Mixed-methods analysis on psychedelic-augmented meditation experiences from a randomized controlled mindfulness retreat

The authors frame the study around a longstanding debate in psychedelic science: acute subjective effects are widely considered important, and they correlate with later therapeutic benefits, but it remains uncertain whether these experiences are causally instrumental to change or simply accompany other drug effects. They note that standard questionnaires such as the Mystical Experience Questionnaire and 5-Dimensions Altered States of Consciousness scale have been criticised for being too fixed, conceptually limited, and unable to capture the full richness of first-person meaning-making. They also point to emerging evidence that meditation can produce altered states resembling psychedelic experiences, and that combining meditation with psychedelics may be informative, particularly because previous work has largely relied on psychometric tools or large public report databases rather than structured experimental narratives. The study set out to test whether natural language processing, specifically BERTopic topic modelling, could be used alongside manual qualitative analysis to examine phenomenological interview data from a double-blind, placebo-controlled randomised controlled trial conducted during a 3-day mindfulness retreat. The researchers aimed to compare the experiential landscapes of DMT/harmine and placebo-enhanced meditation, identify shared and divergent themes, and determine whether unsupervised NLP could detect latent phenomenological patterns that might be missed by conventional theory-driven approaches. The paper is presented as an ecologically valid analysis of structured interview data from experienced meditators in a controlled retreat setting, with relevance for understanding psychedelic phenomenology and the role of extra-pharmacological factors such as set and setting.

The study analysed qualitative interview data from a double-blind, placebo-controlled, between-subject randomised controlled trial conducted in 2023. Forty healthy participants were randomly assigned to DMT-harmine or placebo, with the groups matched for gender and reported to be similar at baseline. They attended one of two structurally identical 3-day meditation retreats at the Zen meditation centre Stiftung Felsentor in the Swiss Alps. The retreat consisted of preparation on day 1, drug or placebo administration on day 2, and integration on day 3. Participants practised sitting meditation, walking meditation, mindful physical work, short breaks and meals throughout the retreat. On the second day, participants received incremental sublingual doses of DMT-harmine or matched placebo in double-blind fashion while continuing meditation practice. Each active tablet contained 30 mg of DMT and 30 mg of harmine, and four tablets were given at 30-minute intervals for a total dose of 120 mg of each compound. The placebo tablets were taste-matched with sucralose, menthol and peppermint. Acute subjective experience was also assessed repeatedly with psychometric scales at eight time points during that day, although the present paper focuses on the interview data rather than those scale outcomes. The qualitative dataset came from phenomenological interviews inspired by the Micro-Phenomenological Interview method, which is designed to elicit detailed chronological descriptions of a specific lived experience while minimising retrospective distortion. Interviews were conducted after the second retreat phase. Because of staff and time constraints, 28 participants were interviewed in total, drawn randomly from the pool: 13 from the verum condition and 9 from placebo are reported here for the NLP analyses, while all 24 transcripts appear to have been used in the manual analysis. Some interviews were conducted in person shortly after the experience, and others online up to 11 days later. All interviews were audio recorded and the study was double-blind at the time of data collection. For the automated analysis, transcripts were cleaned and segmented in Python using spaCy and sentence-based segmentation. The researchers used BERTopic, an unsupervised topic-modelling approach built around sentence embeddings, UMAP dimensionality reduction, HDBSCAN clustering, CountVectorizer tokenisation, class-based term frequency-inverse document frequency weighting, and maximal marginal relevance to refine topics. Nouns and adjectives were emphasised, stop words were removed, and very incoherent or non-informative clusters were merged into a single noise topic. The model was applied first to the full corpus and then separately to the verum and placebo transcripts. For the manual component, two researchers independently coded the transcripts using a hybrid approach based on predefined codes informed by established psychometric scales and the team’s assumptions, while also allowing inductive new codes to emerge. Sentence-level coding was used, and intercoder disagreement was iteratively resolved with mediation until overall agreement was about 90%. A frequency-based thematic analysis then focused on seven main codes and related subcodes, including psychometrics, control, sensations, meaningfulness, well-being/life satisfaction, side effects and valence, together with contextual factors such as placebo versus substance. The manual and automated findings were compared and discussed within the research team.

