Scientists used AI to analyze psychotherapy sessions — and the results were surprising

New research provides evidence that certain patterns of speech and expressions can predict the bond between therapists and their patients. The study, published in iScience, utilized artificial intelligence to analyze the use of personal pronouns and hesitations in speech during psychotherapy sessions.

The motivation behind this research stemmed from a longstanding challenge in psychotherapy: accurately assessing and enhancing the therapeutic alliance. This alliance is the mutual understanding and partnership between a therapist and their patient, recognized as a cornerstone of effective therapy.

Traditional methods for gauging this relationship have relied heavily on subjective self-reports and observer interpretations, which are time-consuming and often fail to capture the dynamic nature of therapy sessions. With the advent of machine learning in healthcare, researchers at the Icahn School of Medicine at Mount Sinai saw an opportunity to objectively study the elements of therapeutic communication, aiming to identify clear, actionable markers of a strong therapeutic alliance.

The researchers recruited a diverse group of participants, comprising 28 patients undergoing psychotherapy and 18 therapists from outpatient clinics at academic hospitals in New York City. These participants were involved in 28 distinct therapy sessions that spanned a range of psychiatric conditions, ensuring a broad representation of therapeutic encounters.

Before diving into the therapy sessions, the study employed a preparatory phase where patients assessed their alliance with previous therapists and their attachment styles through online surveys. This step was crucial in setting the baseline for understanding individual predispositions towards therapeutic relationships.

The therapy sessions themselves were recorded using two wireless microphones per session, enabling a clear separation of patient and therapist speech for later transcription and analysis. This setup was critical for capturing the nuanced verbal exchanges integral to therapy.

To assess the therapeutic alliance, the study employed the Working Alliance Inventory – Short Form, a validated questionnaire that measures the strength of the alliance from the perspectives of both the patient and the therapist. This tool evaluates key components of the therapeutic relationship, including agreement on therapy goals, tasks, and the emotional bond between patient and therapist.

Additionally, the trust game, a behavioral economics paradigm, was introduced as an innovative measure of trust and reciprocity within the therapeutic relationship. This game involved simulated monetary exchanges between participants, offering a novel, quantitative measure of trust that complemented the subjective assessments from the questionnaires.

The heart of the study’s methodology lay in its use of natural language processing (NLP) techniques to analyze the therapy session transcripts. The researchers focused on identifying specific language features, such as the use of personal pronouns (“I,” “we”) and markers of non-fluency (e.g., “um,” “like”), which were hypothesized to reflect aspects of the therapeutic alliance.

The researchers found that the use of personal pronouns by therapists and patients was intricately linked to the strength of the therapeutic alliance. Specifically, therapists’ use of “we” was associated with lower alliance ratings by patients, particularly in the context of personality disorders. This finding challenges the conventional wisdom that inclusive language naturally fosters a stronger connection.

Similarly, both therapists and patients’ frequent use of “I” was negatively correlated with alliance ratings, highlighting the potential drawbacks of overemphasis on the self within therapy sessions. In line with these findings, the trust game game revealed that patients’ repayment behaviors — a proxy for trust — were negatively correlated with therapists’ use of both “we” and “I.”

The researchers also explored the semantic contexts of pronoun usage. Phrases containing “I” followed by action or thought verbs (e.g., “I think,” “I do”) were frequently used in sessions characterized by lower alliance ratings. This points to the importance of how therapists and patients articulate their thoughts and feelings, with certain patterns of self-referential speech potentially hindering the development of a strong therapeutic bond.

An intriguing aspect of the findings pertained to non-fluency markers, such as hesitations or filler words (“um,” “like”), used by patients. Contrary to the often negative perception of such speech patterns in other contexts, the study found that these markers were positively correlated with higher alliance ratings. This suggests that moments of non-fluency in patients’ speech may contribute to a more authentic and engaged therapeutic interaction, possibly because they reflect genuine thought processes and emotional honesty.

“Our study provides the first computational evidence that both first-person pronoun and non-fluency are potential language markers that are predictive of therapeutic alliance and interpersonal trust during psychotherapy treatment,” the researchers wrote.

While the study offers groundbreaking insights into the language of therapy, the researchers acknowledge several limitations. The study had a relatively small sample and the observational nature of the study means causality cannot be inferred. For instance, while the data suggest a potential negative impact of therapists’ use of “we” on the therapeutic alliance, this correlation does not necessarily imply causation. It’s plausible that a higher frequency of “we” in a therapist’s language could reflect an attempt to bridge a gap and foster a sense of unity with a patient who may be more reserved or distant.

Future research is poised to build on these preliminary insights, potentially incorporating larger sample sizes and exploring the context in which specific language patterns occur to deepen our understanding of effective therapeutic communication.

The study, “A natural language processing approach reveals first-person pronoun usage and non-fluency as markers of therapeutic alliance in psychotherapy“, Jihan Ryu, Stephen Heisig, Caroline McLaughlin, Michael Katz, Helen S. Mayberg, and Xiaosi Gu.