Dynamic Prediction by Landmarking in Recurrent Neurological Stroke Using Box-Cox Transformed Longitudinal Risk Factors
Abstract
In medical research, natural logarithmic transformation was a standard approach to transforming outcomes to achieve a normality assumption, but an alternative transformation may yield a better strategy. The Box-Cox method helped select an outcome transformation to ensure the validity of a Gaussian distribution and related assumptions in regression modeling. In the second chapter of this dissertation study, we extended the Box-Cox method to a linear mixed model with an autoregressive correlation structure. Analyses of longitudinal measurements of neurological patient blood glucose and a simulation study illustrated the benefits of the proposed methodology and the resulting proper transformation. Our study indicated that neurological patient blood glucose should be transformed with inverse square root transformation to satisfy Gaussian distribution before making valid statistical inference. After longitudinal lab variables were transformed properly, we combined the transformed time-dependent lab variables with baseline demographic factors to conduct a dynamic prediction of risk probabilities of recurrent neurological stroke; this work is outlined in the third chapter. We proposed a dynamic prediction strategy with a landmarking process, which is based on an extension of an Andersen-Gill (A-G) intensity model, to study the association between recurrent stroke and longitudinal lab variables after adjusting for patient baseline characteristics. The benefit of this landmarking approach is that predictive probabilities can be made at any time interval during the follow-up period, adapted to the changing at-risk population, and while applying the most recent longitudinal measurements. Simulation studies on recurrent events revealed that this proposed approach has biased estimates but is nevertheless robust against model misspecification. We applied our proposed method to patients with neurological injuries in order to dynamically predict a patient’s risk of experiencing recurrent stroke events during the hospital stay. As outlined in the fourth chapter, we demonstrated the dynamic risk prediction tool for postoperative stroke based on neurosurgical patients admitted to the Neuro Intensive Care Units (NICU). We found that mean arterial pressure (MAP) and diagnosis group were significant predictors of postoperative stroke after adjusting for blood glucose level, history of diabetes, and age. Patients who are initially classified in the Traumatic Brain Injury group with consistently high MAP and blood glucose lab measurements are at greatly increased risk of having a postoperative stroke event in the NICU. Neurological stroke is common in the general population and often results in long-term injury. Prediction of these event probabilities enabled physicians to make early interventions when appropriate. Hence, our proposed methods provided a novel prediction tool in hospital patient care.
Subject Area
Biostatistics|Bioinformatics|Health care management
Recommended Citation
Liu, Zheyu, "Dynamic Prediction by Landmarking in Recurrent Neurological Stroke Using Box-Cox Transformed Longitudinal Risk Factors" (2017). Texas Medical Center Dissertations (via ProQuest). AAI10265237.
https://digitalcommons.library.tmc.edu/dissertations/AAI10265237