Authors:
Rodrigo Braga
1
;
Daniel Osório
1
;
2
and
Hugo Gamboa
1
;
2
Affiliations:
1
Department of Physics, Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa, Monte da Caparica, 2892-516, Caparica, Portugal
;
2
Plux-Wireless Biosignals S.A, Avenida 5 de Outubro 70, 1050-59, Lisboa, Portugal
Keyword(s):
Deep Learning, Machine Learning, Wearable, Photoplethysmography, Sleep Stages, Heart Rate Variation.
Abstract:
Sleep’s impact on mood and health is widely recognized by medical researchers with such understanding disseminating among average people in recent years. The main objective of this work was the development of machine learning algorithms for automatic sleep cycles detection. The features were selected based on the AASM manual, which is considered the gold standard for human technicians. For training the models we used MESA, a database containing 2056 full overnight unattended polysomnographies. With the goal of developing an algorithm that would only require a photoplethysmography (PPG) device to be able to accurately predict sleep stages and quality, the main channels used from this dataset were peripheral oxygen saturation and PPG. Testing the performance of Random forest, Gradient Boosting, Gaussian Naive-bayes, K-Nearest Neighbours, Support Vector Machine and Multilayer Perceptron classifiers, and using features extracted from the dataset, we achieved 80.50 % accuracy, 0.7586 Cohe
n’s kappa, and 77.38% F1-score, for five sleep stages, using a Multilayer Perceptron. To assess its performance in a real-world scenario we acquired sleep data and compared the classifications attributed by a popular sleep stage classification android app and our algorithm, resulting in a strong level of agreement (90.96% agreement, 0.8663 Cohen’s kappa), for four sleep stages.
(More)