Landmark Approach to Improve Prediction Accuracy of Long-Term Survival by Incorporating Time Information from Multiple Short-Term Events

• Activity Number: 53 - New Developments in Survival Analysis
• Type: Contributed
• Date/Time: Sunday, August 8, 2021 : 3:30 PM to 5:20 PM
• Sponsor: Biometrics Section
• Author(s): Resmi Gupta and Jing Ning and Jing Jing Zhang and Wen Li and Sean I Savitz and Liang Zhu and Sori Kim and Amirali Tahanan and Mohammad Hossein Rahbar*
• Companies: University of Texas Health Science Center at Houston and The University of Texas MD Anderson Cancer Center and University of Texas Health Science Center at Houston and University of Texas Health Science Center at Houston and University of Texas Health Science Center at Houston and University of Texas Health Science Center at Houston and University of Texas Health Science Center at Houston and University of Texas Health Science Center at Houston and University of Texas Health Science Center at Houston
• Keywords: Landmark analysis; Long-term survival ; Prediction accuracy ; Short-term events; Simulation studies; Stroke

Abstract

In clinical settings, patients who experience multiple short-term events such as hospital readmission may have substantially different long-term survival outcomes as compared with others. We proposed a flexible varying-coefficient landmark model by incorporating multiple short-term events times information into the prediction of long-term survival. We aimed to predict the risk of long-term survival T? t0 + L, (L>0) among patients survived up to a pre-specified landmark time t0 since the initial admission. The predictive performance of the proposed model was evaluated and compared using predictive measures, Brier score, and, area under the curve. A series of simulation studies confirmed that parameters under the landmark models could be estimated accurately and the predictive performance of the proposed method consistently outperformed existing methods without or partially incorporating multiple short-term event times information. We demonstrate the model’s application using UTHealth Houston stroke registry data. Different values of landmark time points and the long-term survival time since t0 were considered for risk prediction.