Abstract
Obstructive sleep apnea (OSA) is a common sleep disease which may cause many serious health problems, therefore timely diagnosis and treatment could bring important help for patients. The current research on automatic detection through bioelectric signal mainly rely on blood oxygen level (SpO2), ECG signal and airflow, with few EEG based solution proposed. More importantly, most researches on OSA detection did not realize the difference between OSA detection and the general sleep staging classification. Some of them just simply transfer sleep staging methods to OSA detection, and an EEG segment was processed as a whole, ignoring the ambiguity in feature space that some EEG frames may contain both normal fragments and OSA fragments. In this paper, we propose a framework, EEG multi-instance learning network (EEG-MIL) for automatic OSA detection based on EEG signals to alleviate this ambiguity. EEG-MIL is composed of subframe multi-resolution convolution extractor (S-MRCNN) and MIL mapping function, which could extract features from sub-frames and mine the interactive relationship between different instances (sub-frames) and bags (frame) to further distinguish OSA event fragment. Meanwhile, in order to meet the clinic needs, we define instance-level task and bag-level task in OSA events detection, and redefine the evaluation criteria to evaluate the effect of our model more comprehensively. Then we verify the performance of our framework in two public datasets and the private dataset, and provide detailed ablation experiments. We validate our method via 5-fold subject independent cross validation approach. Our model obtains 2–8.6% performance gain compared to other works and achieves the new state-of-the-arts.