[1P4-OS-1b] OS-1

[1P4-OS-1b-04] Proposal of a Hazard Ratio Estimation Method for Survival Outcomes Using Data Collaboration Analysis

〇Tomoru Nakayama¹, Yuji Kawamata², Akihiro Toyoda¹, Yukihiko Okada^2,3 (1. Univ. of Tsukuba, Master’s Program in Service Engineering , 2. Univ. of Tsukuba, Center for Artificial Intelligence Research, 3. Univ. of Tsukuba Institute of Systems and Information Engineering)

Keywords:privacy preserving, data collaboration, propensity score

Many studies have examined how to estimate causal effects of treatments on survival outcomes using survival analysis. Marginal hazard ratios are often employed to quantify these effects. Although aggregating data from multiple institutions improves estimation accuracy, privacy concerns hinder data sharing. Existing privacy-preserving methods that estimate marginal hazard ratios via propensity scores mainly address distributed samples but not situations where both samples and covariates are distributed. Limited covariate availability hampers accurate propensity score estimation. Therefore, this study proposes and validates a framework applying Data Collaboration (DC) analysis to handle distributions of both samples and covariates. Experiments on synthetic data showed that the DC-based method outperforms individual analyses at a single institution and achieves accuracy comparable to centralized analysis, the ideal scenario where all data is directly shared.