Machine learning model for reproducing subjective sensations and alleviating sound-induced stress in individuals with developmental disorders

Itsuki Ichikawa, Yukie Nagai, Yasuo Kuniyoshi, Makoto WadaFrontiers in Psychiatry 2025 · DOI

TL;DR

Researchers have developed a deep learning framework that predicts how individuals with developmental disorders (like ASD and ADHD) perceive distressing sounds and determines exactly how to filter those sounds to reduce stress. By combining audio features with personal sensory profiles, the models achieved correlation coefficients of 0.4 to 0.7 in estimating subjective stress, paving the way for "smart" noise-canceling systems that adapt to an individual's unique nervous system.

What Problem Does This Paper Address?

For many individuals with autism spectrum disorder (ASD) or attention deficit hyperactivity disorder (ADHD), the world is often too loud, too sharp, or too chaotic. This "auditory hypersensitivity" isn't just a nuisance; it causes genuine physical distress and social withdrawal.

While noise-canceling headphones exist, they are "blunt instruments"—they often block out everything, including the voices people want to hear. Furthermore, sensory needs are highly individualized; what is soothing to one person might be aggravating to another. The challenge is creating a system that understands a person's subjective "internal" experience of sound and can autonomously modify the "external" environment to match their comfort level.

Methods

The researchers developed two distinct Deep Neural Network (DNN) models based on two experimental tasks:

  1. The Recollection Task: Participants listened to sounds and adjusted filters to reproduce how they remember perceiving that sound in the past. This quantified their specific hypersensitivity.
  2. The Easing Task: Participants adjusted filters to make the sound as least stressful as possible for long-term listening.

The Architecture: The models utilized a Bi-LSTM (Bidirectional Long Short-Term Memory) network to process audio features extracted via YAMNet (a pre-trained model based on MobileNet). Crucially, the researchers didn't just look at the audio; they fed the model "participant condition" data, including:

  • Sensory Profiles: Scores from the Autism Spectrum Quotient (AQ) and Adolescent/Adult Sensory Profile (AASP).
  • State Variables: Current fatigue levels and sleep duration.
  • Environmental Data: Bluetooth profile types and volume settings.

The model then predicted the optimal filter parameters (e.g., gain, frequency-specific amplification, pitch shifts, and "change suppression" for sudden noises) and the resulting subjective stress ratings.

Datasets

The study used a custom dataset of 62 audio-visual stimulus pairs.

  • Audio: 45 files from Pixabay and 17 original recordings (e.g., crosswalk signals, steady noise).
  • Visual: Each sound was paired with a corresponding photo to provide context.
  • Pre-training: The audio feature extractor (YAMNet) was trained on the AudioSet dataset, a large-scale ontology of sound events.

The data was split by stimulus rather than by user to ensure the model could generalize to new, unfamiliar sounds.

Key Insights and Ideas

  • Stress is Predictable: The study proves that subjective stress isn't random; it can be modeled as a mathematical function of audio characteristics and an individual's psychological/sensory traits.
  • The "Easing" Effect: Participants were able to significantly lower their stress ratings through manual filtering, providing a "ground truth" that an automated system can aim for.
  • Personalization is Non-Negotiable: The models performed differently when trained on the entire group versus just the Developmental Disorder (DD) group. This suggests that a "one-size-fits-all" AI won't work; we likely need "branched" architectures where a base model is fine-tuned for specific neurodivergent subgroups.
  • Beyond Volume: The inclusion of "Change Suppression" (which dampens sudden loud noises) and "Water Effects" (low-pass filtering) showed that alleviating stress is about texture and timing, not just turning down the volume.

Key Results

  • Stress Reduction: Subjective stress ratings were significantly lower in the Easing task compared to the Recollection task across all five evaluation categories (Overall, Painful, Distracting, Anxious, and Impeditive).
  • Model Accuracy: The DNN successfully estimated stress ratings with a correlation coefficient (Spearman) between 0.4 and 0.7.
  • Filter Prediction: The model was particularly good at predicting settings for "Amplify (All)," "Band Reject," and "Flanger" filters.
  • Error Rates: The Mean Absolute Error (MAE) remained relatively low, suggesting the model's predictions were close to the participants' actual preferred settings.

Limitations and Future Work

The authors identify several hurdles before this hits your smartphone:

  • Sample Size: With only 57 participants, the model's ability to cover the vast spectrum of neurodiversity is limited.
  • Task Complexity: Participants had to manage over 10 filters simultaneously, which is cognitively demanding and may introduce noise into the data.
  • Real-world Latency: The current study was conducted in a controlled environment. A real-time wearable would need to process environmental audio with near-zero latency.
  • Longitudinal Data: Sensory sensitivity fluctuates. Future models need to account for how a user's needs change over weeks or months, not just a single day.

The ultimate goal? Integrating this logic into "smart earplugs" or AR glasses that automatically "re-render" the acoustic world in real-time to keep the wearer's stress levels in a healthy zone.

Summary updated

Abstract

These findings suggest that the prediction model for the Easing task can potentially be developed into a system that automatically reduces sound-induced stress through auditory filtering. Similarly, the model for the Recollection task could be used as a tool for assessing auditory stress. To establish a robust support system, further data collection, particularly from individuals with DD, is necessary.
BibTeX
@inproceedings{ichikawa2025machine,
  author = {Itsuki Ichikawa and Yukie Nagai and Yasuo Kuniyoshi and Makoto Wada},
  title = {{Machine learning model for reproducing subjective sensations and alleviating sound-induced stress in individuals with developmental disorders}},
  booktitle = {Frontiers in Psychiatry},
  year = {2025},
  doi = {10.3389/fpsyt.2025.1412019},
}
Figure