0955 Evaluating state-of-the-art algorithms for sleep-wake classification using wrist-worn wearable devices

Abstract Introduction Over the last 40 years, a variety of algorithms have been proposed to classify sleep-wake from wrist acceleration data. Input features into these activity counts or raw acceleration. The range single heuristic rule logistic regression machine learning and deep learning. purpose this work is evaluate compare accuracy against polysomnography (PSG) annotations on common dataset sleep laboratory. Methods Newcastle PSG was used various prediction concurrent in thirty second epochs. This contains 28 patients, 27 which had data for both right left one only. Twenty participants at least disorder. Sleep disorders included idiopathic hypersomnia, restless leg syndrome, apnea, narcolepsy, paralysis, nocturia, obstructive RBD, parasomnia insomnia. Results domain adversarial convolutional neural network (DACNN) model showed best overall results (sens=83.9, spec=57.6, f1=81.7, WASO-RMSE=80.9). next performing according WASO-RMSE Cole-Kripke (sens=81.4, spec=50.3, f1=78.0, WASO-RMSE=90.4). followed by van Hees (sens=83.6, spec=47.5, f1=79.1, WASO-RMSE=91.1). fourth Random Forest (spec=77.5, sens=55.5, f1=76.4, WASO-RMSE=93.0). fifth Sadeh (sens=82.6, spec=49.7, f1=78.5, WASO-RMSE=93.6). sixth LSTM (sens=78.6, spec=58.9, f1=77.8, WASO-RMSE=93.9). worst CNN (sens=79.8, spec=54.2, WASO-RMSE=99.8). * sens=sensitivity, spec=specificity, WASO=wake after onset, RMSE=root mean squared error Conclusion DACNN algorithm outperformed all other classification Despite being vastly different input features’ types complexity, performed similarly dataset, with f1 scores ranging 76.4 79.1 90.4 99.8. Support (if any)