Psychedelic-induced experiences: New study of machine learning models could help therapists guide patients

“Language models learn feeling and substance from 11,000 psychoactive experiences” is a research project that took nearly 12,000 drug reports with 52 different molecules and combined them with 3 robotic models to characterize all sense and feeling trajectories as they appear during psychoactive experiences.

The lack of a unified framework to measure changes in consciousness to accommodate each constrained treatment researchers Sam Freesun Friedman and Galen Ballentine to use two machine learning techniques based on deep transformers for natural language processing (called BERT) and superficial canonical correlation analysis (CCA) combined with Erowid’s publicly available psychoactive natural language testimonials, to trying to find answers.

The first model predicted 28 sentiment dimensions in each story, validated with annotations from clinical psychiatrists. A second model was programmed to predict biochemical information (pharmacological and chemical class, molecule name, receptor affinity) as well as demographic information (sex, age) from testimonies.

Finally, CCA linked the affinities of the 52 drugs for 61 receptor subtypes with words through testimonials, revealing 11 latent receptor experience factors, each mapped onto a 3D cortical atlas of receptor gene expression.

Together, the three machine learning methods explain a neurobiologically sensitive and temporally sensitive portrait of subjective drug-induced experiences.

As the researchers explained, these mutually confirming patterns point to an underlying structure of psychoactive experience dominated by the distinction between the lucid and the banal, but also sensitive to the effects of specific drugs.

For example, MDMA was uniquely linked to feelings of “love” mid-experience. Powerful psychedelics like DMT and 5-MeO-DMT were associated with “mystical experiences”.

Other tryptamines were associated with an emotional constellation of “Surprise”, “Curiosity”, and “Realization”.

The authors hope that applying these models to real-time biofeedback (like EEG) with zero-hit learning that tunes the sentimental trajectory of experience through changes in audiovisual outputs will allow practitioners to guide the course. therapeutic sessions, maximizing benefit and minimizing harm to patients.

Find the full research here.

Sherry J. Basler