[Paper Review] The use of neural networks in the analysis of sleep stages and the diagnosis of narcolepsy.

Jens B. Stephansen, Aditya Ambati+20 more|arXiv (Cornell University)|Oct 5, 2017

Sleep and Wakefulness Research56 references19 citations

TL;DR

This study proposes a deep learning approach using neural networks to automate sleep stage scoring from polysomnography, generating detailed hypnodensity graphs that outperform human scorers (87% accuracy vs. consensus). It achieves 91% sensitivity and 96% specificity for Type-1 Narcolepsy diagnosis, with 99% specificity when combined with HLA-DQB1*06:02 typing, enabling potential home-based diagnosis and reducing clinical workload.

ABSTRACT

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 instead of 30 second 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.

Motivation & Objective

• To automate sleep stage scoring in polysomnography records using deep neural networks, reducing reliance on time-consuming manual scoring by trained technicians.
• To develop a hypnodensity graph representation that conveys richer information than traditional hypnograms by modeling sleep stage probabilities at high temporal resolution.
• To create a reliable, automated marker for Type-1 Narcolepsy diagnosis based on abnormal sleep-stage transitions, improving diagnostic accuracy and consistency.
• To validate the method in independent datasets and assess its potential for use in home sleep studies, increasing accessibility and reducing clinical burden.

Proposed method

• Trained a deep neural network on approximately 3,000 polysomnography recordings, including both normal and abnormal sleep data, to predict sleep stages at 5-second intervals.
• Generated hypnodensity graphs by modeling the probability distribution of sleep stages over time, offering higher temporal resolution than standard 30-second epoch scoring.
• Used a T1N marker based on the presence and frequency of abnormal sleep-stage overlaps (e.g., rapid transitions into REM sleep) to detect Type-1 Narcolepsy.
• Combined the neural network-based T1N marker with HLA-DQB1*06:02 genetic testing to further improve diagnostic specificity.
• Validated model performance against consensus scoring from six trained technicians on a cohort of 70 subjects.
• Evaluated diagnostic performance on independent datasets to ensure generalizability and robustness.

Experimental results

Research questions

• RQ1 Can a deep neural network achieve higher accuracy than human scorers in automated sleep stage scoring using polysomnography data?
• RQ2 Can a neural network-generated hypnodensity graph provide more informative sleep staging than traditional hypnograms?
• RQ3 Can an automated marker based on abnormal sleep-stage transitions reliably detect Type-1 Narcolepsy with high sensitivity and specificity?
• RQ4 Does combining the neural network-based T1N marker with HLA-DQB1*06:02 status further improve diagnostic specificity?
• RQ5 Can this approach be effectively applied to home sleep studies, reducing the need for in-clinic polysomnography?

Key findings

• The best-performing neural network model achieved 87% accuracy in sleep stage scoring, surpassing the consensus accuracy of six human scorers.
• The model successfully scored sleep stages at 5-second intervals, significantly improving temporal resolution compared to the standard 30-second epoch.
• The T1N marker based on abnormal sleep-stage overlaps achieved 91% sensitivity and 96% specificity in independent validation datasets.
• Incorporating HLA-DQB1*06:02 typing increased the specificity of the T1N diagnosis to 99%, enhancing diagnostic confidence.
• The method demonstrated strong generalization across independent datasets, supporting its potential use in clinical and home-based sleep monitoring.
• The approach reduces time and effort in sleep clinics and enables automated diagnosis of Type-1 Narcolepsy using home sleep studies.