Across the full corpus, the BERTopic analysis yielded 28 distinct topics plus one noise topic. The researchers reported a complex semantic landscape with both shared and divergent themes across the two conditions. About 82.1% of topics were centred on subjective effects. The intertopic distance map suggested both overlap and separation: some topics formed dense clusters that were semantically close, while four more distant topic islands indicated condition-specific or more distinct themes. The authors describe this as a large thematic diversity score and interpret it as evidence of meaningful experiential differentiation. When the verum transcripts were analysed alone, BERTopic produced 24 distinct topics plus one noise topic, with 20 topics shared with the across-group model. The placebo-only analysis produced 11 distinct topics plus one noise topic. The paper states that the verum group showed greater thematic overlap with the across-group analysis, and that both analyses displayed two main hubs around subjective effects. The placebo analysis also surfaced a placebo-only topic labelled “Placebo effect”. In the manual coding analysis, all 24 transcripts were included. The most frequent top-level codes were Control, Mental Awareness and Psychometrics, with global counts of 346, 312 and 164 respectively. Four new top-level codes emerged inductively, leading to a final code system of 39 codes and subcodes. The manual analysis produced nearly three times as many coded sentences in the verum group as in placebo: 886 versus 320. This difference was especially marked for codes reflecting subjective effects, including the psychometrics subcodes and sensations. Mental Awareness appeared in 96% of interviews. The authors report that awareness of mind and body was salient in both groups, but the verum condition contributed more of the codes for visual, auditory and cognitive aspects. Emotional codes were present in 85% of interviews, with 76.2% of all emotional codes in the verum group. For cognition, the verum-to-placebo count ratio was particularly pronounced at 189 to 39. Bodily Awareness appeared in 92% of interviews and was also almost twice as frequent in the verum group. Sensory alterations were more common under DMT-harmine. Across all interviews, 71% contained auditory codes and 54% contained visual codes. Auditory experiences were more than twice as frequent in the verum group, ranging from heightened or “pure” hearing to synaesthesia and auditory pareidolia. Visual experiences were likewise much more common in the verum condition, with only 3 of 38 visual codes appearing in placebo. These ranged from increased light sensitivity to visual synaesthesia, pseudo-hallucinatory phenomena and complex imagery with eyes open or closed. The text also reports that positive and negative emotional valence were coded, but the extracted passage breaks off before giving the full numerical result for this section. Later in the Results, the authors state that 82.5% of participants guessed their group correctly after the retreat day-2 experience. They also note that verum participants tended to report more intense, diverse and acute subjective effects, whereas placebo participants reported more body-centric, familiar and context-shaped experiences. The placebo narratives contained more references to coping, containment and retreat-context factors, while verum narratives converged more strongly on acute psychedelic-like effects. The authors characterise this contrast as a focusing effect under DMT-harmine, with vivid and sometimes overwhelming experiences concentrating attention and language around acute effects.

The authors argue that the study shows the value of combining unsupervised NLP with manual qualitative analysis to study psychedelic subjectivity in a small, structured trial dataset. They say the automated topic models captured both established acute psychedelic effects and context-specific themes tied to the retreat setting, and that the findings demonstrate the sensitivity of NLP methods to experiential details that may fall outside traditional questionnaire-based frameworks. They also emphasise that the mixed-method approach allowed triangulation: the manual coding and BERTopic outputs converged on a strong group effect, supporting the view that the observed differences were not merely investigator artefacts. The authors place their findings in the context of earlier work suggesting that language can reveal aspects of psychedelic experience that psychometric scales may miss, and that meditative and psychedelic phenomenology can overlap. They interpret the strong use of Buddhist and spiritual language across both conditions as evidence that participants drew on a contemplative framework to make sense of their experiences, which they suggest may have been shaped by the participants’ existing meditation backgrounds and the retreat context. They also note that such language may serve as a bridge for articulating ineffable states, and may be relevant for therapeutic integration. At the same time, they stress experiential divergence between conditions. In their interpretation, DMT-harmine produced a more intense, diverse and acute subjective effect profile, whereas placebo-enhanced meditation was less heterogeneous and more closely tied to familiar meditative and bodily experiences, often framed through coping and containment rather than surrendering to novelty. They suggest that the placebo condition still contained meaningful contextual and experiential effects, and that environmental influences may have been active in both groups even if they were expressed differently. The authors acknowledge several limitations. They note that some topic clusters were less coherent and that unsupervised topic modelling can yield less interpretable outputs when datasets are small. They caution that effectiveness depends on sample quality and size, and recommend care when applying BERTopic to limited corpora. They also point out that the manual coding process was potentially susceptible to investigator bias, especially because coders were aware of experimental assignments and because verum interviews were longer and more vivid. Finally, they state that the retreat setting, while increasing ecological validity, makes it difficult to separate pharmacological effects from contextual influences, so the findings should be treated as exploratory and specific to this experimental context rather than broadly generalisable. The authors suggest that future work should continue to use data-driven, multi-method approaches and should pay close attention to extra-pharmacological factors, including set and setting, when studying psychedelic outcomes.

The conclusion states that this study is an early step towards more qualitative, less biased, and more data-driven research on psychedelic subjective effects, and argues that transformer-based NLP methods such as BERTopic can complement traditional qualitative approaches by detecting experiential features that theory-driven methods may miss while also reducing time and labour demands. It also emphasises that extra-pharmacological factors appear to strongly shape subjective effects, supports the idea that experiences induced by psychedelics or mindfulness-based practices may be central to therapeutic and wellbeing benefits, and calls for stronger phenomenological framing and improved psychometric approaches in future research.