Skip to Main content Skip to Navigation

Automatic classification of human sleep recordings combining artifact identification and relevant features selection

Abstract : This thesis engages in automatic analysis of human sleep. It mainly focuses on the development of an automatic system for classification of polysomnographic recordings, composed of three signals: EEG, EOG and EMG. This thesis proposes a complex classification system, which is capable to deal with various artifacts possibly present in the physiological signals and which uses the most relevant parameters computed from the analyzed signals.
The first part of this thesis presents a procedure to automatically identify eight common artifacts in 2-sec segments of the analyzed signals. Then, a strategy is applied in order to evaluate quality of the signals characterizing each 20-sec epoch of the recording.
In the second part of this thesis, an iterative feature selection method is proposed and applied on a large database of polysomnographic recordings, so as to select the most relevant parameters that will serve as inputs for the automatic classifier.
Then, as a result of the two first parts, a complex two-step sleep/wake stages automatic classification system is proposed. In a first step, an artifact detection system selects the artifact-free polysomnographic signals in the epoch to be scored. In the second step, the features selected as the most relevant are extracted from the artifact-free signals and classified using a neural network classifier chosen among a bank of four classifiers, which differs one from the others by the signals used. Thus, the final classification system allows classification using relevant features computed from artifact-free signals, without loosing many.
Document type :
Complete list of metadatas

Cited literature [65 references]  Display  Hide  Download
Contributor : Lukas Zoubek <>
Submitted on : Sunday, June 1, 2008 - 12:48:56 PM
Last modification on : Thursday, November 19, 2020 - 12:59:39 PM
Long-term archiving on: : Friday, May 28, 2010 - 6:08:00 PM


  • HAL Id : tel-00283929, version 1


Lukas Zoubek. Automatic classification of human sleep recordings combining artifact identification and relevant features selection. Automatic. Université Joseph-Fourier - Grenoble I, 2008. English. ⟨tel-00283929⟩



Record views


Files downloads