[Paper Review] Local dynamic mode of Cognitive Behavioral Therapy

Victor Ardulov, Torrey A. Creed+2 more|arXiv (Cornell University)|Apr 28, 2022

Mental Health Research Topics4 citations

TL;DR

This paper proposes a novel framework for automated cognitive behavioral therapy (CBT) competence assessment by modeling therapist-client interactions as local dynamical systems using windowed Dynamic Mode Decomposition with control (DMDc). It extracts dynamic modes from short dialogue segments and correlates them with CTRS scores, achieving up to 68.7% F1 in classifying session-level therapist competence, demonstrating that local dynamics carry interpretable, actionable signals for real-time feedback.

ABSTRACT

In order to increase mental health equity among the most vulnerable and marginalized communities, it is important to increase access to high-quality therapists. One facet of addressing these needs, is to provide timely feedback to clinicians as they interact with their clients, in a way that is also contextualized to specific clients and interactions they have had. Dynamical systems provide a framework through which to analyze interactions. The present work applies these methods to the domain of automated psychotherapist evaluation for Cognitive Behavioral Therapy (CBT). Our methods extract local dynamic modes from short windows of conversation and learns to correlate the observed dynamics to CBT competence. The results demonstrate the value of this paradigm and outlines the way in which these methods can be used to study and improve therapeutic strategies.

Motivation & Objective

• Address the critical need for scalable, equitable access to high-quality mental health care by reducing reliance on expert-led, time-intensive therapist evaluations.
• Overcome limitations of traditional CTRS scoring by enabling timely, data-driven feedback to clinicians during therapy sessions.
• Develop a method that captures momentary therapeutic strategies through the dynamics of therapist-client interactions, rather than static features.
• Enable real-time, interpretable feedback by modeling dialogue as a control-affine dynamical system with local, windowed analysis.
• Demonstrate that local dynamic modes derived from BERT embeddings and DMDc can predict CTRS competence with high fidelity, supporting clinical training and quality improvement.

Proposed method

• Apply windowed Dynamic Mode Decomposition with control (DMDc) to short segments of therapist-client dialogue, treating each window as a local dynamical system.
• Use BERT embeddings of utterances to represent linguistic content and extract state vectors for each interaction window.
• Model the interaction as a control-affine system, where therapist utterances are control inputs and client responses are state transitions.
• Perform eigenvalue decomposition on the DMDc operator to extract local dynamic modes that represent momentary therapeutic strategies.
• Aggregate dynamic modes across windows using sum or average pooling, followed by classification using models like GNB, LR, or SVM.
• Train and evaluate models on session-level CTRS scores, using T+C (text + client response) inputs to capture both therapist and client dynamics.

Experimental results

Research questions

• RQ1 Can local dynamic modes extracted from short windows of CBT dialogue predict therapist competence as measured by the CTRS?
• RQ2 Do the dynamic modes derived from DMDc analysis of therapist-client interactions carry interpretable, clinically relevant signals about therapeutic strategies?
• RQ3 How do different input modalities (e.g., text-only vs. text + client response) and aggregation strategies affect model performance in predicting CTRS scores?
• RQ4 Can the dynamical system framework support real-time tracking of session quality and enable timely feedback during therapy?
• RQ5 Is the linear dynamical system assumption sufficient to capture meaningful therapeutic dynamics, and how does it compare to more complex models?

Key findings

• The T+C (text + client response) input configuration achieved the highest performance, indicating that both therapist language and client response dynamics are critical for competence prediction.
• The best-performing model achieved an F1 score of 0.6868 on the 'pt' (patient) sub-score, demonstrating strong discriminative ability for session-level competence.
• Models using a window size of w=8 and nλ=7 eigenvalues performed best, suggesting that capturing moderate temporal context improves predictive power.
• The sum aggregation method outperformed average and mean pooling, particularly when combined with logistic regression, indicating that cumulative dynamic patterns are informative.
• The GNB + LR + sum + w=8 + nλ=7 configuration achieved the highest F1 (0.6868), showing that simple, efficient models can effectively leverage dynamic modes.
• The results suggest that local dynamic modes can serve as interpretable, real-time indicators of therapeutic strategy quality, supporting future feedback systems.