Analysis of Online Sleep Apnea Data Streams for Mobile Platforms
Steffen Lien · NORA - Norwegian Open Research Archives · 2016
Sleep apnea is a sleeping disorder characterized by disruption in the natural breathing cycles during sleep. The disorder is common and is linked together with other disabilities and disorders that have serious ramifications to a person’s health. Diagnostic tools such as the polysomnography and portable monitoring devices can record physiological signals to determine a diagnose and severity rating of a patient. Before a diagnose can be determined, the recorded data needs to be manually analyzed by physicians which is a costly and time consuming process. With data mining, machine learning and event analysis the process can be redefined to be more efficient, automatic and require a fraction of the resources. Smart phone devices we have today are very capable of running advanced and complex software, and in some terms can be comparable to laptops. Data mining is a concept of classifying and predicting large amounts of data based on patterns and data analysis built using models. Esper is an open source Complex Event Processing engine and library component that use event series analysis to extract information and patterns in various types of data streams. In this thesis we design and implement four commonly used data mining methods: K-Nearest Neighbor, Support Vector Machine, Artificial Neural Network and Decision Tree. Along the data mining methods, we also design and implement two different detection methods in combination with statistical methods such as a Moving Average and Standard Deviation using the Event Processing Language utilized by Esper. We test both sets of classification methods on input data from three different database sources from PhysioNet. In addition, we analyze the performance utilization of the implementation with comparison with smart phone hardware to gather knowledge that can help develop a future automatic diagnosis application on smart phones. By streaming the data sets from a client to a server running Esper we achieve an accuracy of 90.89\\% using Decision Tree in a combination of four non-invasive signals. Overall accuracy results from all the data mining methods above 85\\% accuracy and close to 90\\% except for three signal combinations. Results from the two apnea detection methods score an accuracy of 93.26\\% and 93.13\\% for the Moving Average and Standard Deviation respectively. From the performance measurements we conclude that smart phones that is labeled as budget in terms of hardware and price is suitable to run data mining methods. The results overall show that both data mining methods and our designed detection method based on Esper combined with the performance metrics are suitable to implement as a standalone automatic diagnosis application on budget smart phones.