News
Article
FDA Approves HoneyNaps’ Artificial Intelligence Sleep Disorder Diagnosis Software
Author(s):
The AI-based SOMNUM technology uses diagnostic algorithms based on multi-channel/time series sleep biosignal data to deliver fast, accurate diagnostic information for providers.
Ji Ho Choi, MD, PhD
The FDA recently approved HoneyNaps’ SOMNUM artificial intelligence (AI) sleep disease analysis algorithm, the first AI solution designed for diagnosing sleep disorders.1
SOMNUM, supported by eXplainable Medical AI technology, utilizes deep learning-based AI to conduct a real-time analysis of vast volumes of multi-channel/time series biosignals. The technology is believed to surpass conventional video image reading systems for biosignals, according to the company. HoneyNaps also noted that this technology is setting a new standard for accuracy and transparency in the field.
"Like the AlphaGo case, which defeated humanity, this FDA approval is a very important event and a turning point in the field of sleep medicine in Korea. In the future, AI reading technology for biosignals is expected to play a very important role, similar to AI autonomous driving technology in cars. Furthermore, with the continuous improvement of biosignals AI reading technology, it will be possible to detect or predict some cardiovascular, neurological, and muscular diseases beyond the diagnosis of sleep disorders,” Ji Ho Choi, MD, PhD, head of the Center for Sleep Medicine at Soonchunhyang University Bucheon Hospital in Korea, said in a statement.1
For more context, sleep biosignals typically include electroencephalograms, electrooculograms, chin and leg electromyograms, electrocardiograms, respiratory airflow and effort, oxygen saturation, posture, and snoring. These biosignals are continuously monitored during sleep to assess sleep status and sleep disorder diagnosis by polysomnography (PSG). Reading the results from these assessments can take up to 3 to 4 hours; however, AI can reduce time and efforts in this area. Although the use of AI has potential, the complex and heterogeneous nature of biosignals makes attempting to improve these systems and achieving the same output as a human difficult.
READ MORE: Axsome Therapeutics Postpones Data Readout for Phase 3 SYMPHONY Trial
"The FDA has recently strengthened its review of AI-based medical devices, and we passed the review in 3 years by conducting clinical trials with 400 subjects including U.S. citizens directly, rather than through an agency, from the validation stage. This is an opportunity for us to further enhance our technology, such as adding diagnostic functions for cardiovascular and neuromuscular diseases, and to accelerate our expansion into the global market,” Tae Kyoung Ha, general representative director at HoneyNaps, said in a statement.1
At the recent 2023 SLEEP Annual Meeting, held June 3-7, in Indianapolis, Indiana, findings from a study on the new algorithm conducted by coauthor Ha and colleagues demonstrated better agreement by 0.068 compared another deep learning algorithm.2 The study selected and compared 30 PSG data with the results of another deep learning algorithm with 30 participants. The kappa value of deep learning algorithms showed 0.77 (95% CI, 0.759-0.78) compared with the new proposed algorithm, which was 0.838 (0.832, 0.843). Additionally, the proposed algorithm provided the classification results such as the type of sleep disorder and provided explanations of the event detection.
In June 2022, findings from a study published in Medicinia conducted byChoi, Ha, and colleagues verified the accuracy of automated sleep-stage scoring based on a deep learning algorithm.3 In the study, a total of 602 polysomnography datasets from participants (men, n = 397; women, n = 205) aged between 19 and 65 years (mean, 43.8; SD = 12.2) were included. The performance of the new proposed model was assessed based on kappa value and bootstrapped point-estimate of median percent agreement with a 95% bootstrap CI and R = 1000. The new model was trained using 482 datasets and validated using 48 datasets; however, only 72 random datasets were used for testing.
The new proposed model demonstrated good concordance rates with manual scoring for stages of wakefulness (94%), nonREM1 (83.9%), nonREM2 (89%), nonREM3 (92%), and REM (93%). The average kappa value was 0.84. Researchers observed a high overall agreement between the automated deep learning algorithm and manual scoring in stages of wakefulness (98%), nonREM1 (94%), nonREM2 (92%), nonREM2 (99%), and REM (98%) and total (96%) using the bootstrap method.
