A Quantized Parsimonious CNN Model for Sleep Polysomnogram Data Streams
An estimated 50 to 70 million Americans suffer from chronic sleep disorders. Humans spend at least a third of their lives sleeping, and quality sleep is vital for adequate brain performance, mood, and health. Getting less than the recommended 7 to 9 hours of sleep per night increases the risk of many diseases and disorders, such as insomnia, narcolepsy, and restless legs syndrome. To diagnose sleep disorders, sleep specialists attach sensors to various parts of a person's body to collect signals known as polysomnography. Typically, sleep experts manually grade sleep stages, which is a complex, costly, and problematic process. Since the accuracy of sleep stage classification depends on the technicians' experience and level of exhaustion, there is often less than 90 % agreement among them. The increased use of wearable and edge devices enables the daily and continuous monitoring of potential sleep orders in natural environments but also challenges the deployment of complicated models such as the convolutional neural network (CNN), due to the lower memory requirement of these edge devices. This paper contributes to this field in two folds. First, un-like existing complicated CNNs, a pre-quantized parsimonious CNN model, i.e., 1D CNN with only six layers, was developed. Compared with five typical but less complicated models across training, testing, and validation sets of sleeping stage data streams of EEG signals, this parsimonious CNN achieves the stage classification accuracy range of 95-99 %. Second, built upon this highly accurate parsimonious CNN model, a quantized parsimonious CNN of different bits was further developed and evaluated to be used on wearables or edge devices. Although the overall accuracy of this quantized CNN decreased from this pre-quantized CNN, which is true for quantized models, this new quantized parsimonious CNN achieved an accuracy as high as about 88-98 % of EEG signals for specific sleep stages. These findings, on one hand, demonstrate that this parsimonious CNN shows a less complex structure with similar or higher model performance over existing CNNs based on their reported benchmark performance (e.g., accuracy of 85-90 %), and a consistent performance improvement over existing less complex approaches, which could potentially enable more efficient sleep diagnosis; on the other hand, this quantized parsimonious CNN could be deployed with reasonable accuracy, potentially useful for certain sleeping stages particularly based on EEG signals in natural environments.
