Constructing a prediction model of thyroid dysfunction making use of a set of time-series routine tests data

2022 Fiscal Year Final Research Report

• Project/Area Number: 19K12206
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 62010:Life, health and medical informatics-related
• Research Institution: Tohoku Medical and Pharmaceutical University
• Principal Investigator: Aoki Sorama 東北医科薬科大学, 薬学部, 助教 (40584462)
• Project Period (FY): 2019-04-01 – 2023-03-31
• Keywords: 甲状腺機能異常症 / 機械学習 / 人工知能 / 医療統計 / 基本的検査 / 時系列解析 / スクリーニング

Outline of Final Research Achievements

In order to prevent thyroid dysfunctions from being overlooked, we constructed a machine-learning-based prediction model for thyrotoxicosis and hypothyroidisms by learning a combination of frequently measured blood test items, as a preliminary step to measuring thyroid-related hormones. We have also demonstrated that it is possible to introduce that model into actual hospitals by running it on the cloud.
Additionally, by adding a feature that expresses the rate of annual change from the values at the last examination, and turning it into a time-series model, we have successfully reduced the false-positive rate of the hyperthyroidism prediction model by more than half. On the other hand, even by turning it into a time-series model, the accuracy of the hypothyroidism prediction did not improve, and no improvement was observed even after trying various learning methods for the model.

Free Research Field

医療情報学

Academic Significance and Societal Importance of the Research Achievements

甲状腺機能異常症見逃し防止のための機械学習モデルについて、クラウド化により実際の病院における3健診施設にて稼働させることができたことで、人工知能の応用化が多忙な健診施設においても可能であることを示せたと考えている。また、本モデルはこのうち対象者数が最大であった施設（東北公済病院）にて見逃されていた患者25例の発見にも成功しており、患者QOLの向上にも繋がることを実証した。さらに検査値の時系列情報について、速度として取り扱うと有用である疾患と有用でない疾患の差異を限定的ながら明らかにできたことで、臨床検査値の取り扱いに関する知見を深めることができた。