Machine learning, functional connectivity, and suicidality in posttraumatic stress disorder

Abstract
Objective
Individuals with posttraumatic stress disorder (PTSD) are at elevated risk for suicide and prior work suggests that patterns of connectivity within brain networks such as the default mode, attention, and central executive networks are associated with risk of suicide. Suicide risk remains difficult to predict despite existing standardized screening tools. Existing screening tools frequently rely on individuals’ awareness of their internal experience and willingness to disclose thoughts and plans related to suicide. Objective biomarkers of suicide risk are, therefore, urgently needed. Powerful machine learning approaches have been used to identify behavioral correlates of suicide risk in PTSD. However, there have been no published studies leveraging machine learning and functional neuroimaging data to identify neural biomarkers predictive of suicidality specifically in individuals with PTSD. To address this significant gap, we conducted the present investigation to examine predictive neural correlates of suicidality in a high-risk group of individuals with PTSD by leveraging machine learning tools coupled with functional neuroimaging data.

Methods
Participants were 65 adults with histories of childhood abuse-related PTSD and varying levels of dissociative symptoms. The Clinician Administered PTSD Scale for DSM-5 was used to diagnose PTSD, and item 9 of the Beck Depression Inventory-II was used to index suicidality. Participants also completed a 20 min functional magnetic resonance imaging scan. We identified 92 individualized regions of interest for each participant from which we computed functional brain connectivity matrices. We then trained a support vector regression machine learning model to predict suicidality from patterns of functional connectivity.

Results
Our model identified seven connections that were central to predicting suicidality. Increased connectivity between regions in visual and dorsal attention networks was linked to decreased suicidality. In contrast, increased connectivity between central executive and attention networks was linked to increased suicidality.

Conclusion
This study is the first of its kind in PTSD and identified patterns of brain functional connectivity that predicted suicidality among this high-risk group. In particular, connectivity between visual and dorsal attention networks may be linked with increased working memory and problem-solving capacity, thereby decreasing suicide risk. In contrast, greater connectivity between central executive and attention networks may be a sign of cognitive rigidity that leads to increased suicide risk. Overall, these findings underscore the importance of diagnosis-specific investigations to fully elucidate neural predictors of suicidality.