Machine learning models predict PTSD severity and functional impairment: A personalized medicine approach for uncovering complex associations among heterogeneous symptom profiles

Objective: Posttraumatic stress disorder (PTSD) is a debilitating psychiatric illness, experienced by approximately 10% of the population. Heterogeneous presentations that include heightened dissociation, comorbid anxiety and depression, and emotion dysregulation contribute to the severity of PTSD, in turn, creating barriers to recovery. There is an urgent need to use data-driven approaches to better characterize complex psychiatric presentations with the aim of improving treatment outcomes. We sought to determine if machine learning models could predict PTSD-related illness in a real-world treatment-seeking population using self-report clinical data. Method: Secondary clinical data from 2017 to 2019 included pretreatment measures such as trauma-related symptoms, other mental health symptoms, functional impairment, and demographic information from adults admitted to an inpatient unit for PTSD in Canada (n = 393). We trained two nonlinear machine learning models (extremely randomized trees) to identify predictors of (a) PTSD symptom severity and (b) functional impairment. We assessed model performance based on predictions in novel subsets of patients. Results: Approximately 43% of the variance in PTSD symptom severity (R2avg = .43, R2median = .44, p = .001) was predicted by symptoms of anxiety, dissociation, depression, negative trauma-related beliefs about others, and emotion dysregulation. In addition, 32% of the variance in functional impairment scores (R2avg = .32, R2median = .33, p = .001) was predicted by anxiety, PTSD symptom severity, cognitive dysfunction, dissociation, and depressive symptoms. Conclusions: Our results reinforce that dissociation, cooccurring anxiety and depressive symptoms, maladaptive trauma appraisals, cognitive dysfunction, and emotion dysregulation are critical targets for trauma-related interventions. Machine learning models can inform personalized medicine approaches to maximize trauma recovery in real-world inpatient populations.