(PS10-25) Machine Learning in the Prediction of Posttraumatic Stress: Application of Random Forests

Background: The definition of clinical trauma has been refined across editions of the Diagnostic and Statistical Manual of Mental Disorders (DSM) in an effort to clarify the characteristics of qualifying events for posttraumatic stress disorder. Less studied, however, are the methods used to quantify the severity of traumatic events. While the assessment of severity is crucial in testing hypothesized dose-response relations between the intensity of exposures and subsequent outcomes, research is limited by inconsistent conceptualizations of event severity and problematic measurement. Common operationalizations of severity include raw counts of lifetime exposure types (e.g., sum of endorsed domains on the Life Events Checklist), the frequency of exposure (e.g., number of times an event occurred), the duration of exposure (e.g., period of chronic victimization), the severity of corresponding injuries, and the intensity of peritraumatic responses (e.g., ratings of fear, helplessness, certainty of death). While existing metrics capture different aspects of severity, all exhibit limitations when considered in isolation. The aims of the current study are to utilize novel, machine learning techniques to (a) identify unique combinations of severity indicators that maximally predict post-trauma outcomes and (b) use results to inform current conceptual models of trauma severity.


Method: Planned analyses will utilize archival data obtained from undergraduate survivors (N = 486) completing interview-based mental health assessments as part of an ongoing study. Participants in the existing set are primarily female (n = 376, 77.9%) and non-Hispanic White (n = 385, 81.9%). Index events identified during interview measures include sexual assault (40%), accident and disaster trauma (26%), physical and/or threatened violence (19%), unexpected injury or death of a loved one (12%), and military trauma (3%). Predictors extracted from the archival set will be participant demographics, trauma characteristics, and mental health history. Trauma characteristics include event type, frequency, duration, age of exposure, injury, and subjective emotional response across self-report and interview measures. Functional outcomes will include interview-based PTSD (CAPS-5), posttrauma cognitions (PTCI), negative affectivity (PANAS, BDI-II), and overall global functioning (i.e., Barkley Functional Impairment Scale). Preliminary analyses indicate moderate symptoms of PTSD with a high degree of variability in CAPS-5 scores (M = 19, SD = 13.8). Random forests, an ensemble learning method that aggregates many decision tree models fit to bootstrapped samples of the original data, will be used to predict post-trauma outcomes. Variable importance metrics will be calculated to identify the set of predictors most central to accounting for variability in each outcome; 5 -fold cross-validation will be used to mitigate overfitting. Findings from this study will establish an exploratory foundation for future conceptual exploration and analysis aimed at testing indicators of event severity.