A Large Language Model Assembles a Mental Health Patient's Problem Network from Therapy Conversations Alone

Artificial intelligence is now capable of assembling a mental health patient's individual problem network based solely on therapy conversations. The work of American and German researchers shows that a large language model can identify key psychological processes and their interconnections from therapy speech without the traditionally required, long-term follow-up data.

In current psychotherapy, there is an increasing focus on individualization: treatment is tailored from the patient's own 'network', where different thoughts, emotions, and behavioral patterns are interconnected. Building such a network usually requires dense, longitudinal measurement, which is often unrealistic in clinical practice.

Researchers have now constructed a so-called end-to-end method chain that creates the network directly from therapy recordings. They collected transcripts of 77 therapy conversations and manually annotated them with 3,364 psychological processes and their various dimensions. A psychological process refers to, for example, ways of interpreting oneself and others, regulating emotional states, or behavioral patterns.

With these, as if teaching examples, the large language model was taught to recognize both the processes and their properties using so-called in-context learning. The AI learned the task by being given enough annotated examples, without actual long-term retraining. According to the study, the method achieved a high level of performance.

The results suggest that in the future, therapy could support case formulation and treatment planning with networks produced automatically. This could facilitate the adoption of network-based, individualized treatments even in places where resources for extensive measurement are limited.

Source: Using Large Language Models to Create Personalized Networks From Therapy Sessions, ArXiv (AI).