1 - (SYM 77) A Precision Medicine Framework for Predicting Response to Evidence Bases Therapies for PTSD: A Machine Learning Approach

Background: Although several gold-standard interventions exist for PTSD (e.g., prolonged exposure therapy (PE), cognitive processing therapy (CP), etc.), a significant proportion of patients fail to achieve successful response to treatment. Thus, there is an urgent need for improved knowledge about who will likely benefit from evidence-based interventions for PTSD. Predictions may be improved by embracing a precision medicine approach that incorporates advances in computational psychiatry such as machine learning.

Methods: In the current study, a 10-fold nested cross-validation elastic net machine learning approach was used to predict both primary (i.e., PTSD response) and secondary (i.e., reliable change in guilt, depression response, and anger response) treatment outcomes across four clinical trials (n=482) in which patients received either CPT or PE. In instances of class imbalance for the outcome variables, the SMOTE procedure was adopted.

Results: Results indicated that machine learning models were able to successfully predict changes in guilt and anger (AUCs= 0.84 to 0.71), yet they were not able to do so for PTSD and depression response status (AUCs = 0.57 to 0.59). Across the models, childhood physical abuse was the strongest and most consistent predictor of poor treatment outcomes.

Discussion: Overall, the study highlights how machine learning can predict distinct outcomes with differential success. Although machine learning models did not significantly outperform chance in distinguishing traditional responder status based on PTSD and depression symptoms, predictive modelling techniques may offer incremental value in determining who might achieve benefit with respect to reductions in guilt and anger. Future directions include refining precision medicine prediction efforts by way of considering a broader array of predictor modalities (e.g., electronic health records, biological factors, etc.) to augment model performance.