USES OF CONTINUOUS-TIME MARKOV CHAIN TO DESCRIBE LONGITUDINAL PATIENT-REPORTED OUTCOMES FOR SURVIVAL PREDICTION AND DIMENSION REDUCTION

A patient-reported outcome (PRO) is a type of outcome reported directly from patients, and it has been widely used in medical research and clinical trials to measure a patient’s symptoms, health-related quality of life, physical functioning, and health status. Previous studies have linked PROs to survival outcomes, but most of them only used the PRO information at baseline or at a specific clinical time point [1, 2]. Even though some of these studies collected longitudinal PROs, only few of them evaluated the association between the longitudinal PROs and a survival outcome. One of the major challenges in longitudinal PRO studies is to address the individual heterogeneity in PRO repeated measurements. Due to the fact that PRO is reported directly from patients, and different patients may have different experiences, longitudinal PROs have been often observed with individual heterogeneity, yet current methods [3-5] are not able to account for the individual heterogeneity. Therefore, in this research, we developed three methods using two-state Continuous-Time Markov Chain (CTMC) to summarize longitudinal PRO. The primary summary used is the estimated state transition rates, which serve as summary statistics to depict longitudinal PRO patterns at the individual level. These transition rates can also be incorporated into survival models as predictors or into factor analysis as observed variables. Specifically, in the first two papers, we developed prognostic models that contained baseline covariates and a longitudinal process in two survival models, Weibull Regression and Cox Proportional Hazard Regression, with different estimation approaches. Simulation studies were conducted to validate the proposed methods, and the proposed models were then applied to two PRO studies separately, with both using repeated PRO measurements during the treatment period in cancer patients to predict the survival outcomes that happened after the treatment. In the third paper, we then integrated two-state CTMC with factor analysis to evaluate the usage of CTMC in PRO symptom clustering. This study showed that CTMC could well summarize the longitudinal PRO information during the treatment period of cancer patients. The underlying construct of patient-reported symptoms had also met our expectations from clinical experience