Fig. 2: The framework of our proposed method.
From: Stable Cox regression for survival analysis under distribution shifts
a, Stage 1: independence-driven sample reweighting. The sample reweighting module decorrelates the spurious correlations among covariates. b, Stage 2: weighted Cox regression. The module utilizes the learned sample weights W to reweight the event of each subject in the partial log-likelihood loss. The weighted partial log-likelihood loss could drive the model to utilize stable variables S to make predictions rather than unstable variables V (that is, \({\hat{\beta }}_{w}(V\,)=0\)).
