Assessing Risk of Health Outcomes From Brain Activity in Sleep

Abstract

Background and objectives: Patterns of electrical activity in the brain (EEG) during sleep are sensitive to various health conditions even at subclinical stages. The objective of this study was to estimate sleep EEG-predicted incidence of future neurologic, cardiovascular, psychiatric, and mortality outcomes.

Methods: This is a retrospective cohort study with 2 data sets. The Massachusetts General Hospital (MGH) sleep data set is a clinic-based cohort, used for model development. The Sleep Heart Health Study (SHHS) is a community-based cohort, used as the external validation cohort. Exposure is good, average, or poor sleep defined by quartiles of sleep EEG-predicted risk. The outcomes include ischemic stroke, intracranial hemorrhage, mild cognitive impairment, dementia, atrial fibrillation, myocardial infarction, type 2 diabetes, hypertension, bipolar disorder, depression, and mortality. Diagnoses were based on diagnosis codes, brain imaging reports, medications, cognitive scores, and hospital records. We used the Cox survival model with death as the competing risk.

Results: There were 8673 participants from MGH and 5650 from SHHS. For all outcomes, the model-predicted 10-year risk was within the 95% confidence interval of the ground truth, indicating good prediction performance. When comparing participants with poor, average, and good sleep, except for atrial fibrillation, all other 10-year risk ratios were significant. The model-predicted 10-year risk ratio closely matched the observed event rate in the external validation cohort.

Discussion: The incidence of health outcomes can be predicted by brain activity during sleep. The findings strengthen the concept of sleep as an accessible biological window into unfavorable brain and general health outcomes.

Background

Good sleep and healthy life are closely associated. For example, people with dementia have difficulty falling asleep and have reduced delta oscillations (1–4 Hz) in their brain waves (EEG) during deep sleep compared with matched controls without cognitive problems. People with depression or anxiety tend to have insomnia and increased sleep fragmentation compared with people without depression or anxiety while controlling for covariates. People with atrial fibrillation or congestive heart failure have fragmented sleep beyond the frequently present central and obstructive sleep apnea compared with controls. The ability to use physiologic measurements of sleep, such as EEG, to predict future incident health outcomes is significant because it could allow early interventions to prevent unfavorable outcomes. Because sleep is not only a window into sleep health but also a causal determinant, such interventions could also aim to improve sleep quality, which is a major goal of sleep medicine and the pharmaceutical industry. One approach is to develop measures of sleep quality that are associated with clinical outcomes by design. Along these lines, there are several recently introduced sleep-based biomarkers measuring different aspects of sleep quality that have been related to outcomes, including the sleep EEG-based brain age index (BAI), which attempts to measure the biological (as opposed to chronologic) age of the brain; the odds ratio product measuring sleep depth; cardiopulmonary coupling measuring sleep stability; and hypoxic burden measuring the extent of hypoxia due to apnea across the night. However, these biomarkers of sleep quality were not designed to, and do not explicitly predict future health outcomes.

Methods

The EEG montage can be either six channels: F3-M2, F4-M1, C3-M2, C4-M1, O1-M2, and O2-M1, or two channels: C3-M2 and C4-M1. The EEG should be resampled to 200 Hz, notch filtered to remove line frequency, and then band-pass filtered from 0.3 to 35 Hz. To remove artifacts, we excluded 30-second epochs containing absolute signal amplitudes higher than 500 µV or containing flat signal (standard deviation less than 0.2 µV) lasting longer than 5 seconds.

For each 30-second epoch, we extracted features from both spectral and temporal domains of sleep EEG. Spectral features included mean and kurtosis (a measure of extreme values to represent bursts) of band powers across 2-second sub-epochs from delta (1 Hz ≤ f < 4 Hz), theta (4 Hz ≤ f < 8 Hz), alpha (8 Hz ≤ f < 12 Hz), and sigma (12 Hz ≤ f < 16 Hz) bands and their band power ratios. Temporal features included density, amplitude, frequency, and duration of spindles; amplitude and density of slow oscillations (0.5–1.25 Hz); and coupling of spindles and slow oscillations.

Spindle and slow oscillation patterns were detected using Luna during NREM sleep. Specifically, spindles were detected based on a wavelet method, with a central frequency of 13.5 Hz and a wavelet cycle number set to 12. Slow oscillations were detected by first band-pass filtering between 0.2 Hz and 4.5 Hz, followed by detecting positive-to-negative zero crossings in the filtered signal, and then choosing intervals between 0.8 and 2 seconds having negative peak higher than 1.5 times the median voltage across all zero crossings and peak-to-peak amplitude higher than 1.5 times the median.

For each 30-second NREM epoch, we first extracted 57 features, and then each feature was averaged across all NREM epochs to represent the whole night. Similarly, for each 30-second REM epoch, we first extracted 21 features, and then each feature was averaged across all REM epochs. In addition, we included age, sex, body mass index (BMI) at the baseline sleep study, and 5 categories of medications that can affect sleep: benzodiazepines, antidepressants, sedatives, antiseizure drugs, and stimulants. In total, t here are 86 input features to summarize the information over the whole night of sleep.

Data Description

The extracted features for 8672 participants (8673 in the paper, 1 not mapped to the de-identified BDSPPatientID). Note the date of visits are shifted for de-identification (DOVshifted). The features include band powers, spindle characteristics, slow oscillation characteristics, and spindle-slow oscillation coupling metrics. The data contains more features than those described in the Methods section above, because it also contains features from the W stage.

Usage Notes

The data is in CSV format, with "," as the separator. It is a common and public format, that can be opened in many softwares.

Ethics

Data collection and sharing is performed under Institutional Review Board (IRB) approvals and data sharing agreements among participating hospitals, with waiver of the requirement for informed consent. The data is generated as part of usual patient care. All data is de-identified. 

Acknowledgements

During this research, Dr. Westover was supported by the Glenn Foundation for Medical Research and the American Federation for Aging Research through a Breakthroughs in Gerontology Grant; the American Academy of Sleep Medicine through an AASM Foundation Strategic Research Award; and grants from the NIH (R01NS102190, R01NS102574, R01NS107291, RF1AG064312, RF1NS120947, R01AG073410) and NSF (2014431). Dr. Thomas was supported by an AASM Foundation Strategic Research Award and by the NIH (RF1AG064312, R01NS102190).

Conflicts of Interest

Dr. Westover is a co-founder, scientific advisor, and consultant to  Beacon Biosignals and has a personal equity interest in the company.
R.J. Thomas discloses (1) patent and license/royalties from MyCardio, LLC, for the ECG-spectrogram; (2) patent and license/royalties from DeVilbiss-Drive for an auto-CPAP algorithm; and (3) consulting for Jazz Pharmaceuticals, Guidepoint Global, and GLG Councils. Other authors declare that they have no conflict of interest.