Content Coding of Psychotherapy Transcripts Using Labeled Topic Models.

Area of Science:

• Computational linguistics
• Psychology
• Machine learning

Background:

• Psychotherapy relies on linguistic interactions between patients and providers.
• Current methods for analyzing these conversations are labor-intensive, prone to errors, and difficult to scale.
• Objective and scalable methods are needed to summarize psychotherapy session content.

Purpose of the Study:

• To compare the performance of a labeled latent Dirichlet allocation (L-LDA) model against a baseline lasso regression model for automated psychotherapy session coding.
• To evaluate the models' ability to predict session-level codes and identify specific text passages representative of codes.

Main Methods:

• Utilized a publicly available psychotherapy corpus of patient-provider conversation transcripts.
• Applied the L-LDA model to learn text-code associations and predict codes.
• Compared L-LDA with a lasso logistic regression model using predictive accuracy and area under the curve (AUC).

Main Results:

• The L-LDA model achieved higher predictive accuracy for session-level codes (average AUC of 0.79) compared to the lasso logistic regression model (average AUC of 0.70).
• Both models could identify specific talk-turns related to symptom codes, though human coders remain more reliable for fine-grained analysis.
• L-LDA demonstrated superior performance over discriminative methods for session-level coding.

Conclusions:

• The L-LDA model presents a promising objective and scalable approach for automated psychotherapy session coding.
• This method shows potential for improving the efficiency and accuracy of psychotherapy research and practice.
• Further development is needed to match human coder reliability for fine-grained, talk-turn level analysis.