Sleep apnea detection with multi-input convolutional neural networks

Sleep apnea is the most common disease among respiratory disorders and its degree is determined by the AHI index. In the detection of sleep apnea, many physiological parameters such as ECG, EOG, EMG, oxygen saturation (SAO2), air flow are examined. These parameters are recorded for at least 6 hours or more at night and a diagnosis is made by specialist physicians. Sleep apnea can have serious consequences, including death, when not detected. In such cases, Convolutional Neural Networks offer an effective alternative to traditional polysomnography methods. In this thesis study, sleep apnea detection was performed with "Multi-Input Convolutional Neural Networks (MI-CNN)". Using multi-input ESA architectures, appropriate ones of different data types (image, signal, etc. other) were used. In the detection of sleep apnea, DarkNet53 achieved high success with 99.25% scalogram visualization using apnea's respiratory parameters. Thus, an attempt was made to obtain more successful results in the detection of sleep apnea by using an approach that has not been tried before in this field in the literature.