Exploring Label Noise Reduction Techniques for Sleep Stage Classification Using Wearable Devices

Wearable devices offer a portable alternative to polysomnography for sleep stage classification using accelerometer and photoplethysmography (PPG) data. However, the performance of machine learning models heavily depends on the quality of the reference data. Consequently, the presence of incorrectly labeled data undermines the performance of these models. In this work, we conduct an exploratory analysis of different label noise reduction methods, including duration window and Isolation Forest. We evaluate the impact of these techniques on sleep stage prediction using several machine learning classifiers. Our results provide insights into the effectiveness and characteristics of label noise reduction methods for improving sleep stage classification.