한국 성인에서 심뇌혈관 질환 위험도 예측 ; Prediction of Cardiovascular Risk in Korean Adults: Cox proportional hazards model and privacy-preserving deep learning ; Cox 비례위험모형과 프라이버시 보전 딥러닝

학위논문(박사) -- 서울대학교대학원 : 의과대학 의학과, 2022. 8. 윤형진. ; While artificial intelligence (AI) has demonstrated extensive potential in various domains, developing AI models in healthcare is limited because of medical data privacy restrictions. Privacy preserving deep learning (PPDL) attempts to train AI models while protecting data privacy, and one type of PPDL training algorithm named federated learning (FL) has been gaining attention as a solution. However, FL is based on model synchronizations that require automated FL platforms, which may be difficult to implement in some medical institutions with closed networks. We present transfer iterative learning (TIL), a decentralized PPDL training algorithm based on transfer learning which can achieve model performances comparable to FL at lower communication cost. Therefore, the main purpose of this study is to preliminarily analyze predictive modeling applying the conventional statistical method using data from the National Health Insurance Service, and to analyze the patient data implemented in the Common Data Model (CDM) of a single institution. It is to implement FL and TIL by simulating with multicenter data and compare the performance. For predictive modeling with clinical significance, although many prediction models for cardiovascular diseases (CVDs) have been developed and validated for Western populations, the development of CVD prediction models for Asians has been slow. Our cohort study retrospectively analyzed the incidence of CVD that occurred between January 1, 2009, and December 31, 2016, in all Koreans who underwent national health screening. This dataset included 21,581,796 adults between the ages of 40 and 79 years (10,412,947 men, 11,168,849 women) without CVD at baseline. The primary outcome, CVD, was defined as the development of any of the following: acute coronary syndrome, cerebral infarction, and cerebral hemorrhage, as defined with health insurance claims data. The prediction model was constructed by Cox proportional hazard regression and validated ...