Source-linked AI summary
Sleep Staging from Electrocardiography and Respiration with Deep Learning
Haoqi Sun, Wolfgang Ganglberger, Ezhil Panneerselvam, Michael J. Leone, Syed A. Quadri, Balaji Goparaju, Ryan A. Tesh, Oluwaseun Akeju, Robert J. Thomas, M. Brandon Westover
TL;DR
Sleep staging typically relies on EEG, but ECG and respiratory signals may carry staging information and can be available when EEG analysis is difficult. This study trains CNN-LSTM systems on PSG signals; ECG+ABD achieves the highest kappa, including 0.762 for Wake vs. NR vs. REM.
Problem
Sleep staging relies primarily on EEG, while autonomic patterns suggest ECG and respiratory signals may also distinguish stages, including in settings where EEG is difficult to analyze.
Source: Introduction · p. 2Method
The study trains CNN-LSTM models using ECG, chest or abdominal respiratory effort, and combinations of these signals from PSG recordings.
Source: Introduction · p. 2, Methods · p. 4Results
ECG+ABD had the highest kappa across evaluated groupings, reaching 0.762 for Wake vs. NR vs. REM.
Source: Results · p. 8Takeaways & Limitations
ECG and respiratory effort provide substantial information about sleep stages, with performance varying to some extent by age, apnea-hypopnea index, and sleep study type.
Source: Discussion · p. 14The dataset includes only adults, so generalizability to pediatric groups requires additional study.
Source: Discussion · p. 13
Abstract
from arXiv · showhide
Study Objective: Sleep is reflected not only in the electroencephalogram but also in heart rhythms and breathing patterns. Therefore, we hypothesize that it is possible to accurately stage sleep based on the electrocardiogram (ECG) and respiratory signals. Methods: Using a dataset including 8,682 polysomnographs, we develop deep neural networks to stage sleep from ECG and respiratory signals. Five deep neural networks consisting of convolutional networks and long short-term memory networks are trained to stage sleep using heart and breathing, including the timing of R peaks from ECG, abdominal and chest respiratory effort, and the combinations of these signals. Results: ECG in combination with the abdominal respiratory effort achieve the best performance for staging all five sleep stages with a Cohen's kappa of 0.600 (95% confidence interval 0.599 -- 0.602); and 0.762 (0.760 -- 0.763) for discriminating awake vs. rapid eye movement vs. non-rapid eye movement sleep. The performance is better for young participants and for those with a low apnea-hypopnea index, while it is robust for commonly used outpatient medications. Conclusions: Our results validate that ECG and respiratory effort provide substantial information about sleep stages in a large population. It opens new possibilities in sleep research and applications where electroencephalography is not readily available or may be infeasible, such as in critically ill patients.
Introduction
Sleep staging has traditionally relied on EEG, but sleep-related autonomic patterns suggest that ECG and respiration may also carry useful information. The paper develops deep-learning methods to stage sleep from these non-EEG signals.
Introduction
- Sleep analysis commonly uses EEG to distinguish wake, REM, and NREM, with NREM further divided into N1, N2, and N3.Slow-wave oscillations characterize deeper NREM, while spindles and theta oscillations characterize lighter NREM.Source: Introduction · p. 2
- Autonomic activity couples with cortical activity during sleep, and characteristic heart-rate and breathing patterns suggest sleep stages may be staged from non-EEG signals.REM includes respiratory and heart-rate fluctuations, while wake is associated with low-frequency heart-rate variability and large movements.Source: Introduction · p. 2
- ECG and respiration may be practical alternatives because ECG is often continuously recorded and wearable devices increasingly measure both signals.EEG can be highly abnormal in medically ill populations, making standard analysis difficult.Source: Introduction · p. 2
- The study trains CNN-LSTM networks on ECG and respiratory signals using a large clinical dataset spanning varied ages, medications, and sleep disorders.The approach builds on prior deep-learning sleep staging from EEG, reported at 87.5% accuracy and Cohen’s kappa 0.805.Source: Introduction · p. 2
Methods
The methods use PSG-derived ECG and respiratory effort to predict the sleep stage of each 30-second epoch from signal context and sequence patterns. Multiple single- and paired-signal CNN-LSTM models are trained and evaluated on held-out PSGs.
