4 - (SYM 38) Following Netflix' Example: How a Recommendation System Could Help Improve PTSD Treatment Outcomes

Recommendation systems, such as those used by Netflix for content curation, could help improve mental health treatment. For example, personalized algorithms could be developed around common treatment targets, such as negative posttrauma cognitions in Cognitive Processing Therapy (CPT), and provide clinicians with specific recommendations for cognitions to target in treatment that would lead to the most amount of PTSD symptom change. This presentation details the development of such a recommendation system for CPT.

Data from 1,331 veterans who completed 3- and 2-week intensive PTSD treatment programs (ITPs) were used for analyses. Prior research has shown that individuals improve by approximately 20 PCL-5 points. To build the recommendation system, we first used latent profile analysis to identify different groups of individuals based on their endorsement of individuals Posttrauma Cognition Inventory (PTCI) items as well as PTSD (PCL-5) and depression severity (PHQ-9) at baseline and mid-treatment. We then employed machine learning techniques (Elastic Net and Random Forest) to determine specific PTCI item changes at pre- and mid-treatment within these clusters in which change predicted the greatest amount of change in PTSD severity from pre- to mid- as well as mid- to post-treatment. Predicted changes in PTCI items were then ranked in terms of their importance for each patient and presented to clinicians.

To test the potential effectiveness of the recommendation system, we conducted several case simulations, which suggested that following the system's recommendations could lead to additional reductions in PCL-5 scores for individual cases above the average symptom improvement. For instance, in Case 1 of a simulated combat veterans, focusing on items related to isolation and the inability to feel normal emotions again was recommended. Following the recommendations was predicted to result in an additional PCL-5 reduction of approximately 12 points. Similarly, clinicians working with Case 2’s with similar characteristics were recommended to target items related to feelings of being dead inside and constant vigilance. Simulations suggested that this would improve their symptoms by an estimated additional 9 PCL-5 points above the standard symptom improvement.

The presentation will conclude with a focus on the development of a user-friendly interface for clinicians as well as discussion of a prospective study to assess the effectiveness of recommendation system.