Loading page…
The assessment aid aims to free clinician time while preserving human judgment. Early performance was encouraging, but the workshop projects remain proofs of concept.
Listen to this story
At a VA Boston enterprise AI workshop, researchers tested a generative-AI prototype intended to organize evidence from PTSD assessments, not make diagnoses. Dr. Amar Mandavia said early results appeared comparable to clinician performance, but adding training data made the system too rigid; VA provided no scores or test design, so that comparison is not evidence of diagnostic accuracy. The setback highlights a practical model-development tradeoff for clinical tools, while the work remains a proof of concept constrained by available models and resources.
The Clinician-Administered PTSD Scale for DSM-5 (CAPS-5) can take up to an hour to administer, with additional time needed to interpret results.
Mandavia’s goal is to free clinician time and reach more Veterans while keeping clinical judgment central to diagnosis.
VA Boston was the first VA medical center to host the enterprise AI workshops, whose projects also addressed scheduling, transportation, research support, and other workflows.
VA Boston brought clinicians, researchers and technology experts together to explore how AI might help Veteran care—including the demanding work of assessing post-traumatic stress disorder. The projects remain proofs of concept. In its October 6 account, VA News described both an encouraging initial result and a setback: more training data made the PTSD tool too rigid.
A PTSD diagnosis requires specialized expertise and a structured assessment that can take up to an hour to administer. Interpreting the results takes additional time. The workshop’s PTSD project explored whether generative AI could help providers analyze and organize that information more efficiently, while leaving clinical judgment central to the decision.
Dr. Amar Mandavia, a clinical psychologist and informatician with VA’s National Center for PTSD, explained that providers use validated assessment tools. These include the Clinician-Administered PTSD Scale for DSM-5, or CAPS-5, which requires them to bring together substantial amounts of information and apply professional judgment. His goal is to free clinician time and reach more Veterans, not replace clinical expertise.
Our hope is that we can free up some clinician time and still be confident whether a Veteran meets diagnostic criteria.
Dr. Amar Mandavia, clinical psychologist and informatician, VA National Center for PTSD
Mandavia said early testing showed the PTSD tool performing at a level comparable to a clinician. But adding more training data made it too rigid. VA’s account provides no numerical results or testing details for that comparison, so it should not be read as an established measure of diagnostic accuracy.
Some of the team’s larger ambitions also exceeded what current models and resources could support. Mandavia said hands-on testing helped reveal those boundaries.
VA Boston was the first VA medical center to host the enterprise AI workshops. Each project started with a challenge identified by frontline employees, with teams seeking practical solutions that could eventually scale across VA.
Other workshop projects explored operating-room scheduling, Veteran transportation, research support, and additional clinical and administrative workflows. Dr. Kieran Warner, VA Boston’s chief health informatics officer, described the aim as supporting clinical decisions, improving care coordination and finding efficiencies that create value for Veterans.
Loading discussion...
Join the conversation
Explain where consistency should give way to clinical judgment.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.