Study On Voice-assisted Diagnosis Of Patients With Obstructive Sleep Apne

Purpose: Obstructive sleep apnea(OSA)is a chronic sleep breathing disorder characterized by loud and irregular snoring,nighttime awakening or a sense of suffocation,daytime drowsiness,and sleep disturbance,which can even affect normal life in severe cases.Polysomnography is currently the gold standard for diagnosing obstructive sleep apnea,but due to the high cost,time-consuming examination process,and uneven medical conditions that limit diagnosis in many patients,there is a need for easier,faster,and more effective means of rapid screening.Methods: Based on experimental phonological theory,this research focuses on the vocal features of OSA patients.Adult males with complains of snoring were selected for the study.They were asked to produce sounds of test with Super Nasal-Oral Ratiometry System(SNORS)and Electroglottography(EGG)in both sitting and supine positions.Comparing the characteristics of speech,voice,and aerodynamic in patients with OSA.With thresholds of AHI=30,it classified the participants into sever OSA group and non-severe OSA group.Seven machine learning models are used for training.The characteristics of the best classification model were visualized and interpreted for analysis by SHAP.Results: Patients were divided into severe groups(n=88,AHI > 30 events/h)and nonsevere group(n=35,AHI≤30 events/h).The receiver operating characteristic(ROC)analysis,suggested that the most distinguishable acoustic characteristic was the bandwidth of the first formant(B1)for the vowel in the nasal n-syllable in the seated position,with an area under the curve(AUC): 0.648.The AUC of the harmonic-to-noise ratio(HNR)of the sustained vowel ü in the seated position is 0.753.The aerodynamic feature maximum mixed airflow speed of the vowel in the fricative sh[(?)] syllable in the seated position were able to differentiate the severe group from the non-sever group with the AUC: 0.620.The XGBoost model outperformed the other six machine learning model with an AUC of 0.876,an accuracy of 0.7,and a Brier score of 0.2.Conclusion: Phonological features base on speech signal processing can be an ideal way of screening test for patients with OSA.By choosing syllables or continuous speech streams,and combining aerodynamics,which can serve as an auxiliary reference for early OSA diagnosis.