Obstructive Sleep Apnea Detection Using Data Mining Technique

Obstructive apnea (OSA) may be a common, but severely under-diagnosed disorder that affects the natural breathing cycle during roll in the hay periods of reduced respiration or no airflow in the least. As a first step towards the goal, we explore whether a limited subset of the physiological signs used in traditional OSA diagnosis, together with automatic classification, may be used to detect apnea occurrences in this study. We examine the effects of five data mining algorithms in classifying epochs of data from the PhysioNet Apnea-ECG and MIT-BIH datasets as interrupted or normal breathing. This research focuses on respiratory signals from the nose, abdomen, and chest, as well as oxygen saturation. We calculate the accuracy, sensitivity, specificity, and Kappa statistics of classification with data mining algorithms for any combination of these signals. With a collection of respiration data from both the chest and nose as input data, we reach an accuracy of 96.6 percent for Apnea-ECG, and an accuracy of more than 90 percent for other signal combinations. Surprisingly, these good results may also be obtained using the basic KNN approach. Because of noise, lesser size, and some class imbalance, the findings for MIT-BIH are lower.