Deep Learning for Sleep State Detection Using CW Doppler Radar Technology

Sleep is an essential function of the body to protect the mental and physical health of an individual. Monitoring and understanding the sleep patterns can help preventing diseases and designing tailored treatments for patients with sleep disorders. The goal of this work is to detect sleep and wake stages using deep learning techniques from radar data provided by the Swiss company Sleepiz AG. The aim is to understand the applicability of deep learning techniques for the sleep staging task and to evaluate the classification performance as well as time-requirements of the proposed approaches. For this purpose, raw signals and a time-frequency representation of the signals were used as input for several deep learning approaches to classify single segments of the input signal. Then, a sequential deep learning approach was introduced to generate predictions by processing an entire recording and generating predictions per each time step that were dependent on the previous and next inputs. The best model, named RawSeqEnsemble, uses night recordings data as input and it automatically generates predictions for the entire recording in a single moment. RawSeqEnsemble outperforms the existing model at Sleepiz AG based on a conventional random forest classifier applied on hand-engineered. Specifically, RawSeqEnsemble yielded approximately to 10 and 11 percentage points increase in Cohen’s kappa and wake F1-score, respectively, on a test set with 73 recordings. The results show that, using a sequential deep learning method, it is possible to learn spatial and temporal features directly from raw Sleepiz data based on radar-technology without the need of hand-crafted features. Also, the proposed approach not only generalizes well on different recordings, but it also predicts classes about 70 time faster for an entire recording on the same hardware