Time Series Analysis for Sleep Apnea Detection Using Machine Learning

Abstract

This thesis investigates the applicability of various machine learning models on sleep apnea diagnosis using blood oxygen saturation measured with SpO2. Specifically, it examines the machine learning models k-NN, SVM, random forest and fully connected neural networks. The main focus is to determine whether this approach is a realistic and reliable diagnostic tool for sleep apnea detection. Additionally, the thesis aims to identify which machine learning model is best suited for this task. To evaluate this, the classification metrics precision, recall and F1-macro will be used. The utilized dataset contained 994 subjects, from which nine features were extracted after preprocessing. By evaluating the classification metrics of the developed models, the general conclusion is that fully connected neural networks are the most suitable for diagnosing sleep apnea, while k-NN models are the least suitable. In order to establish a proof of concept, a wearable device capable of measuring the oxygen saturation called EmotiBit was utilized to simulate the diagnosis.