Technical Summary: REST-ASMR Dataset

Problem Statement

Research into Autonomous Sensory Meridian Response (ASMR)—a somatosensory phenomenon characterized by pleasant tingling and cardiovascular slowing—has been significantly hindered by a lack of standardized, open-access, multimodal datasets. Existing studies often rely on private datasets or self-reported data, which limits the ability to model the complex, non-linear relationships between audiovisual triggers and physiological responses. Furthermore, there is a scarcity of resources that simultaneously capture high-resolution physiological data, continuous subjective annotations, and synchronized stimuli to distinguish ASMR-induced relaxation from general relaxation (e.g., nature stimuli).

Methodology

Dataset Composition and Participants

The REST-ASMR (Response to Environmental & Sensory Triggers) dataset was collected from 34 healthy university students (18 men, 16 women; mean age 25.2 ± 4.6 years). The study utilized a counterbalanced block design where participants viewed eight 60-second video clips stratified into two categories:

1. ASMR Triggers (N=4): Non-vocal, binaural clips featuring brushing and tapping sounds, with visual content cropped to exclude facial features to isolate acoustic-tactile mechanisms.
2. Environmental Controls (N=4): Nature sequences (brook, ocean, bonfire, wind) devoid of human presence, serving as a non-social relaxation baseline.

Data Acquisition

The experiment lasted approximately 12 minutes per participant, structured with baseline rest, two 4-minute stimulus blocks (separated by washout rest), and final rest.

• Physiological Data: High-resolution Photoplethysmography (PPG) was recorded at 2000 Hz using a Biopac SS4LA transducer on the left middle finger to minimize motion artifacts from the dominant right hand.
• Behavioral Data: Participants provided continuous subjective ratings in real-time using a discrete 4-point Likert scale ("none" to "very") via keyboard. They rated "Tingling Intensity" for ASMR videos and "Pleasantness" for Nature videos at a 1000 Hz sampling rate.
• Stimuli: Delivered via Presentation software with precise timestamp logging, 60 Hz visual refresh, and dichotic audio delivery.

Data Processing and Alignment

A multi-stage preprocessing pipeline synchronized the modalities into a dense 10 Hz time-series:

1. Latency Correction: Adjusted for system initialization delays, clamping the first trial's valid duration to 55.7 seconds to prevent data leakage.
2. State-Aware Buffering: The first 5 seconds of each clip served as a warm-up buffer; the final keypress in this window established the initial state.
3. Dense Upsampling: Sparse keypress events were upsampled to 10 Hz using zero-order hold logic.
4. Physiological Processing: Raw PPG signals were bandpass filtered (0.5–35 Hz), sliced based on latency-corrected timestamps, downsampled to 10 Hz via Fourier resampling, and Z-score normalized per clip to mitigate inter-participant variability.

Machine Learning Validation

To validate the dataset's utility, a baseline predictive model was trained to perform frame-by-frame binary classification (Tingle vs. No Tingle).

• Feature Extraction: Visual features were extracted using ResNet-18 (512-dim), acoustic features using MFCCs and spectral properties (31-dim), and physiological features using normalized PPG (1-dim). These were concatenated into a 544-dimensional vector.
• Architecture: A two-layer Bidirectional Long Short-Term Memory (BiLSTM) network with a hidden dimension of 128.
• Validation Protocol: A strict, double-independent 4-fold cross-validation was employed. In each fold, 25% of participants and a specific pair of stimuli (one ASMR, one Nature) were withheld from training and validation to ensure no data leakage and to test generalization to unseen subjects and stimuli.
• Ablation Studies: Unimodal models (Video-only, Audio-only) and a non-temporal XGBoost baseline were compared against the full multimodal fusion model.

Key Results

Stimulus Efficacy and Behavioral Agreement

• Responder Rate: 97.06% (33/34) of participants successfully reported tingling sensations during at least one ASMR trial.
• Inter-Subject Agreement: Continuous behavioral traces showed statistically significant temporal alignment across participants for ASMR stimuli (p < 0.05), with strong correlations (e.g., Video 3: r = 0.539, p < 0.001).
• Physiological Correlation: A significant divergence was found between conditions. ASMR trials showed a negative correlation between subjective tingle intensity and pulse rate (r = -0.0921), indicating parasympathetic deceleration. In contrast, Nature trials showed a slight positive correlation (r = 0.0556).

Predictive Performance

• Frame-Level Classification: The full multimodal BiLSTM model achieved a Global Accuracy of 75.51% and a Macro F1-score of 71.86% under double-independent cross-validation.
• Specificity: The model achieved 100% Nature Specificity, correctly identifying zero false positives in control trials, confirming it learned ASMR-specific patterns rather than general positive affect.
• Video-Level Classification: Using a peak-detection strategy (requiring >15% of frames to be predicted as "tingle"), the model achieved 100% accuracy in distinguishing ASMR trials from Nature trials across all held-out test sequences.
• Statistical Significance: Permutation tests confirmed the results were not due to chance (p < 0.0002), with a large effect size (Cohen's d = 5.65).
• Ablation Findings: The full multimodal model significantly outperformed unimodal baselines (Video-only: F1=68.12%; Audio-only: F1=69.94%), validating the necessity of integrating physiological data with audiovisual cues. The temporal model (BiLSTM) also outperformed the non-temporal XGBoost baseline (F1=62.72%), highlighting the importance of temporal dynamics in ASMR modeling.

Significance and Claims

The paper claims that REST-ASMR constitutes the first open-access resource to combine synchronized high-resolution PPG, continuous subjective ratings, and audiovisual stimuli in a comparative design including both ASMR and Nature controls.

Its primary significance lies in providing a dense temporal foundation for:

1. Affective Computing and Multimodal Research: Enabling the development of models that can distinguish ASMR-specific neurophysiological signatures from general relaxation.
2. Personalized Models: Supporting the creation of personalized audio-medicine and digital therapeutics for stress and insomnia by identifying predictive physiological biomarkers.
3. Generative AI: Offering a framework for physiology-guided content generation and adaptive systems that optimize heart rate reduction through real-time feedback.

The authors emphasize that the dataset's rigorous technical validation, including strict leakage-free cross-validation and demonstrated biological plausibility (cardiovascular deceleration), makes it a robust benchmark for advancing interdisciplinary studies in sensory neuroscience and human-computer interaction.