Semiparametric model averaging method for survival probability predictions of patients

Abstract

In biomedical and clinical research, predicting the survival probabilities for patients is a core task. Accurate survival probability predictions can help physicians make better treatments or prevention plans for patients. A novel semiparametric proportional hazards model averaging prediction technique is introduced to address this problem. Under the potential partly linear additive structures, the conditional survival probabilities of individuals can be predicted by the weighted averages of submodels which are estimated by maximizing the partial likelihood functions. The selection of weights is a crucial part of model averaging since the weights can affect the accuracy of survival probability prediction. A Brier score type criterion is employed to choose the optimal model averaging weights and the rate of convergence of the selected weights is studied. In addition, the finite sample performance of the proposed method is evaluated via abundant simulation studies. To further illustrate the effectiveness of the proposed approach, the model averaging is applied to heart failure data.