Pap Therapy and Sleep Stage Classification with Statistics, Signal Processing, and Deep Learning

Abstract

This paper focuses on PAP (Positive Airway Pressure) therapy compliance and automated sleep staging. These are related to sleep medicine, data science, and signal processing. During a sleep study, an EEG (electroencephalogram) recording is labeled with sleep stages by a technician. The sleep stage labels combined with other recordings, observations, and analysis are used to diagnose a patient with obstructive sleep apnea. PAP therapy is the primary treatment for obstructive sleep apnea. Compliant use of the PAP thereapy is considered 4 hours of use on 70% of nights. The PAP therapy focus is a retrospective study on patients from OSA In Home (formerly Sleep Centers of Middle Tennessee) to find the impacts of the dispenser of PAP therapy between durable medical equipment (DME) supplier (DME group) versus those provided directly by an integrated sleep practice (ISP group). The ISP group had a significantly higher rate of PAP adherence at 30 days (71% vs 66%; P = .004), 90 days (66% vs 56%; P < .00001), and 1 year (52% vs 33%; P < .00001) following initiation of PAP therapy, relative to the DME group. There was a significantly higher duration of PAP use among the ISP group at 30 days (357 vs 345 minutes; P = .002), 90 days (348 vs 319 minutes; P < .00001), and 1 year (312 vs 164 minutes; P < .00001). The automated sleep stageing focus is a study on REM sleep classification with interpretable models. The interpretability of the model is based on relating the filters of the CNN (Convolutional Neural Network) part of the model to clincal sleep markers used by technicians to grade EEGs during a sleep study. The model is created using CNNs, full connected neural networks, and GRUs (gated recurrent unit). The data is a unprocessed or raw single channel EEG from ”The Sleep-EDF Database [Expanded]”. The best results produced 97% accuracy, 93% precision, and 89% recall. The data is then preprocessed by DWT (multilevel discrete wavelet transform) with 45 different mother wavelets and reconstructed as input to the CNN. The best f1 score with DWT preprocessing and with raw data are 95% with coiflet 4 mother wavlet and 93% respectively. The filters of the CNNs are compared with Spearman's rank coefficient to find filters that are most correlated in each EEG frequency band. The most correlated filter was preprocessed with Daubechies order 8 mother wavelet had an aboslute value of the Peasron's corelation coefficient of greater than 0.3415 with 15% of the other filters related to the same frequency band.