Classificação de estágios de sono utilizando técnicas de aprendizado de máquina automatizado

Abstract

Sleep analysis using wearable sensors has become a promising alternative for non-invasive monitoring, enabling sleep stage estimation from physiological data such as heart rate and movement. This study presents a machine learning model capable of classifying different sleep stages. The preprocessing steps included outlier removal using the IQR method, data normalization with MinMaxScaler, and feature extraction with the Catch22 feature set. The model was trained using leave-one-out cross-validation, ensuring a robust and independent evaluation for each participant. The AutoGluon framework was employed to test multiple classification algorithms, including Random Forest, Extra Trees, LightGBM, CatBoost, and neural networks. The results were compared with findings from the literature, demonstrating that combining of AutoML models using features extracted from heart rate and movement can achieve competitive accuracy in sleep stage classification.