1122 A Deep Learning Model for Inferring Sleep Stage from a Flexible Wireless Dual Sensor Wearable System Without EEG

Abstract Introduction In-lab polysomnography (PSG) is costly and difficult to scale due to a need for specialized personnel for data acquisition and annotation. Numerous novel wearable devices without electroencephalography (EEG) have been developed to improve scalability of data acquisition. However, validated automated approaches to data annotation, including sleep staging are needed. Here, we apply deep learning approaches to the problem of sleep staging using data from the ANNE One (Sibel Health, Evanston, IL), a minimally intrusive flexible wireless dual sensor system measuring chest electrocardiography (ECG), triaxial accelerometry, and temperature, and finger photoplethysmography (PPG). Methods We obtained wearable sensor recordings from 281 adults undergoing concurrent clinical polysomnography at a tertiary care sleep lab. PSG recordings were scored according to AASM criteria. PSG and wearable sensor data were automatically aligned using their ECG signals with alignment confirmed by visual inspection. We trained a neural-network model to predict both 3-class (Wake, NREM, REM) and 2-class (Wake, Sleep) sleep stage classifications using a randomly selected 85% of the recordings and tested the model on the remaining recordings. We applied the model to ambulatory wearable sensor recordings from 233 older adults at risk for dementia. Our neural-network employed a convolutional-encoder and autoregressive-decoder architecture. In addition to time domain signals, we also engineered frequency domain features as well as selected scalar and metadata features as input to our model to improve performance. Ensembling of model variants was performed. Results Our approach achieved a 2-class macro-F1 of 0.718 with a sensitivity of 0.760 and specificity of 0.763 and a 3-class macro-F1 of 0.585 (wake precision 0.564 accuracy 0.745; NREM precision 0.886 accuracy 0.634; REM precision 0.258 accuracy 0.671). Our feature engineering and training techniques offered a 9% performance improvement from the time domain signals only baseline given the same neural network architecture, while ensembling different model variants offered a further 4% performance improvement. Conclusion A deep learning model can infer sleep stage from an EEG-less flexible wireless system and can be successfully applied to data from older community-dwelling adults at high risk for dementia. Support (if any) The Centre for Aging and Brain Health Innovation, Canadian Institutes of Health Research, National Institute on Aging