TL;DR
The BEATS dataset is a massive, "in-the-wild" compilation of psycho-physiological data from 30 participants over 21 days of daily life. By pairing raw wearable sensor data (heart rate, skin conductance, etc.) with real-world music listening habits, it provides a unique window into how music regulates our physiology outside the artificial confines of a laboratory.
What Problem Does This Paper Address?
While we know music has therapeutic value for everything from stress regulation to Parkinson’s disease, the underlying biological "why" remains a bit of a black box. Most existing research suffers from two major flaws:
- The Lab Effect: Most studies happen in controlled, sterile environments using short 60-90 second song snippets. This doesn't reflect how people actually listen to music—while working, commuting, or relaxing at home.
- Data Fragmentation: Previous datasets often lack raw, multimodal physiological signals (like electrodermal activity or triaxial acceleration) or fail to account for individual differences like personality and musical background.
The authors address this by taking the research into the "wild," capturing the messy, complex relationship between music and the human body during real-world daily routines.
Methods
The researchers developed an open-source framework that integrates wearable sensing with music streaming.
- Hardware: Participants wore the Empatica EmbracePlus, a research-grade smartwatch that continuously recorded Photoplethysmography (PPG), Electrodermal Activity (EDA), skin temperature, and movement (ACC).
- Stimulus Selection: Instead of letting participants pick any song, the team used a "computationally grounded" approach to select Spotify playlists. They used text embeddings and a weighted relevance score—balancing popularity, acoustic criteria (e.g., instrumental vs. vocal), and functional keywords—to categorize music into three buckets: Affect Regulation (relaxation), Eudaimonic (introspection), and Goal Attainment (focus).
- Data Pipeline: A custom monitoring system used Spotify API calls to log exactly when a track started and ended. To ensure high-quality "ground truth," participants received short email surveys twice a day to confirm if the music actually helped them achieve their intended state (e.g., "Did this help you focus?").
Datasets
The primary contribution is the BEATS (Biomarkers, Experience and Affect from Therapeutic Soundscapes) dataset.
The authors provide a comparison with existing benchmarks (like DEAP or PMEmo), highlighting that BEATS is the first to offer multi-week, "in-the-wild" recordings with full listener control and raw multimodal signals.
Key Insights and Ideas
- Acoustic Fingerprints: Using logistic regression, the authors found that different "therapeutic" music functions have distinct acoustic profiles. For example, "Goal Attainment" music (for focus) has a higher density of note events, while "Eudaimonic" music (for meditation) is characterized by smoother, less abrupt progressions.
- The "Lingering" Effect: One of the most interesting findings was that physiological changes—specifically in skin conductance and temperature—remained closer to the "regulated" baseline even after the music stopped. This suggests that music's physiological impact isn't just a momentary reaction but a sustained regulatory effect.
- Context Matters: Physiological markers like Heart Rate Variability (HRV) and motion were significantly lower during music listening compared to the hour before or after. This validates that participants were successfully using the music to enter specific, stable states (like deep focus or relaxation) during their daily lives.
Key Results
- Scale: The dataset includes over 10,725 hours of raw physiological recordings.
- Alignment: 262 hours of that data directly overlap with active music listening sessions across 844 distinct sessions.
- Engagement: "Goal Attainment" was the most popular use case, accounting for 176.4 hours of listening.
- Validation: A logistic regression classifier achieved a mean F1 score of 0.785 in distinguishing between the three music functions based solely on acoustic features, proving the playlists were distinct and well-categorized.
- Survey Success: The study achieved a 90.8% response rate on post-listening surveys, with most users confirming the music supported its intended purpose.
Limitations and Future Work
The authors are transparent about the challenges of "in-the-wild" data. Because the environment wasn't controlled, factors like ambient temperature or physical activity (walking vs. sitting) can "pollute" physiological signals like EDA.
Future work will likely involve developing more sophisticated machine learning models to "de-noise" these real-world signals. Additionally, while the dataset is large, it only includes 30 participants; expanding this to a more diverse population could reveal how cultural background or specific health conditions (like anxiety) further modulate the music-physiology link.
Summary updated