Automated Detection and Classification of Sleep Spindles using Machine Learning and Signal Processing Techniques

¹ Department of Mechatronics & Biomedical Engineering, Air University, Islamabad, Pakistan
² IRC-IMR, King Fahd University of Petroleum and Minerals (KFUPM), Dhahran 31261, Saudi Arabia
³ School of Information Engineering, Jiangxi University of Science and Technology, Ganzhou, Jiangxi 341000, China.

^a sundhufaiza@gmail.com
^b hafizzia@au.edu.pk
^c) Corresponding author: zeashan.khan@kfupm.edu.sa
^d ajm@jxust.edu.cn

Abstract

Sleep spindles are biomarkers of sleep quality and cognitive process, especially memory consolidation. These spindles, which are related to thalamocortical circuits, manifest as 11-15 Hz sinusoidal waves with a duration range of 0.5 to 3 seconds in EEG recordings. There is evidence that schizophrenia and thalamic stroke have reduced sleep spindles. This study aims to classify and automate the detection of sleep spindles from raw physiological signals using computational methods and AI. While global efforts are underway to accurately detect sleep spindles using AI, most involve classical signal processing techniques like wavelet transforms and STFT. Few studies have comprehensively compared different AI methods for spindle classification. Further insight into how the precision, recall, and F1 scores of spindle parameters can be used to advance cognitive performance is also required. In this research, Google Colab AI is innovatively utilized to devise a multi-stage approach to sleep spindle extraction from EMG recordings, offering a non-invasive, more precise way of quantifying muscle activity and cerebral responses while sleeping. Bandpass filtering, STFT, and wavelet transforms will provide the main parameters: spindle frequency (12-16 Hz), amplitude, duration, and spectral density. Detailed consideration of AI approaches will be conducted for automatic spindle recognition. To contrast the accuracy and generalizability of classification models like SVM, KNN, RF, and CNN, these models are cross-tested across datasets. Cross-validation and precision, recall, and F1 scores helped identify the best classification algorithm. This research can be integrated with the wearable technology for real-time monitoring and intervention to the next level, providing important insights into sleep spindle classification and analysis.