Photo Credit: bsd studio

The following is a summary of “Employing large language models for emotion detection in psychotherapy transcripts,” published in the May 2025 issue of Frontiers in Psychiatry by Lalk et al. 

The following is a summary of “Employing large language models for emotion detection in psychotherapy transcripts,” published in the May 2025 issue of Frontiers in Psychiatry by Lalk et al. 

Researchers conducted a retrospective study to train a large language model (LLM) for detecting emotions in German psychotherapy transcripts to predict symptom severity and therapeutic alliance.  

They utilized a publicly labeled dataset of 28 emotions and fine-tuned a pre-trained LLM for emotion classification and applied the model to 553 psychotherapy sessions involving 124 patients. Using machine learning (ML) and explainable artificial intelligence (AI), the symptom severity and therapeutic alliance based on the detected emotions like depressions were predicted.  

The results showed that the fine-tuned LLM achieved an F1 _macro  score of 0.45, an accuracy of 41% (Accuracy = 0.41), and a Cohen’s Kappa of 0.42 (Kappa = 0.42) across 28 emotions. The ML model predicted symptom severity with a correlation coefficient of r = 0.50 (95%-CI: 0.42, 0.57) and alliance scores with r = 0.20 (95%-CI: 0.06, 0.32). Key predictors for symptom severity included approval, anger, and fear, while curiosity, confusion, and surprise were most relevant to alliance.  

Investigators concluded that the model virtually predicted symptom severity and therapeutic alliance, underscoring the relevance of both negative and positive emotions in psychotherapy outcomes. 

Source: frontiersin.org/journals/psychiatry/articles/10.3389/fpsyt.2025.1504306/full