A LANGUAGE-DERIVED AFFECTIVE SENTIMENT DIMENSION OF KETAMINE RESPONSE IN DEPRESSION

Background

Ketamine produces rapid antidepressant effects, yet PHQ-9 total scores obscure heterogeneity in how patients improve. As psychiatry shifts toward dimensional, mechanism-informed models, language offers a direct affective readout that traditional scales may miss. Sentiment analysis can capture these signals, but it is unknown whether sentiment follows reliable treatment trajectories or which patient factors shape its change. Using a large real-world cohort, this study evaluates sentiment as a distinct affective dimension of ketamine response and identifies clinical characteristics that moderate its improvement.

Methods

We analyzed a cohort of 535 adults (mean age 42.9 [SD 14.1]; 55% women) with treatment-resistant depression treated with IV ketamine across a large outpatient network, including major depressive disorder (n=491) and bipolar depression (n=44). Most participants (n=535) completed the standard four-infusions-in-14-days protocol; 234 received offprotocol dosing or discontinued early. PHQ-9 scores were collected at baseline (≤30 days pre-treatment), infusion visits #1 and #4, and +3, +9, and +14 days post-treatment. At intake, patients reported comorbidities including anxiety (n=418), OCD (n=72), PTSD (n=172), SUD (n=56), and suicidal ideation (n=289), as well as current psychotropic use such as antidepressants (n=345), antipsychotics (n=71), benzodiazepines (n=159), nonbenzodiazepine sedatives (n=78), mood stabilizers (n=88), stimulants (n=114), and other agents (n=68). Free-text responses collected at the same visits as part of the PHQ-9 were tokenized, mapped to AFINN lexicon scores and averaged to produce a single sentiment score. Sentiment trajectories were modeled using a mixed-effects framework as PHQ-derived factors, with log-transformed time as the primary predictor, identical covariates, and random participant intercepts. Fixed effects included age, sex, diagnosis, comorbidities and medication exposure. Secondary analyses examined moderators comorbidity and medication exposure.

Results

Sentiment increased significantly over time (β=0.94, 95% CI 0.63–1.25, p < 0.001). Higher baseline PHQ severity (β=−0.08, p=0.001) and mood stabilizer exposure (β=−0.80, p=0.015) were associated with lower overall sentiment but did not alter its rate of improvement. Two comorbidities moderated trajectories: anxiety predicted a steeper increase (Time×Anxiety β=0.91, 95% CI 0.18–1.64), whereas OCD markedly blunted improvement (Time×OCD β=−1.25, 95% CI −2.10 to −0.40).

Conclusion

Language-derived sentiment captured an affective dimension of ketamine response not reflected in traditional symptom scores. Although sentiment improved across treatment, its level and trajectory varied by clinical profile, enhanced in patients with anxiety and attenuated in those with OCD. These findings demonstrate that linguistic signals provide complementary insight into emotional change and reveal moderators distinct from PHQbased domains. Integrating language-derived measures with conventional scales may sharpen mechanistic understanding and support more individualized treatment. Overall, these Results underscore the promise of language-based models as scalable tools for phenotyping response and advancing precision strategies in depression.