A fine-grained diagnostic method for sleep apnea based on target detection network with SpO2 and nasal airflow

The extant automatic diagnostic technologies for sleep apnea primarily focus on the accuracy of apneahypopnea index (AHI) estimation and severity classification, ignoring the accurate annotation for abnormal respiratory events. Motivated by target detection principle, we proposed a fine-grained diagnostic method called apnea-hypopnea event detector (AHE-Detector) in this paper. Through fusing time-frequency features of pulse oxygen saturation and nasal airflow signal, our model can accurately identify abnormal respiratory events on the continuous time axis. The model specifically contains three modules: 1) a feature fusion module; 2) a multi-scale feature extraction module and 3) a predictive module for locating and distinguishing respiratory events. We conduct comprehensive validation of our method on two public datasets. The experimental results show that our method achieves considerable performance on three tasks with different complexity, including the task to detect abnormal events, the task to identify hypopnea and apnea, as well as the task to further distinguish apnea event subtypes. Based on the detection results, we realize accurate assessment of two clinical indicators, AHI and apnea hypopnea event duration index. To the best of our knowledge, our method is the first to enable accurate and simultaneous estimation of two clinical indicators associated with sleep apnea. The comparison results further prove that our method excels the extant methods in the aspects of event detection, clinical index estimation and severity classification.