Sleep Stages Classification in Sleep Disorder Patients: Integrating Wearable and Contactless Commercial Devices

Sleep disorders decrease the quality of sleep for affected individuals, potentially leading to serious, negative health effects. Therefore, it is essential to promptly diagnose these disorders and subsequently monitor their progression. The diagnosis of sleep disorders involves the examination of sleep, categorized into distinct stages, with polysomnography (PSG) currently considered the gold standard for assessment. However, PSG has limitations: it is expensive, time-consuming, complicated to operate, obtrusive, and usually only performed on a single night. One possible solution to these limitations is to leverage commercially available wearable and contactless devices that are already capable of providing sleep stages classification. These devices are affordable, easy to use, comfortable, and suitable for multiple nights of use. To investigate this alternative, this study analyzes the data collected from patients with sleep disorders to whom two wearable devices (Fitbit Inspire 2 and Empatica E4) and two contactless devices (Somnofy and Emfit) were added during PSG. The initial step consisted of an in-depth evaluation of the sleep stages automatically provided by Somnofy, Fitbit and Emfit in comparison to the PSG for this particular type of patients. The devices demonstrated an overall accuracy of 67% for Somnofy, 64% for Fitbit and 47% for Emfit. Statistically significant differences were found in all sleep measures, such as total sleep time and REM latency, with particular difficulty in detecting cases of very short sleep stages durations. The dataset was then used to fine-tune some models using signals from Empatica to classify sleep stages. The results aligned with those of the other devices, performing better than Fitbit and worse than Somnofy. Lastly, a novel approach to sleep stages classification was proposed: fusing sleep stages from multiple devices. A random forest was trained to classify the sleep stage of an epoch based on the sleep stages predicted by devices at that epoch. Sleep stages from Somnofy, Fitbit, Emfit, and the fine-tuned Empatica model were incorporated, and all possible combinations of two to four devices were tested. This method achieved the highest accuracy of 73% when fusing Somnofy, Fitbit, and Empatica. It was generally more accurate than the devices used alone, particularly when combining three or four devices. In conclusion, this study demonstrates the potential of using commercially available devices for sleep stages classification. Encouraging results were achieved through the integration of multiple devices. Despite some limitations, these devices represent a promising path toward more comprehensive and accessible sleep monitoring for both healthy individuals and patients with sleep disorders.