Who drops out of exposure therapy? A machine learning mega-analysis of PTSD patients

Posttraumatic stress disorder (PTSD) is a global mental health challenge. Although several types of evidence-based exposure therapies have been recognized as first line interventions for PTSD, high dropout rates remain of critical concern and are a significant barrier towards recovery. Early identification of patient profiles and treatment properties that may contribute to dropout would increase treatment efficacy. However, the body of literature on dropout is marked by large variability. This may be caused by methodological constraints such as limited sample sizes and inadequate statistical techniques. To address those issues, we conducted a machine learning mega-analysis on 803 PTSD patients who received a form of exposure therapy, with the aim to predict who will later discontinue treatment. From various predictors such as age, education, initial symptom severity, trauma type, comorbidities and protocol variations (number of sessions, imaginal or in-vivo exposure, telemedicine or in person), our gradient boosted decision tree model classified dropouts above chance levels (accuracy = 60%). We found that comorbid depression, substance use, telemedicine, lower education and younger age increased the chances of dropping out. However, a substantial portion of the variance remains unexplained, which emphasizes the need to improve data collection strategies to attain larger samples and a wider set of predictors. This improvement is essential for reaching better prediction performance to a level where findings become clinically applicable. This study showcases the promise of machine learning in refining treatment protocols and improving patient retention, facilitating more personalized therapeutic strategies in this critical area of mental health.