Sleep-Wake Classification using Recurrence Plots from Smartwatch Accelerometer Data

Multiple wearable devices are equipped with sensors that capture motion-based sleep information. Using accelerometer sensor data from smartwatches, the literature explores the sleep/wake classification performance through different data representations, such as raw data (sensor time series) and feature extraction. Nevertheless, the representation of time series through Recurrence Plots can produce informative and noise-robust characteristics. In this sense, we propose a method based on Recurrence Plots from smartwatch accelerometer data and leverage the RensNet50 and EfficientNet neural networks to classify sleep/wake stages. Our best result reaches 79.3% balanced accuracy, a gain of almost three percentage points compared to feature extraction of the baseline work. We also explore feature extraction techniques to compare different representations with the Random Forest and Logistic Regression classifiers, achieving up to 85.0% balanced accuracy, surpassing the baseline work, and showing that these techniques also have the potential to improve the Recurrence Plots models.