Supervised machine learning appears to identify 800 different variables that place patients at increased risk for post-traumatic stress disorder (PTSD), according to researchers at NYU.
Supervised machine learning appears to identify 800 different variables that place patients at increased risk for post-traumatic stress disorder (PTSD), according to researchers at NYU. The findings, published online in BMC Psychiatry, support the ability of machine learning to increase prediction versatility as a promising step toward “developing algorithmic, knowledge-based, personalized prediction of post-traumatic psychopathology,” wrote the study authors.
“Our study shows that high-risk individuals who have experienced a traumatic event can be identified less than two weeks after they are first seen in the emergency department,” said Arieh Y. Shalev, MD, Barbara Wilson Professor, Department of Psychiatry, NYU Langone Medical Center, and co-director, Steven and Alexandra Cohen Veterans Center, NYU. “Until now, we have not had a tool—in this case a computational algorithm—that can weigh the many different ways in which trauma occurs to individuals and provides a personalized risk estimate.”
Previous research indicates that PTSD is associated with various, multimodal risk indicators, many of which can be observed shortly after exposure to trauma. However, clinically useful, personalized predictors of PTSD have yet to be identified, in part because completed studies have identified group-level risk factors that overlook within-group heterogeneities and individual paths to PTSD.
To overcome these and other limitations in current research, the study team determined that “forecasting methods of PTSD must accommodate multiple combinations of risk indicators, account for partially available information and use prior knowledge to adjust the relative weights of putative predictors.”
Building on findings from a previous study they conducted, the study investigators sought to evaluate multiple, equivalent, maximally predictive sets of early risk indicators for PTSD by applying a Target Information Equivalence algorithm to uncover “all compact non-redundant sets of items that maximize the prediction of non-remitting PTSD symptom trajectory.” Support vector machines were used to evaluate the accuracy of prediction from each of these sets.
When applied to data collected within 10 days of a traumatic event from adults aged 18 to 70 who had been admitted to the emergency department following potentially traumatic events, the algorithm used in the study identified 789 minimal sets of variables (MBs) before cross-validation that rendered all others non-significant predictors of non-remitting PTSD.
The average number of MBs identified in repeated cross-validations was 800, with an average of 19 features per MB. Thirteen features participated in more than 75% of all MBs. Consistently predictive features included age, emergency department length of stay, head injury, perceived emergency department pain, patient and clinician’s clinical global impression, total PTSD symptom scale and Kessler-6 scores, reporting nightmares, concentration problems, feeling worthless, wanting help, and quality of social support.
“Until recently, we mainly used early symptoms to predict PTSD, and it had its drawbacks,” said Shalev. “This study extends our ability to predict effectively. For example, it shows that features like the occurrence of head trauma, duration of stay in the emergency department, or survivors' expressing a need for help, can be integrated into a predictive tool and improve the prediction…In the future, we hope that we will be better able to tailor treatment approaches based on more personalized risk assessment. PTSD exacts a heavy toll on affected individuals and society.”