Multimode OSAHS Detection Method Based On Deep Learning

Obstructive Sleep Apnea Hypopnea Syndrome（OSAHS）is a physiological activity in-volving multiple systems,and there is correlation between physiological signals of different modes.Compared with single mode,multi-mode data contains more information.It also re-veals the potential correlations between modes.However,the existing multimodal studies lack effective processing methods to explore the potential correlation between the various modes,and are unable to adapt to the individual signal differences and channel differences,thus failing to take into account the accuracy and stability of sleep detection.Therefore,this paper is based on deep learning multi-mode OSAHS detection algorithm to carry out research,and the main contents are as follows:（1）Processing data.For outlier signals,linear interpolation is used to eliminate the zero-level artifacts caused by sensor disconnection.For power frequency interference and myoelec-tric interference,wavelet processing technology is used to remove noise.For baseline drift,convolution moving average filter is used to filter.Divide the data set.（2）The Convolutional Neural Networks（CNN）and the stacked Light Gradient Boosting Machine（LGBM）classifier are proposed to detect OSAHS.This method can better explore and utilize the potential correlation between different modes.Data sets were divided across subjects to resist the influence of feature distribution changes.CNN is used to extract the features of the middle layer to realize deep feature interaction and feature fusion.Combined with stacked LGBM classifier,the representativeness of features is improved to obtain deeper discriminant information for OSAHS detection.The proposed model is applied to apnea-ECG database,and the results show that the layered feature fusion is better than the decision level fusion and single signal detection,which indicates that the proposed model can effectively extract potential correlation information between Sp O₂and ECG.（3）ResNeSt-LGBM network learning potential characterization of OSAHS detection be-tween two different modes Sp O₂and ECG is proposed.In view of the individual differences existing in OSAHS,the cross-quilt method was used to solve the individual differences.In order to solve the problem of channel difference,splitter Attention module is used to extend channel Attention to feature graph group to capture the key time-frequency characteristics of Sp O₂and ECG signals.The global pooling layer is used to keep channel dimensions separate,and at the same time,feature maps are aggregated on spatial dimensions to solve the problem of channel differences.LGBM classifier is used to improve the ability of feature representation.The results show that ResNeSt-LGBM network can effectively interact multimodal features and learn the potential correlation between Sp O₂and ECG signals to achieve multimodal fea-ture fusion.