データの特性を確率モデルに基づき表現し，それをNNへ埋め込むことで解釈性や汎化性を向上させる．
社会的意義としては，生体信号の予測が実現できれば医療モニタリング応用に役立つ．在宅医療を受ける患者は約18万人いるとされる（2017年厚生労働省調べ）．そのような患者に対し，自宅でも計測が容易な血圧や指尖容積脈波などと提案法を組み合わせることにより容態変化を予測することができれば，異常が起こる前に医師に連絡することができスムーズな治療が期待できる．