Noise-Robust Sleep States Classification Model Using Sound Feature Extraction and Conversion

• Main Authors: Ko, Sangkeun, Min, Seongho, Choi, Ye Shin, Kim, Woo-Je, Lee, Suan
• Format: Conference Proceeding
• Language: English
• Subjects: Analytical models; CNN; Convolutional neural networks; Deep Learning; Feature extraction; Image Representation; LSTM; Noise; Noise robustness; Robustness; Sleep apnea; Sound Feature Extraction
• ISSN: 2375-9356
• DOI: 10.1109/BigComp60711.2024.00051

This study proposes an effective state classification model for sleep, which is crucial for improving daily functioning and overall quality of life. We delve into the extraction of auditory features from sleep-related sounds, such as snoring and teeth grinding, and apply five distinct image transformations-Recurrence Plots (RP), Markov Transition Fields (MTF), Gramian Angular Summation Fields (GASF), Gramian Angular Difference Fields (GAD F), and Short-Time Fourier Transform (STFT)-to accurately delineate sleep states. Our research introduces an innovative deep learning model adept at classifying these states based on the images obtained from these transformations. Furthermore, we rigorously test the model's resilience to noise by introducing varying levels (0%, 25%, 50%, and 75%) and observe that the Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model, particularly when combined with the STFT technique, consistently outperforms under all noise conditions, achieving accuracies between 99.55% and 98.98%. The findings of this research significantly contribute to the fields of sleep analysis and the study of sleep disorders, offering a robust framework for understanding and classifying sleep states.