Riyadh, Oct 1, 2023, A recent scientific article titled "Enhancing Obstructive Sleep Apnea Diagnosis with Transformer Neural Network" appeared in the journal "Sensors" on September 15, 2023.
This groundbreaking research was conducted by Malak Almarshad, a PhD student, under the expert guidance of Prof. Ahmed BaHammam from the College of Medicine, and Prof. Saad Al-Ahmadi, Dr. Saiful Islam, and Dr. Adel Soudani from the College of Computer Sciences, all affiliated with King Saud University.
The study introduces an innovative approach to diagnosing obstructive sleep apnea (OSA) by employing a transformer neural network with learnable positional encoding. The primary goal is to enhance the diagnostic accuracy of oximetry for OSA while minimizing the time and costs associated with traditional polysomnography (PSG). Notably, this method provides annotations at a one-second granularity, aiding physicians in interpreting the model's findings.
During the study, the research team experimented with various positional encoding designs as the initial layer of the model. The most promising results were obtained using a learnable positional encoding based on a novel structural autoencoder. Furthermore, the model's adaptability was assessed at different temporal resolutions, ranging from 1 to 360 seconds. Tests conducted on the publicly available OSASUD dataset confirmed the method's superiority over existing solutions, achieving an impressive AUC of 0.89, an accuracy rate of 0.80, and an F1-score of 0.79.
The innovation presented in this paper not only holds the potential to refine the diagnosis of OSA but also to reduce associated healthcare costs. By improving the reliability of OSA detection, this pioneering approach may significantly decrease the number of undiagnosed cases, ultimately leading to better health outcomes for individuals dealing with this sleep disorder.
