New Deep Learning Approach Improves Access to Sleep Diagnostic Testing

Researchers from Georgia Tech and Massachusetts General Hospital created a powerful algorithm that can assess sleep scoring as accurately as a sleep technologist.

A new deep learning approach can automatically analyze and score sleep tests as effectively as sleep technologists, according to researchers from Georgia Tech’s School of Computational Science and Engineering (CSE) and the Neurology Department of Massachusetts General Hospital (MGH).

The breakthrough – outlined in a paper published in the December 2018 issue of the Journal of American Medical Informatics Association (JAMIA) – will enable greater access to critically needed diagnostic testing for the 40 million people in the United States who suffer from chronic long-term sleep disorders.

Currently, it takes a certified sleep technologist one to two hours to manually assess and score a polysomnography (PSG) test, a key test in sleep disorder studies, from just one night of sleep.

The new automated approach uses a combination of deep recurrent and convolutional neural networks (RCNN) and was trained on 10,000 MGH sleep studies, one of the world’s largest clinical data sets of sleep study information. According to the study, the model produces results that are as accurate as those of an experienced sleep technologist in a fraction of the time. 

“Timely and accurate diagnosis of sleep disorders is critical to pursue appropriate treatment and improve health outcomes, yet most sleep disorders remain undiagnosed,” said CSE Associate Professor Jimeng Sun, a member of the research team leading the study. 

“The amount of time required to manually score sleep charts creates a bottleneck in the sleep assessment process, which often prevents patients from having access to sleep diagnostic testing for longer durations [between appointments] or all together.”

MGH researchers and neurologists Dr. Brandon Westover and Dr.Matt Bianchi noticed the unmet need with sleep study access and an opportunity with its data labeling process more than 10 years ago.

“One of our problems with the way medicine is done generally, and in our own specialty of neurology, is that there is a lot of subjectivity,” said Westover. “We realized that sleep scoring was a perfect procedure to automate because the data is routinely labeled by experts.”

Deep RCNN layout for automated polysomnography analysis.

Figure 1. From 'Expert-level Sleep Scoring with Deep Neural Networks' - Deep RCNN layout for automated polysomnography analysis.

After collecting sleep study data at MGH for a decade, Bianchi and Westover began collaborating with Georgia Tech colleagues, Sun and Ph.D. student Siddharth Biswal, to craft a powerful computer algorithm with enough information to adequately represent the diversity of clinical sleep signals, removing the subjective element of sleep data interpretation entirely. 

“In the past, most of the work in designing [sleep study] algorithms focused cleaning data, deciding which features to extract, and finally training algorithms on small data sets. But the dozens of algorithms in the literature that take this approach don’t work in the real world. Our approach was to collect truly big data, from thousands of real patients — not carefully selected research subjects. This rich data let us train a very flexible, powerful machine learning algorithm, directly from the raw data, that works across the entire range of ages and sleep problems that we encounter in patients with sleep problems,” said Westover. 

“We believe this new method will increase the throughput of the sleep clinics that exists now and help extend the reach of sleep medicine to more patients beyond sleep clinics,” he said.

With the research published, the new automated assessment and scoring tool will be helping real-world patients in the short term. According to Sun, a longer-term project will combine new sensors to apply the new approach to at-home methods of sleep diagnostics.


Kristen Perez

Communications Officer I