Natural speech algorithm applied to baseline interview data can predict which patients will respond to psilocybin for treatment-resistant depression

Quantitative analysis of natural speech has advanced in recent years and is increasingly applied to psychiatric problems. Previous work has used automated measures of speech coherence and emotion to identify conditions such as schizophrenia and mood disorders, and to predict clinical trajectories; these studies indicate that language features can serve as objective diagnostic and prognostic markers. However, it remained unclear whether pre-treatment speech patterns could predict who will respond to psychedelic-assisted therapy for depression. Carrillo and colleagues set out to test whether automated natural language analytics combined with machine learning could predict clinical response to psilocybin in patients with treatment-resistant depression (TRD). The study applied emotional-sentiment measures to autobiographical interview transcripts collected before treatment and used a classifier to distinguish healthy controls from depressed patients and to predict which patients later responded to psilocybin treatment.

This was an open-label clinical study, sponsored by Imperial College London, in which patients with TRD received two oral doses of psilocybin (10 mg then 25 mg) one week apart, with psychological support provided before, during and after dosing. Seventeen patients completed the study (mean age 44.59, SD 10.97, 5 females). Eighteen age- and sex-matched healthy controls were recruited separately (mean age 36.44, SD 17.23, 7 females). The primary clinical outcome was the Quick Inventory of Depressive Symptoms (QIDS-16) measured at baseline and again 5 weeks after the 25 mg dose; treatment response was predefined as a ≥ 50% reduction in QIDS score at 5 weeks. The extracted text does not detail other inclusion/exclusion criteria or randomisation because the trial was open-label; the authors direct readers to the main trial publication and online supplement for further procedural details. Participants completed an autobiographical memory test (AMT), a structured interview that elicits specific memories in response to cue words. Patients undertook the AMT approximately 2 weeks before the first psilocybin dose, while controls completed the task at their convenience. All AMT sessions were audio recorded and transcribed. The interviews typically took about 10 minutes to administer. Transcriptions were performed by named research staff (L.F., J.S. and P.A.), and two balanced versions of the AMT were used across subjects. For speech analysis the researchers applied an automated Emotional Analysis algorithm that assigns positive and negative sentiment scores to individual words (decimal values between 0 and 1). For each transcript they calculated two features: AVG P (average positivity across words) and AVG N (average negativity). A Gaussian Naive Bayes classifier, a probabilistic machine-learning algorithm, was used to represent each subject by these two features and to perform two classification tasks: patients versus controls, and responders versus non-responders. A 7-fold cross-validation scheme was applied to estimate classification performance and permutation testing was used to assess whether accuracies exceeded chance; details of the cross-validation and permutation procedures are provided in the supplement according to the extracted text.

The speech analytics distinguished depressed patients from healthy controls primarily via lower use of positive emotional words. Controls had a higher AVG P (0.0532 ± 0.013) than patients (0.0384 ± 0.011), a difference that reached significance (t-test p = 0.0011). AVG N did not differ significantly between groups (p = 0.4). Using AVG P and AVG N as inputs to a Gaussian Naive Bayes classifier with 7-fold cross-validation, the model classified patients versus controls with a mean accuracy of 82.85%, precision 0.82, recall 0.82 (sensitivity 0.82), and specificity 0.83. Permutation testing (1,000 trials) indicated this accuracy was significantly greater than chance (p < 0.05). For the main clinical question, 7 of 17 patients (41%) met the predefined response criterion (≥ 50% reduction in QIDS at 5 weeks) and were labelled responders; 10 were non-responders. Mean AVG P and AVG N did not differ significantly between responders (AVG P 0.0334 ± 0.0132; AVG N 0.0368 ± 0.0088) and non-responders (AVG P 0.0418 ± 0.0091; AVG N 0.041 ± 0.0082). Despite the lack of a simple univariate difference, the same classifier trained on AVG P and AVG N achieved an accuracy of 85% for predicting responders versus non-responders (precision 0.75) using 7-fold cross-validation. Permutation testing placed this accuracy in the upper 97th percentile of the null distribution, indicating significance at p < 0.05. The authors note that responders tended to use fewer emotional words at baseline, especially fewer positive words, which they suggest may reflect greater capacity for change in those patients.

Carrillo and colleagues interpret the findings as evidence that short, structured autobiographical interviews analysed by automated natural language tools can both detect depressive status and predict clinical response to psilocybin in TRD. They emphasise that the AMT required little time to administer (around 10 minutes) yet provided features sufficient for above-chance classification and prediction when combined with a simple machine-learning classifier. The discussion situates these results within broader work showing that language features and the quality of acute psychedelic experiences relate to clinical outcomes; the authors note that psilocybin’s idiosyncratic acute effects have previously been linked to longer-term response. They further argue that, given the low marginal cost of applying automated speech-analysis software, these methods could offer a cost-effective screening tool to identify individuals more likely to benefit from psychedelic treatment. Finally, the authors propose directions for future research, specifically testing the specificity of the observed relationships and assessing whether the predictive associations generalise to other interventions and clinical outcomes.