Proteomic biomarkers of sleep apnea.

STUDY OBJECTIVES: Obstructive sleep apnea (OSA) is characterized by recurrent partial to complete upper airway obstructions during sleep, leading to repetitive arousals and oxygen desaturations. Although many OSA biomarkers have been reported individually, only a small subset have been validated through both cross-sectional and intervention studies. We sought to profile serum protein biomarkers in OSA in unbiased high throughput assay. METHODS: A highly multiplexed aptamer array (SomaScan) was used to profile 1300 proteins in serum samples from 713 individuals in the Stanford Sleep Cohort, a patient-based registry. Outcome measures derived from overnight polysomnography included Obstructive Apnea Hypopnea Index (OAHI), Central Apnea Index (CAI), 2% Oxygen Desaturation index, mean and minimum oxygen saturation indices during sleep. Additionally, a separate intervention-based cohort of 16 individuals was used to assess proteomic profiles pre- and post-intervention with positive airway pressure. RESULTS: OAHI was associated with 65 proteins, predominantly pathways of complement, coagulation, cytokine signaling, and hemostasis which were upregulated. CAI was associated with two proteins including Roundabout homolog 3 (ROBO3), a protein involved in bilateral synchronization of the pre-Bötzinger complex and cystatin F. Analysis of pre- and post intervention samples revealed IGFBP-3 protein to be increased while LEAP1 (Hepicidin) to be decreased with intervention. An OAHI machine learning classifier (OAHI >=15 vs OAHI<15) trained on SomaScan protein measures alone performed robustly, achieving 76% accuracy in a validation dataset. CONCLUSIONS: Multiplex protein assays offer diagnostic potential and provide new insights into the biological basis of sleep disordered breathing.

Design of a deep learning model for automatic scoring of periodic and non-periodic leg movements during sleep validated against multiple human experts.

OBJECTIVE: Currently, manual scoring is the gold standard of leg movement scoring (LMs) and periodic LMs (PLMS) in overnight polysomnography (PSG) studies, which is subject to inter-scorer variability. The objective of this study is to design and validate an end-to-end deep learning system for the automatic scoring of LMs and PLMS in sleep. METHODS: The deep learning system was developed, validated and tested, with respect to manual annotations by expert technicians on 800 overnight PSGs using a leg electromyography channel. The study includes data from three cohorts, namely, the Wisconsin Sleep Cohort (WSC), Stanford Sleep Cohort (SSC) and MrOS Sleep Study. The performance of the system was further compared against individual expert technicians and existing PLM detectors. RESULTS: The system achieved an F1 score of 0.83, 0.71, and 0.77 for the WSC, SSC, and an ancillary study (Osteoporotic Fractures in Men Study, MrOS) cohorts, respectively. In a total of 60 PSGs from the WSC and the SSC scored by nine expert technicians, the system performed better than two and comparable to seven of the individual scorers with respect to a majority-voting consensus of the remaining scorers. In 60 PSGs from the WSC scored accurately for PLMS, the system outperformed four previous PLM detectors, which were all evaluated on the same data, with an F1 score of 0.85. CONCLUSIONS: The proposed system performs better or comparable to individual expert technicians while outperforming previous automatic detectors. Thereby, the study validates fully automatic methods for scoring LMs in sleep.

Neural network analysis of sleep stages enables efficient diagnosis of narcolepsy.

Analysis of sleep for the diagnosis of sleep disorders such as Type-1 Narcolepsy (T1N) currently requires visual inspection of polysomnography records by trained scoring technicians. Here, we used neural networks in approximately 3,000 normal and abnormal sleep recordings to automate sleep stage scoring, producing a hypnodensity graph-a probability distribution conveying more information than classical hypnograms. Accuracy of sleep stage scoring was validated in 70 subjects assessed by six scorers. The best model performed better than any individual scorer (87% versus consensus). It also reliably scores sleep down to 5 s instead of 30 s scoring epochs. A T1N marker based on unusual sleep stage overlaps achieved a specificity of 96% and a sensitivity of 91%, validated in independent datasets. Addition of HLA-DQB1*06:02 typing increased specificity to 99%. Our method can reduce time spent in sleep clinics and automates T1N diagnosis. It also opens the possibility of diagnosing T1N using home sleep studies.

Deep residual networks for automatic sleep stage classification of raw polysomnographic waveforms.

We have developed an automatic sleep stage classification algorithm based on deep residual neural networks and raw polysomnogram signals. Briefly, the raw data is passed through 50 convolutional layers before subsequent classification into one of five sleep stages. Three model configurations were trained on 1850 polysomnogram recordings and subsequently tested on 230 independent recordings. Our best performing model yielded an accuracy of 84.1% and a Cohen's kappa of 0.746, improving on previous reported results by other groups also using only raw polysomnogram data. Most errors were made on non-REM stage 1 and 3 decisions, errors likely resulting from the definition of these stages. Further testing on independent cohorts is needed to verify performance for clinical use.

Comparison of computerized methods for rapid eye movement sleep without atonia detection.

Rapid eye movement (REM) sleep without atonia detection is a prerequisite for diagnosis of REM sleep behavior disorder (RBD). As the visual gold standard method is time-consuming and subjective, several automated methods have been proposed. This study aims to compare their performances: The REM atonia index (RAI), the supra-threshold-REM-activity metric, the Frandsen index, the short/long muscle activity indices, and the Kempfner index algorithms were applied to 27 healthy control participants (C), 25 patients with Parkinson's disease (PD) without RBD (PD-RBD), 29 patients with PD and RBD (PD + RBD), 29 idiopathic patients with RBD, and 36 patients with periodic limb movement disorder (PLMD). The indices were calculated in various configurations: (1) considering all muscle activities; (2) excluding the ones related to arousals; (3) excluding the ones during apnea events; (4) excluding the ones before and after apnea events; (5) combining configurations 2 and 3; and (6) combining configurations 2 and 4. For each of these configurations, the discrimination capability of the indices was tested for the following comparisons: (1) (C, PD-RBD, PLMD) vs (PD + RBD, RBD); (2) C vs RBD; (3) PLMD vs RBD; (4) C vs PD-RBD; (5) C vs PLMD; (6) PD-RBD vs PD + RBD; and (7) C vs PLMD vs RBD. Results showed varying methods' performances across the different configurations and comparisons, making it impossible to identify the optimal method and suggesting the need of further improvements. Nevertheless, RAI seems the most sensible one for RBD detection. Moreover, apnea and arousal-related movements seem not to influence the algorithms' performances in patients' classification.