[ad_1]
Sleep apnea, a disorder that occurs when a person's breathing is interrupted while asleep, affects approximately 22 million Americans. The problem is that the majority of cases – 80% – are undiagnosed and that, if they are not treated, sleep apnea may increase the risk of coronary heart disease, heart attack, heart failure and stroke.
One area of study – the Automatic Snoring Clbadification, or ASSC – aims to develop a sleep apnea-based diagnosis method based on snoring (sleep apnea is characterized by repetitive episodes airflow decreased or completely stopped). However, despite the progress made in recent years, the labeled data on which ASSC systems can be formed is lacking.
That's why researchers at Imperial College London, the University of Augsburg and the Technical University of Munich have asked in a new article ("Snore-GANs: Improvement of automatic clbadification of humming noise with synthesized data "), to develop data to fill the gaps in the actual data. (For the uninitiated, GANs are two-part neural networks made up of generators that produce samples and discriminators that attempt to distinguish between generated samples and real-world samples.) The Rich Data Set was then used to form an ASSC.
"When increasing data, we aggregate data from all [GANs]and randomly select pool data that is then merged into the original training set. In doing so, it is planned to expand the diversity of augmented data from different sources. [GANs], Explain the authors of the document.
To validate their method, they used a publicly available dataset – the Munich-Pbadau Snore Sound Corpus (MPSSC) – to rank the location of vibrations in the upper airways during snoring, starting with the recordings existing reviews of three German medical centers performed during the year. clinical examinations between 2006 and 2015. The snoring selected were clbadified by medical experts in otolaryngology based on the results of video recordings, and the annotated samples were separated into sets of training, development and Independent test subject.
So, what is the price of the approach? The authors of the paper indicate that they have succeeded in generating data sharing a distribution from the original data, which has resulted in an increased amount of training data without the need for human annotation. In addition, the combination of synthesized and original data has improved the performance of the clbadifier.
"In the future, we will continue to collect more data on snoring from different hospitals and patients to increase the size and diversity of data, which will allow us to re-evaluate the proposed methods," he said. Researchers. "In addition, more advanced systems … will be proposed and evaluated in our subsequent work to improve acoustic sequence generation models."
Source link
