A computational approach to measure the linguistic characteristics of psychotherapy timing, responsiveness, and consistency

Abstract

Although individual psychotherapy is generally effective for a range of mental health conditions, little is known about the moment-to-moment language use of effective therapists. Increased access to computational power, coupled with a rise in computer-mediated communication (telehealth), makes feasible the large-scale analyses of language use during psychotherapy. Transparent methodological approaches are lacking, however. Here we present novel methods to increase the efficiency of efforts to examine language use in psychotherapy. We evaluate three important aspects of therapist language use - timing, responsiveness, and consistency - across five clinically relevant language domains: pronouns, time orientation, emotional polarity, therapist tactics, and paralinguistic style. We find therapist language is dynamic within sessions, responds to patient language, and relates to patient symptom diagnosis but not symptom severity. Our results demonstrate that analyzing therapist language at scale is feasible and may help answer longstanding questions about specific behaviors of effective therapists.

Fig. 1. Therapist speech phase-dependence.

The dynamic nature of therapist speech, grouped by language feature category. It represents trends in therapist language over time after aggregating across therapists. LIWC = Linguistic Inquiry and Word Count, a dictionary-based lexicon that maps words and word stems to psychologically relevant categories. EmoLex = Word-Emotion Association Lexicon, a list of English words mapped to crowdsourced sentiment annotations. We performed smoothing/interpolation between discrete points at the level of temporal quintiles using a natural cubic spline. See Fig. 2 for per-feature examples of these trends viewed without smoothing.

Fig. 2. Therapist and patient language within-session changes.

Quantitative assessment of changes in therapist language features over time, as well as within-quintile differences between patient and therapist language. b and c show examples of patient and therapist language features that converged over time. d illustrates a case where patient and therapist language features diverged over time. a highlights a language feature that was significantly different between therapist and patient and neither converged nor diverged over the course of the session. The center line of each boxplot shows the median value for that time bin, while the lower and upper bounds of the box indicate the first quartile (25th percentile) and third quartile (75th percentile), respectively. The lower and upper “whiskers” extend to 1.5x the interquartile range (IQR) beyond the lower and upper quartile, respectively. Observations outside this range are displayed as independent points. All differences annotated with asterisks (*) are significant at level α = 0.05 after controlling for multiple hypothesis tests via the Benjamini-Hochberg procedure. p-value annotation: Non-significant (ns): 0.01 < p ≤ 1.0; *0.01 < p ≤ 0.05; **0.001 < p ≤ 0.01; ***0.0001 < p ≤ 0.001; ****p ≤ 0.0001.

Fig. 3. Therapist responsiveness patterns at the level of individual sessions.

Illustration of significant directional associations between patient language and therapist language in four sessions, each representing a unique patient-therapist dyad. Language features are colored by feature group (see Table 2). Edges are colored according to the average partial correlation coefficient. a illustrates an example of one patient-therapist dyad in which there was just one significant association: increases in patient rate of speech, as measured in words per second, were associated with decreases in therapist rate of speech, and vice versa. b shows a patient-therapist dyad in which the patient’s past-oriented speech and rate of speech had opposite effects on the therapist’s rate of speech. c demonstrates a case where decreases in the patient’s rate of speech led to increases in a diverse array of therapist language features, or vice versa. d highlights a patient-therapist dyad with varied significant associations: increased patient use of third-person plural pronouns (‘“They” Pronouns’) drove increased therapist use of third-person plural pronouns (‘“They” Pronouns’), increased use of positive language by the patient (“Positive”) was associated with increased use of checking for understanding phrases by the therapist (“Checking for Understanding”), etc. These are four of the 73 network diagrams produced, one for each session/patient-therapist dyad.

Fig. 4. Therapist responsiveness patterns aggregated over all sessions.

The number of times a particular type of association between patient language features and subsequent/accommodating therapist language features was found, across all sessions. Patient language features are on the left, therapist language features on the right. For the purposes of illustration, only associations that were found in at least 4 patient-therapist dyads are displayed (see Supplementary Fig. 2 for a similar plot containing all significant associations). There were 72 such associations from 43 unique patient-therapist dyads, of which 24 involved changes in the patient’s rate of speech (“Words per Second”). Language features are colored by feature group (see Table 2). Edges are colored according to the average partial correlation coefficient amongst all patient-therapist dyads in which that association was found. For example, 12 patient-therapist dyads exhibited a significant negative association between patient rate of speech and therapist rate of speech, such that increases in the patient’s words per second (“Words per Second”) were associated with subsequent decreases in the therapist’s words per second (“Words per Second”) and/or vice versa (i.e., decreases in the patient’s words per second were associated with subsequent increases in the therapist’s words per second).