RIYADH: A recent scientific research paper titled, “Adoption of Transformer Neural Network to Improve Diagnostic Performance of Oximetry for Obstructive Sleep Apnea” was published in the journal “Sensors” introduced a novel method for diagnosing obstructive sleep apnea, using a transformer neural network with learnable positional encoding.
The primary objective of the research is to amplify the diagnostic efficacy of oximetry for OSA and minimize expenses and time linked with traditional polysomnography (PSG), according to Saudi Press Agency.
Distinctively, the method offers annotations at one-second granularity, facilitating physicians in interpreting the model’s results.
The research was conducted by Ph.D. student Malak Almarshad, under the supervision of Prof. Ahmed BaHammam from the College of Medicine and Dr. Saiful Islam, Prof. Saad Al-Ahmadi, and Dr. Adel Soudani from the College of Computer Sciences, all at King Saud University.
In the study, the team experimented with different positional encoding designs as the initial layer of the model. The most promising results were derived from a learnable positional encoding built on an autoencoder with structural novelty.
Moreover, the model’s adaptability was assessed with various temporal resolutions, ranging from one to 360 seconds. Tests conducted on the public OSASUD dataset confirmed the superiority of the method over existing solutions, boasting an AUC of 0.89, an accuracy of 0.80, and an F1-score of 0.79.
Effects of Research Paper
The innovation presented in the research paper has the potential not only to refine OSA diagnosis but also to reduce associated costs.
By enhancing the reliability of OSA detection, this groundbreaking approach may drastically decrease undiagnosed cases, paving the way for improved health results for individuals grappling with this sleep disorder.