Implementing deep learning on edge devices for snoring detection and reduction.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Signal Processing

Background:

• Snoring affects sleep quality and is linked to various health issues.
• Existing snoring detection methods often require high computational resources, limiting their use on edge devices.
• Real-time snoring mitigation solutions are needed for accessible sleep health management.

Purpose of the Study:

• To introduce MinSnore, a novel deep learning model for real-time snoring detection and reduction.
• To design a computationally efficient model deployable on low-configuration edge devices.
• To evaluate the performance and real-world effectiveness of the MinSnore system.

Main Methods:

• Developed MinSnore by integrating MobileViTV3 blocks into a Dynamic MobileNetV3 backbone, combining CNNs and transformers.
• Pre-trained the model on 46,349 audio files using Self-Supervised Learning with Barlow Twins (SSL-BT).
• Fine-tuned the model on 17,355 segmented audio clips and evaluated its performance on a Raspberry Pi.

Main Results:

• Achieved high accuracy (96.37%), precision (96.31%), recall (94.12%), and F1-score (95.02%) in snoring detection.
• Demonstrated a reduction in snoring duration during real-world experiments when deployed on a Raspberry Pi.
• The model exhibited minimal computational overhead, suitable for edge device deployment.

Conclusions:

• MinSnore offers a highly accurate and efficient solution for real-time snoring detection and reduction.
• The model's low computational requirements make it ideal for deployment on low-cost edge devices.
• This work presents a significant advancement in addressing sleep-related health issues through accessible technology.