An AI Intervention Recommendation System for Psychotherapists

Abstract

In psychotherapy, therapists usually decide on interventions based on limited information gathered exclusively during therapy sessions.

This project presents a decision-support system (DSS) that models individual psychotherapeutic processes. It combines responses from the Therapy Process Questionnaire (TPQ), daily diary entries, and a brief pre-/post-session evaluation, structured around five therapeutic dimensions defined in the nonlinear change model [1]. Using mathematical modelling and machine learning, the DSS forecasts shifts in those same dimensions — problem severity, therapeutic success, motivation to change, emotions, and insight. Based on those forecasts, the DSS recommends and explains clients' personalized interventions to help therapists tailor each session’s focus.

We created a machine learning pipeline for the DSS, which simulates each client's therapy process. First, we used a large language model to generate diverse synthetic clients and their data. The data included daily diary entries, scores for TPQ dimensions, and a pre‑/post‑session questionnaire every seventh day.

Then we extracted features from each diary that included word and character count, type-to-token ratio, sentiment, readability, lexical diversity, and syntactic complexity. These features were fed into a Random‑Forest regressor that forecasted near‑term shifts on each TPQ dimension. From the features, sentiment consistently proved to be the most influential predictor of the TPQ dimensions.

We expanded the feature space with raw factor scores, Euclidean distances, signed deltas and cosine similarities between therapist‑pre, therapist‑post and client‑post evaluations. Then a script merged questionnaire data, trained Random‑Forest models, and calculated feature importance with permutation tests. This revealed which questionnaire answers best predicted the next day’s TPQ scores.

When real client data is available, our pipeline can be used to evaluate the predictive power of our algorithms, which, if successful, can predict changes in the TPQ dimension. Additionally, we developed a functional prototype for the web application built on a Supabase backend. Therapists and clients can register, complete the TPQ online, enter daily diary entries and, based on that have their five dimension profiles automatically configured.

Limitations of our current work include reliance on synthetic data, which restricts our ability to fully evaluate the algorithmic accuracy. Our future work could involve integrating advanced natural language processing tools, such as BERT, potentially improving the predictive quality of the model. This project translates concepts from non-linear dynamics into a practical tool for therapy, supporting more informed and adaptive clinical decision-making.

References

[1] G. Schiepek, B. Aas, and K. Viol, “The Mathematics of Psychotherapy: A Nonlinear Model of Change Dynamics,” Nonlinear Dynamics Psychol. Life Sci., vol. 20, no. 3, pp. 369–399, Jul. 2016.