A Joint Logistic Regression and Markov Chain Model with Application in Predicting 6-Month Outcome after Severe Traumatic Brain Injury
Traumatic brain injury is a major public-health problem worldwide. It is considered one of the most disabling injuries since it can causes long-term irreversible damage. Because TBI can lead to mortality and has detrimental effects on outcomes, determining prognosis of patients with TBI in the acute injury period is fundamental. Prognostic models for long-term functional outcomes on patients with TBI have been built in the past. Gold standard models were built on admission characteristics but are not useful for clinical-decision making because they have not always achieved good performance. Discrimination ability of these prognostic models could be improved by incorporating secondary insults that occur during the acute phase of injury. In this research, we built joint prognostic models for binary and ordered categorical outcomes that simultaneously combine baseline or non-dynamic characteristics with longitudinal covariate information. The longitudinal data is modeled with continuous-time Markov chains, and the transition rates that characterize the process are included as predictors in a logistic regression model to predict the response. We applied these models to populations of patients with TBI to predict a long-term functional outcome measured in two different scales, and showed that the information provided by physiological changes during the first five days post-injury helped improved predictive power over models based on baseline data only. When the outcome was modeled in an ordinal scale, we observed an increase in statistical power to detect significant effects of the Markov chain on the outcome. Lastly, we aimed to build prognostic models for a long-term binary outcome on a large population of patients with TBI, using as predictors baseline characteristics as well as several summary measures of injury severity and physiological data computed over the first 24 hours after injury. We hypothesized that patient physiology would be a major contributor of outcome prognosis. The transition rates of the previously developed model were included as subject features in this model. Our model achieved 84% area under the ROC curve in an independent population, and in contrast to our hypothesis, baseline and injury severity variables were found to be the most important predictors by the first day post-injury.^
Rubin, Maria Laura, "A Joint Logistic Regression and Markov Chain Model with Application in Predicting 6-Month Outcome after Severe Traumatic Brain Injury" (2017). Texas Medical Center Dissertations (via ProQuest). AAI10685264.