Natural language signatures of psilocybin microdosing

This double-blind placebo-controlled trial (n=34) assessed natural language as a resource to identify speech produced under the acute effects of psilocybin microdoses (0.5g dried mushroom), focusing on variables known to be affected by higher doses: verbosity, semantic variability and sentiment score. Verbosity and sentiment scores significantly differed between groups suggesting that microdosing can be identified from natural speech.

Abstract

“Serotonergic psychedelics are being studied as novel treatments for mental health disorders and as facilitators of improved well-being, mental function and creativity. Recent studies have found mixed results concerning the effects of low doses of psychedelics (microdosing) on these domains. However, microdosing is generally investigated using instruments designed to assess larger doses of psychedelics, which might lack sensitivity and specificity for this purpose. Following a double-blind and placebo-controlled experimental design, we explored natural language as a resource to identify speech produced under the acute effects of psilocybin microdoses, focusing on variables known to be affected by higher doses: verbosity, semantic variability and sentiment scores. Except for semantic variability, these metrics presented significant differences between a typical active microdose of 0.5 g of psilocybin mushrooms and an inactive placebo condition. Moreover, machine learning classifiers trained using these metrics were capable of distinguishing between conditions with high accuracy (AUC close to 0.8). Our results constitute first proof that low doses of serotonergic psychedelics can be identified from unconstrained natural speech, with potential for widely applicable, affordable, and ecologically valid monitoring of microdosing schedules.”

Authors: Camila Sanz, Federico Cavanna, Stephanie Muller, Laura A. de la Fuente, Federico Zamberlan, Matias Palmucci, Lucie Janeckova, Martin Kuchar, Facundo Carrillo, Adolfo M. Garcia, Carla Pallavicini & Enzo Tagliazucchi

Summary

Abstract

A double-blind and placebo-controlled experimental design was used to investigate the effects of low doses of psilocybin mushrooms on verbosity, semantic variability and sentiment scores. Machine learning classifiers were trained to distinguish between active and placebo conditions with high accuracy.

Psychedelic microdosing involves consuming small amounts of serotonergic compounds, approximating the perceptual threshold, two to four times per week. It is used for the self-treatment of mental health disorders such as depression or anxiety, and has been suggested as a model for the clinical use of psychedelics. Studies supporting the positive effects of microdosing are generally self-selected and lack adequate control conditions. Double-blind and placebo-controlled experimental designs have found less support for positive outcomes of microdosing.

Studies of full doses of psychedelic substances might not capture the subtler effects induced by microdosing, thus raising the need for novel methods to study low doses. We adopted a non-invasive approach based on natural language processing to investigate the contents of the drug-elicited experience, and to investigate the modulation of language production itself, which can be informative of drug action beyond what is reported during the subjective acute effects. There are no works to date applying NLP to the study of microdosing. We conducted a double-blind placebo-controlled experiment with 0.5 g dried material of Psilocybe cubensis mushrooms and obtained two metrics: verbosity and semantic variability. We also investigated the mean sentiment score to account for the purported effects microdosing on mood.

Participants were enrolled via social media or word of mouth and reported past experiences with serotonergic psychedelics. They were also asked to sign an informed consent form.

Participants in this study stopped their habitual use of psychoactive drugs (including alcohol and caffeine) during the duration of the experiment. They did not receive financial compensation for their participation. The experiment followed a double-blind placebo-controlled design, where participants consumed either capsules with the active dose or the placebo. They were interviewed on Fridays, approximately 2.30 h after dosing, and were asked to identify the condition corresponding to that week.

A dose of 0.5 g of ground dried Psilocybe cubensis mushrooms contained 0.32 mg of psilocybin, 0.48 mg of psilocin, 0.025 mg of baeocystin and 0.0063 mg of norbaeocystin. The active ingredients may have been lost during the month between the end of the experiment and the chemical analysis. We recorded and manually transcribed interviews with participants, and afterwards reviewed the transcripts by a member of the research team. We computed verbosity scores by counting the number of words produced by the participants when answering the questions of the interview.

We computed a metric of semantic variability based on the distance between consecutive words spoken by the subjects, and then calculated the variability of this time series to correspond to the semantic variability.

Sentiment analysis was implemented in Python using a sentence-level model pre-trained for Spanish. It consisted of standard text pre-processing steps, a univariate feature selector and a Multinomial Naive Bayes classifier trained using labeled data obtained from different websites, comprising approximately 1 million training samples.

2.4. Statistical analyses

Statistical analyses were conducted using non-parametric Mann-Whitney U tests to determine the difference in median verbosity, semantic variability, and MSS between groups and subgroups determined by the conditions active dose vs. placebo, and blinded vs. unblinded. Boxplot representations were used to graphically summarize the results.

Machine learning models were trained using the scikit-learn library to classify samples according to the experimental condition and its unblinding. Six models were implemented, two to distinguish between experimental conditions and their blinding, and four to perform the same classification but restricted to four combinations. We divided the samples into 3 folds via stratified cross-validation, and used the average of the area under the ROC curve (ROC AUC) across folds to determine the model performance. We repeated this procedure 100 times with and without shuffling of the sample labels.

We did not find statistically significant differences in semantic variability between the active dose and the placebo condition. Participants under the active dose exhibited higher verbosity in their answers to all questions in the interview. Participants under the active dose reported longer answers and more positive sentiment in their speech when reporting changes to perception, mood and creativity, compared to participants under the placebo condition.

3.1.2. Blind vs unblind

Blinded and unblinded participants showed similar verbosity and MSS values, with blinded participants showing higher median verbosity values.

We used Mann-Whitney U tests to compare verbosity, semantic variability and sentiment scores for each subgroup. When restricted to blinded participants only, the classifier yielded an AUC value of 0.48 0.13, which did not differ significantly from the value obtained by shuffling labels. However, when restricted to unblinded participants only, the classifier yielded an AUC value of 0.51 0.17)

Discussion

We measured the effects of psychedelic microdosing on natural speech using semantic variability, verbosity, and sentiment scores. No significant differences were found between groups.

Machine learning was used to classify participants into active and placebo groups. Results showed that active doses of psilocybin did not affect semantic variability, suggesting that low doses might not be effective in enhancing creativity. Microdosing users generally report an improvement in their mood, which is consistent with increased MSS of natural speech, suggesting that microdosing could be capable of inducing a state of positive mood, which is consistent with increased enthusiasm, motivation and energy during the acute effects of the microdose.

18 The comparison between blinded and unblinded participants suggests that expectations play an important role in the perceived effects of microdosing, but that these effects are not apparent in the comparison between blinded and unblinded participants restricted to the active dose. Natural language processing tools can be used to reveal significant differences and obtain a robust classification of conditions in patients taking low doses of psychedelics. In this study, we used natural language processing to characterize natural language produced under effects of low doses of psilocybin. We found that certain markers from unstructured and unconstrained speech are compatible with improved mood of the participants, and can be extracted remotely and automatically.

Acknowledgments

We thank Sol Pérez Vázquez for her assistance with the logistics of this study, which was funded by the Ministry of Health of the Czech Republic and the Grant Agency of the Czech Republic.