Q&A: Electrocardiography-based sleep stage scoring ‘on par’ with polysomnography

Key takeaways:

• Electrocardiography data can be put into a neural network to score/determine sleep stages.
• With electrocardiography, fewer electrodes are needed during testing vs. polysomnography.

Polysomnography is the current gold standard in sleep testing, but a new method has achieved similar performance in sleep stage scoring via an electrocardiography-based deep learning neural network, according to a press release.

The method is coined cardiosomnography (CSG) and means a sleep study conducted with just electrocardiography (ECG/EKG).

Rather than having a clinician score/determine sleep stages, which is the case in polysomnography (PSG), researchers behind CSG explained that ECG data can be put into a neural network to determine sleep stages.

Additional benefits of CSG include its ability to reduce the number of electrodes an individual needs to wear while sleeping and its allowance for at-home testing rather than sleep lab testing.

To learn more about the technology behind CSG and what clinicians need to implement this method, Healio spoke with Adam Jones, PhD, former graduate student at the University of Southern California, and Bhavin R. Sheth, PhD, associate professor of electrical and computer engineering at the University of Houston.

Healio: What is a neural network?

Jones and Sheth: At its most basic, a neural network is an algorithm that somewhat mimics the structure of brains (interconnected layers of neurons) to process inputs and produce an output.

In this case, our network takes as its input the ECG recording for the entire night, the age and sex of the individual, and the time that the recording began. The network then produces a “score” for each 30-second period of the night. Scoring, in this case, means to determine which sleep stage (ie, Wake, N1, N2, N3, REM) an individual is in. However, these labels will be somewhat unfamiliar to the lay public. Instead, in current commercial devices, they will often list “light” sleep (N1 + N2), “deep” sleep (N3), and REM (the stage where dreaming is more likely to occur). One reason current commercial devices skip N1 is because it is notoriously difficult to classify unless using PSG — or now, cardiosomnography (CSG).

Healio: How did your system perform vs. PSG?

Jones and Sheth: We demonstrated with an extensive meta-analysis of PSG studies that CSG performs on par with PSG for sleep staging. It is this on par performance that supports the claim that this could complement or replace PSG in numerous cases.

Furthermore, we show that the best “electroencephalography (EEG)-less” studies and devices (“EEG-less” = no brain waves recorded, such as the Oura Ring, Apple Watch and Fitbit) perform poorly compared with PSG or CSG.

Healio: With your method, an individual in need of a sleep study only has to wear two electrodes. How many electrodes does an individual typically have to wear during a PSG test? What are the benefits of reducing the number of electrodes a patient must wear?

Jones and Sheth: The number of electrodes that are worn during PSG depends on several factors, including available equipment and if it is of research or clinical interest. This number is usually 12 to 36, but it can go much higher.

What is more important than the number, however, is their location. Traditional PSG relies extensively on measuring EEG, or brain waves. These necessitate caps that can hold the electrodes in place over the skull within the hair. Furthermore, PSG has electrodes placed near the eyes, chin, across the chest and on one leg. With all the electrodes, tape and wires, it is quite cumbersome for the sleeper. Additionally, if those electrodes disconnect during the night (which is very easy with EEG electrodes that must be placed over hair), a technician will have to wake the sleeper up to reconnect them.

One of the main benefits of CSG revolves around the reduction in electrodes and in their more comfortable locations. Now, using a single lead (ie, two electrodes and possibly a separate ground) of ECG/EKG, many of the clinical and research requirements for PSG can be duplicated. Because ECG is a very strong signal, the requirements for the electrode connection with the skin are minimized — meaning a more comfortable setup.

Healio: What do clinicians need to do to implement the method you developed? Do health care providers need training? Is the equipment available?

Jones and Sheth: For the enterprising clinicians, and some have already reached out to us, they can use the ECG equipment they already have to record and score sleep using CSG. However, it will likely take some time for this to become widely used in clinical settings, purely because, unlike research settings, clinical facilities and staff are more familiar with FDA-approved turn-key devices and software. However, there is nothing limiting them from using our software with their current equipment.

On the other hand, research settings, which are more accustomed to having to “roll their own solutions,” could be expected to easily make use of CSG right away.

Our overarching goal is to get this into use as widely as possible. Therefore, we have provided the code that was used in the research so that others can use it free of charge at https://cardiosomnography.com/.

Healio: How would at-home sleep tests benefit patients?

Jones and Sheth: One of the first use-cases that we propose for this method would be to take many of the PSG studies that are currently conducted in clinics and bring them into the home. The familiar environment alone could alleviate one of the paradoxical issues of current sleep studies: the unfamiliar environment is different to sleep at home (in some cases, difficult and in others, counter-intuitively better). Moreover, the study of sleep in remote populations, which is a known tool in basic sleep research, can be better served with CSG in our opinion.

The next major use case would be to transform sleep studies from one-off diagnostic tests into long-term clinical monitoring. In the paper we discuss the possibility of long-term monitoring of individuals at-risk for developing dementia, as there are well-documented sleep disturbances that are correlated with a diagnosis. However, by monitoring sleep more frequently, it might be possible to track the efficacy of interventions and medications, and to respond accordingly.

Healio: What is the next step in this research?

Jones and Sheth: We think our findings are both groundbreaking and robust; however, many questions remain unanswered as to what is happening within the neural network to allow it to do this so well. As alluded to earlier, decades of researchers have tried and failed to replace PSG with ECG (or other similar sensors) at an equivalent performance. The next step will be to try to better understand what the network is focusing on in the heart data to achieve its remarkable performance.

Additionally, since CSG does not monitor directly the state of the brain per se, we suggest our work argues for the involvement and engagement of the body, and not just the brain, in sleep.

Healio: Do you have a timeline for filing for patents or FDA approval? Do you expect eventual commercialization?

Jones and Sheth: We have no interest in commercial development, but we hope that some enterprising company or companies take this finding as an impetus to develop more commercial wireless ECG devices. We expect that they could look and feel quite similar to current heart rate chest straps. However, the major difference would be that heart rate straps only transmit the heart rate, which is akin to the tempo of a piece of music, instead of a recording of the music.

Editor’s note: On Aug. 16, 2024, Adam Jones, PhD, alerted Healio that there are two commercial wireless ECG devices in the 'heart rate chest strap' form factor already on the market: H10 (Polar) and Movesense MD (Movesense).

References:

• Groundbreaking approach to sleep study expands potential of sleep medicine. https://uh.edu/news-events/stories/2024/july/07022024-sheth-sleep-staging-monitoring.php. Published July 2, 2024. Accessed July 8, 2024.
• Jones AM, et al. Comput Biol Med. 2024;doi:10.1016/j.compbiomed.2024.108545.