Researchers built a pipeline that reads therapy transcripts and spits out a map of a client's psychological issues and how they connect.
The team annotated 8,028 utterances across 77 transcripts from six clients. An LLM first flagged clinically relevant moments in the conversation, then sorted them into categories. A second step grouped those categories into clusters and named them, and a final step drew connections between clusters to form a network - the kind therapists have traditionally built by hand from weeks of diary-style symptom tracking. Evaluators judged the results largely useful, though they did not always agree with each other on how well the model performed, and no one has yet tested whether the AI-built networks actually lead to better treatment.
This matters because personalized treatment planning - matching therapy techniques to a specific client's symptom patterns - has been stuck behind a data problem. Collecting the intensive longitudinal surveys needed to build these networks statistically is slow and often impractical in a normal clinical practice. If a single transcript can produce a comparable map, that removes a major bottleneck to giving more clients tailored care.
It is a proof of concept, not a product: six clients is a small sample, and the researchers themselves want it tested against the statistical networks it is meant to replace before anyone trusts it with real treatment decisions.