For patients with post-traumatic stress disorder (PTSD), written exposure therapy (WET) — in which patients write about a traumatic event under the guidance of a therapist — has shown to be a quick and effective way to lessen psychological distress. However, there are more patients who could benefit from this treatment than there are therapists who have been taught these skills. That is because there’s a gap between the number of therapists who would like to use these skills and trained therapy consultants who can help teach them.
Debra Kaysen, a professor of psychiatry and behavioral sciences at Stanford University, hopes to shrink that gap and improve patient outcomes with a tool called TherapyTrainer. This app uses AI to simulate both a therapy teacher and patients with PTSD to give therapists an on-demand environment to practice these new skills before using them with patients.
“I see AI as a potential major contributor to eliminating this bottleneck in therapy training,” said Kaysen, whose work on TherapyTrainer has been published in Cognitive and Behavioral Practice. “This offers real-time feedback from the AI consultant, and therapists can try again if they want more practice. We can’t do that with real patients!”
Building TherapyTrainer
The purpose of TherapyTrainer is to develop an app that can provide 24/7 access to the same type of expert consultation and clinical practice that can be hard to accommodate with real providers alone, said Elizabeth Stade, a clinical psychologist and research scientist at the Stanford Institute for Human-Centered AI and the lead author on this research.
In TherapyTrainer, therapists experience five mock sessions with AI patients where they can implement WET techniques and then receive feedback from an AI consultant on how well they adhered to the method’s guidelines.
Both the AI consultant and AI patients were developed on OpenAI’s GPT models using prompt engineering as well as WET training materials and input from WET experts. For the patients, Stade and colleagues developed and trained the AI to embody two different profiles: a Latino male combat veteran in his 30s and a middle-aged Black woman with a history of sexual trauma, intimate partner violence, and substance-use disorder.
The team developed the AI consultant in an interactive process with actual WET consultants to ensure that prompts given to the model would solicit appropriate responses that are very aligned with the WET model.
“We gave the model instructions like ‘You are an expert in written exposure therapy. Your job is to review the conversation that the therapist is having with the patient and provide guidance on how to improve the delivery of written exposure therapy,’ ” said Stade.
The TherapyTrainer prototype was then tested in three phases. First, the researchers collected initial impressions from master’s and doctoral clinicians who had varying levels of familiarity with WET techniques. Then they brought the app to a WET workshop where clinicians were able to test it and provide feedback. Finally, they conducted structured user interviews.
Enabling Better Care
Across these different testing phases, Stade said the team received a wide range of feedback.
“User testing is so, so valuable,” she said. “Therapists told us about real problems they encountered while using TherapyTrainer — addressing these things will be crucial for making TherapyTrainer useful in real-world clinical settings.”
For example, some of the feedback the team received was about TherapyTrainer’s user interface. Some therapists found it unnatural to use digital messaging in the sessions instead of speaking with the patient. Therapists also said they wanted more guidance on how to start using TherapyTrainer and wanted a WET manual built into the app as a reference.
They also offered feedback about the accuracy of the AI consultant, said Stade. In particular, the tool had a tendency to suggest improvement on skills that were not core to WET. Stade said this is likely because the app uses an AI model trained broadly on the internet. For example, it suggested the therapists instruct patients to use deep breathing skills to calm themselves when feeling distressed, whereas, in WET, therapists help patients learn that distress naturally decreases over time if they refrain from using avoidance or coping strategies. The team is working now on ways to improve this in the app’s next iteration.
For TherapyTrainer’s next steps, the team will conduct a randomized clinical trial that compares therapist knowledge of WET and confidence in their skills with patient outcomes when a therapist receives WET consultation from a real psychologist versus from the AI consultant.
“If we could show that TherapyTrainer is just as good at delivering consultation as humans, even if it’s not better, that would actually be really meaningful,” said Stade.
This research developed out of the Center for Responsible and Effective AI Technology Enhancement of Treatments for PTSD.
