Language markers of emotion flexibility predict depression and anxiety treatment outcomes

TL;DR

The study tackles the challenge of predicting non-response to anxiety and depression treatment in real-world care by leveraging passively collected linguistic data from extensive teletherapy transcripts. A transformer-based small language model extracts utterance-level emotions, which are then clustered by a state-space method (VISTA) to reveal distinct emotion-dynamics trajectories, subsequently interpreted via temporal networks. Two groups emerge: Improving ($n=8{,}230$) and Non-response ($n=3{,}813$), with the non-response group showing sadness and fear as dominant drivers and higher odds of deterioration, while the improving group displays balanced positive and negative emotions. The findings demonstrate that emotional inflexibility, as reflected in dynamic linguistic markers, can serve as scalable, interpretable predictors for risk stratification and adaptive care in precision mental health, with an open-source framework enabling replication and extension.

Abstract

Predicting treatment non-response for anxiety and depression is challenging, in part because of sparse symptom assessments in real-world care. We examined whether passively captured, fine-grained emotions serve as linguistic markers of treatment outcomes by analyzing 12 weeks of de-identified teletherapy transcripts from 12,043 U.S. patients with moderate-to-severe anxiety and depression symptoms. A transformer-based small language model extracted patients' emotions at the talk-turn level; a state-space model (VISTA) clustered subgroups based on emotion dynamics over time and produced temporal networks. Two groups emerged: an improving group (n=8,230) and a non-response group (n=3813) showing increased odds of symptom deterioration, and lower likelihood of clinically significant improvement. Temporal networks indicated that sadness and fear exerted most influence on emotion dynamics in non-responders, whereas improving patients showed balanced joy, sadness, and neutral expressions. Findings suggest that linguistic markers of emotional inflexibility can serve as scalable, interpretable, and theoretically grounded indicators for treatment risk stratification.

Figures (5)

• Figure 1: High level depiction of our methodology. Over the course of 12 weeks, patients participate in message-based psychotherapy. We infer patient emotions in each talk turn using a transformer-based language model. VISTA clusters the patients based on temporal dynamics of emotions. For each cluster, we create emotion networks to identify central emotions.
• Figure 2: The result of VISTA with the returned clusters and noiseless trajectory (in red) determined by the optimal parameters along with randomly selected time series from each cluster. The x-axis is in week, while the y-axis is the emotion score in $[0,1]$ returned by DistilRoBERTa-base.
• Figure 3: Likert plots of initial symptoms differences between emotion clusters in Improving (Cluster 1) and Non-response (Cluster 2) groups. Significant difference between the two clusters, as measured by the Mann-Whitney U test, are denoted with asterisks (*), with *** indicating $p < 0.0001$, ** $p < 0.001$, and * $p < 0.05$, where $p$ is the Bonferroni-corrected p-value. Percentages indicate patients in the cluster below (left) or above (right) the clinical threshold (2+) for that symptom.
• Figure 4: Temporal networks for both of the clusters returned by VISTA. Nodes represent emotions extracted from patient transcripts, while edges represent predictive temporal connections among symptoms. Blue edges represent positive association, with red representing negative association. Thicker edges represent stronger associations.
• Figure 5: Out-Expected Influence in temporal network for both of the clusters returned by VISTA.