Feature Learning and Data Compression of Biosignals Using Convolutional Autoencoders for Sleep Apnea Detection
| Autor: | Rim Haidar, Bryn Jeffries, Irena Koprinska |
|---|---|
| Rok vydání: | 2019 |
| Předmět: |
0301 basic medicine
medicine.diagnostic_test business.industry Computer science Dimensionality reduction Apnea Sleep apnea Pattern recognition Polysomnography medicine.disease Convolutional neural network 03 medical and health sciences 030104 developmental biology 0302 clinical medicine medicine Artificial intelligence Respiratory system medicine.symptom business Hypopnea Feature learning 030217 neurology & neurosurgery Data compression |
| Zdroj: | Neural Information Processing ISBN: 9783030367077 ICONIP (1) |
| DOI: | 10.1007/978-3-030-36708-4_14 |
| Popis: | Sleep apnea is a medical condition that can be diagnosed from events in respiratory biosignals, and many supervised machine learning techniques can readily be applied to automate this task. Opportunities to use unsupervised techniques to identify different variants (phenotypes) of sleep apnea from such data require feature reduction techniques that capture essential details, ideally without requiring expert knowledge that might be biased to expected outcomes. Convolutional neural networks have shown successful results in detecting apnea events using raw respiratory data. In this work, we propose the use of convolutional autoencoders to compress and learn features from biosignals for sleep apnea analysis. We test reducing the original signals into latent space representations of a range of sizes, that are then used in conjunction with convolutional neural network classifiers. We compare their performance to down-sampling and principle component analysis feature reduction methods. We demonstrate that apnea and hypopnea events can be accurately detected even when the signals are reduced to a latent space representation 2–3% of the original size. We show that with a simple classifier architecture and very short training time, the reduced features from the convolutional autoencoders can give high performance in detecting apnea events. Our results are useful for the design of low-cost and portable devices for sleep monitoring. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|