Programs
- M. Tech. in Automotive Engineering -
- Fellowship In Interventional Pulmonology and Pulmonary Critical care - Fellowship
Publisher : EPSCICON 2018 - 4th International Conference on Power, Signals, Control and Computation
Campus : Coimbatore, Kochi
School : School of Engineering, School of Medicine
Department : Electrical and Electronics, Neurology
Year : 2018
Abstract : Sleep Apnea is a serious sleep disorder which is characterized by breath cessation during sleep. Repeated cessation can result in reduced oxygen supply to the brain and the rest of the body parts leading to poor sleep quality. Apnea cases are mostly unrecognized owing to the fact that the sleep studies are time and money consuming. Furthermore, the unavailability of high hospital standards and the lack of specialized persons, especially in rural areas have added up to the reason for the non-recognition of apnea. High accurate apnea detection systems were developed which made use of features extracted from the Rapid Eye Movement (REM) stage and thus demanded the persons complete 8 hour sleep. As an initial stage of disease screening, the patient may be uncomfortable to spend his complete night for testing. This paper has brought a system which can detect apnea as soon as the patient goes into his sleep rather than waiting him to enter to the REM stage. This requirement is accomplished by removing the stage dependency of features through Nuisance Attribute Projection (NAP), a used technique for removal of channel nuisances in speaker recognition system. Using the time and frequency domain features of heart rate and respiratory rate variability as input to Support Vector Machine (SVM) classifier, a baseline system of 50% absolute accurate system was developed. NAP was performed on the training features. A weight matrix was developed which brought together those features which belong to same classification label, but separated by the sleep stage and patient-specific variations. Individual systems were developed which removed the stage variations and patient-specific variations. An improvement in classification accuracy of 12.5% and 7.5% absolute were observed when stage and patient-specific variations were removed respectively. Ignoring the N1 stage features a further improvement of 20% absolute was observed. © 2018 IEEE.