Music for Psilocybin Therapy and the Temporal Arc of Psychedelic Experience

Technical Summary: Music for Psilocybin Therapy and the Temporal Arc of Psychedelic Experience

Problem Statement
Psilocybin therapy (PT) relies heavily on music to structure the 4–6-hour therapeutic session. The prevailing theoretical framework, derived from historical guidelines (e.g., the Bonny Method, Johns Hopkins protocols), posits that music should be acoustically and affectively aligned with the pharmacodynamic "onset–peak–return" arc of the psychedelic experience. This model suggests that onset music should be stable and predictable, peak music emotionally expansive, and return music grounding and familiar. However, despite this widespread theoretical consensus, there is no empirical evidence confirming that contemporary PT playlists actually exhibit consistent, measurable acoustic trajectories corresponding to these phases. Current playlist curation relies largely on clinical intuition rather than objective acoustic criteria, leaving a gap between theoretical design and actual implementation.

Methodology
The authors conducted a large-scale computational analysis of 46.9 hours of audio comprising 369 unique tracks from eight widely used PT playlists (drawn from Imperial College London, Johns Hopkins University, Chacruna Institute, Copenhagen University, and COMPASS Pathways). The study employed a multi-layered machine learning approach to test for phase differentiation:

1. Affective Analysis: The authors estimated perceived emotional arousal and valence using Spotify's API and the Music2Emo model across the full duration of the playlists.
2. High-Level Descriptor Analysis: Tracks were annotated with nine high-level musical descriptors (tempo, energy, danceability, acousticness, instrumentalness, loudness, valence, liveness, speechiness) via the Spotify API. Logistic regression models were trained to classify tracks into onset, peak, or return phases based on these features.
3. Low-Level Audio Feature Analysis: To capture subtler acoustic properties, 65 low-level features from the ComParE feature set (e.g., spectral centroid, MFCCs, RASTA bands) were extracted at both 30-second segment and whole-track levels. Both logistic regression and Random Forest classifiers were trained to predict phase labels.
4. Unsupervised Clustering: K-means clustering was applied to the low-level feature space without phase labels to determine if intrinsic acoustic groupings naturally correspond to the onset–peak–return structure.
5. Granular Analysis: The same supervised and unsupervised methods were applied individually to each of the eight playlists to assess whether phase structure exists within specific curation sets but fails to generalize across the dataset.
6. Playlist Differentiation: Random Forest models were trained to identify the source playlist of a track to determine if distinct curatorial styles exist between the different protocols.

Key Results
The analysis yielded consistent evidence that current PT playlists lack robust acoustic organization aligned with the theoretical onset–peak–return model:

• Emotional Content: Estimated arousal and valence showed no systematic phase-based organization across the majority of playlists.
• High-Level Descriptors: Logistic regression models using high-level descriptors failed to distinguish phases significantly better than a majority-class baseline (average macro F1 score of 0.44 vs. chance of 0.33).
• Low-Level Features (Supervised): Random Forest models trained on low-level ComParE features achieved modest, statistically significant above-chance performance at the track level (average macro F1 = 0.58; p = 0.0018). However, misclassification rates remained high (e.g., 62% of onset tracks misclassified), and feature importance analysis (SHAP values) indicated that phase differentiation relied on distributed, subtle acoustic signals (e.g., specific MFCCs and RASTA bands) rather than dominant markers.
• Unsupervised Structure: K-means clustering revealed a near-total absence of intrinsic phase structure. Silhouette scores were very low (< 0.25), and agreement with true phase labels (ARI, NMI) was near zero, suggesting that listeners cannot naturally separate tracks into phase-based acoustic categories without explicit prior labeling.
• Within-Playlist Consistency: When analyzed individually, only 3 of the 8 playlists showed phase classification performance significantly above baseline, and even then, accuracy was poor. Unsupervised clustering within individual playlists also failed to reveal coherent phase groupings.
• Playlist Differentiation: While playlists could be distinguished from one another slightly better than chance, classification accuracy remained low (mean misclassification of 79% for most playlists), indicating that different protocols draw from a shared pool of acoustically similar music.

Significance and Claims
The paper claims to provide the first large-scale computational characterization of music playlists used in psychedelic-assisted therapy. The primary significance of the work is the demonstration that the dominant theoretical principle of PT music programming—acoustic alignment with the onset–peak–return arc—is not consistently realized in contemporary practice.

The authors argue that while music remains a critical component of the therapeutic experience, the mechanism of its efficacy may not depend on rigid, generalized acoustic trajectories. Instead, the findings suggest that therapeutic music may function more effectively through subjective emotional resonance, personalization, and context-dependent factors rather than adherence to a standardized acoustic template. The study concludes that current curatorial practices are not aligned with their theoretical basis, raising the need for future frameworks that move beyond coarse temporal models toward more flexible, personalized, or fine-grained psychological targets. The work does not propose new playlist designs but rather challenges the assumption that existing playlists are acoustically optimized according to established theories.