Machine learning classifier solving the problem of sleep stage imbalance between overnight sleep.

Abstract

Feature extraction follows the American Academy of Sleep Medicine (AASM) sleep score manually and applies it to machine learning with a focus on the generalization of sleep data to enable data-centric artificial intelligence. In real-world clinical testing, the manual scoring of sleep stages is time-consuming and requires significant expertise. Additionally, it is subject to interobserver subjective bias. Machine-learning techniques offer a way to overcome these limitations through automation. However, machine learning for sleep phase prediction can perform poorly for small classes. If the distribution of the training data was unbalanced, the model was trained with a bias toward the majority class. To address this, we experimented with loss function adjustment and resampling methods that assign more weight to the prediction errors of minority classes in sleep scoring to determine how to overcome the data imbalance problem. Machine learning can also be used to compare the accuracy of each channel in identifying electrodes, which should be monitored more closely in real-world clinical testing. Owing to the small amount of data available for machine learning in this study, we used various machine learning classifiers by increasing or decreasing the dataset using sampling techniques and weighting different classes of sleep stages. In our experiments, the best-performing model for classifying sleep stages had an accuracy of 91.9%, kappa of 0.899, and F1-score of 86.9%.