Methods
- The analysis includes 8,682 PSGs after exclusions, with five sleep stages annotated in non-overlapping 30-second epochs.The PSGs were acquired at Massachusetts General Hospital from 2009 to 2016 and included ECG plus chest and abdominal effort channels.Source: Methods · pp. 2–3
- Each prediction targets the middle 30-second epoch of a 270-second window, providing 120 seconds of signal context on either side.Longer context was used because heart-rhythm and respiration changes often span longer timescales.Source: Methods · p. 4
- Five models use ECG, chest effort, abdominal effort, ECG+chest, or ECG+abdomen; artifact epochs are excluded from CNN training but retained for LSTM context.ECG is represented by R-peak timing, while respiratory signals are down-sampled to 10 Hz.Source: Methods · p. 4
- Each model combines a CNN for epoch-level features with an LSTM for temporal patterns across consecutive epochs.The LSTM uses bidirectional context, concatenating activations from the forward and backward directions.Source: Methods · p. 4
- For paired signals, pretrained modality-specific CNN layers are fixed, their outputs are concatenated, and a further CNN subnetwork processes the combined features.The architecture then feeds CNN outputs into a bidirectional LSTM and a dense output layer.Source: Methods · p. 4, Figure 1 · p. 5
- The PSGs are split into 6,682 training, 1,000 validation, and 1,000 testing records; CNNs are trained before LSTMs using CNN outputs.Training uses cross-entropy, and model settings are selected by validation objective.Source: Methods · p. 5
- Training weights each 270-second input by the inverse frequency of its sleep stage to reduce dominance of common stages such as N2.The paper notes that people spend about 50% of sleep in N2 and 20% in N3.Source: Methods · p. 6
- Testing assesses five-stage and two three-super-stage classifications with confusion matrices and Cohen’s kappa; kappa confidence intervals use 1,000 bootstrap samples.The three-stage groupings are awake-or-drowsy versus sleep versus REM, and awake versus NREM versus REM.Source: Methods · p. 6
Results
ECG plus abdominal respiration yielded the strongest staging agreement, while performance varied with participant and study characteristics. The discussion connects stage grouping to sleep biology and identifies limits in the study’s scope and evaluation.
Results
- With ECG+ABD, accuracy was 81.8% for wake, 55.8% for N1, 66.3% for N2, 66.6% for N3, and 92.2% for REM; errors clustered among adjacent stages.N3 was often classified as N2, while N1 was often classified as wake or N2.Source: Results · p. 6, Figure 2 · p. 6
- ECG+ABD had the highest Cohen’s kappa: 0.6 for all five stages, 0.74 for W+N1 vs. N2+N3 vs. REM, and 0.762 for Wake vs. NR vs. REM.The testing set contained 1,000 PSGs.Source: Results · p. 8, Table 2 · p. 8
- Performance was lower in older participants, higher-AHI groups, and split-night studies, but severe-apnea kappa remained 0.574 for five stages and above 0.7 for three super-stages.The reported group comparisons were statistically significant at the 0.05 level.Source: Results · p. 8, Table 3 · p. 8
- Individual-PSG kappas varied substantially, from around -0.1 at the low end to around 0.9 at the high end.The distribution was right-skewed.Source: Results · p. 9, Figure 5 · p. 9
- Typical-recording examples showed visible correspondence between signal spectrograms and predicted stages, but heart-rate variability features sometimes disagreed with EEG-based staging.In one example, loss of very-low-frequency power was classified as N3 while EEG-based staging included both N2 and N3.Source: Results · p. 10
Discussion
- The networks were reported robust across commonly used medications and OSA/PLMS; ICU or hospital use would require training with enough ICU patients receiving varied heart-rate or blood-pressure medications.Source: Discussion · pp. 12–13
- ECG-and-respiration staging performed below EEG-based staging; grouped stages may better reflect sleep biology, while similar N2/N3 autonomic signatures can complicate EEG-based distinctions.The discussion notes that stable N2 and N3 may have very similar or identical cardiorespiratory signatures.Source: Discussion · p. 13
- The study included adults only; 30-second scoring, limited network interpretability, and a single train-validation-test split constrain the analysis.The authors note that cross-validation across multiple folds would be less biased.Source: Discussion · pp. 13–14
Disclosure Statement
The disclosure lists the authors’ financial and professional interests related to sleep medicine and ECG-based sleep technologies.
Disclosure Statement
- RJT reports patents, licensing, and royalties from MyCardio for an ECG-based sleep-quality and sleep-apnea method, plus intellectual property for a CO2 device.The disclosure also notes GLG consulting in general sleep medicine.Source: Disclosure Statement · p. 14
- BG reports that the work was done while at Massachusetts General Hospital and that BG is now a Novartis data scientist.Source: Disclosure Statement · p. 14
Supplementary Figures
The supplementary materials show signal segmentation, supplementary visualizations of signal and network features, and subgroup kappa tables for different inputs.
Supplementary Figures
- The supplementary figures include Cohen’s kappa histograms across input signals and sleep-stage groupings.Source: Figure S2 · p. 17
- Figure S3 shows Cohen’s kappa when zero-mean Gaussian jitter is applied to ECG R peaks.Source: Figure S3 · p. 18
- Figures S4–S6 visualize last-layer activations for ECG, chest respiration, and abdominal respiration, with each point representing a 270-second epoch.Source: Figure S4 · p. 18, Figure S5 · p. 19, Figure S6 · p. 19
- Figures S7–S9 show first-convolution-layer kernels for ECG, chest respiration, and abdominal respiration; Figures S10–S12 show high-probability ABD, CHEST, and ECG examples across sleep stages, not necessarily from the same recording.Source: Figure S7 · p. 20, Figure S8 · p. 20, Figure S9 · p. 21
Supplementary Tables
- Tables S1–S4 report Cohen’s kappa by group in the 1,000-PSG testing set for ECG, chest, abdominal, and ECG-plus-chest inputs.The supplementary-table text also includes the labels “W vs. NR vs.” and “Periodic limb movement (per hour).”Source: Table S1 · p. 23, Table S2 · p. 24, Table S3 · p. 25