Sequential feature selection methods for Parkinsonian human sleep analysis

• Published in: 2009 17th Mediterranean Conference on Control and Automation pp. 1468 - 1473
• Main Authors: Fairley, J., Georgoulas, G., Vachtsevanos, G.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.06.2009
• Subjects: Automatic control; Automation; Biology computing; component; Data mining; Diseases; Electrodes; Electroencephalography; Feature Subset Selection; Human Sleep; Humans; Parkinson's Disease; Pathology; Sleep
• ISBN: 1424446848; 9781424446841
• DOI: 10.1109/MED.2009.5164754

This paper focuses on the selection of quantitative features from the polysomnogram (psg) to enhance automated/computerized approaches in analyzing the human sleep cycle for pathology identification. Validation of the utilization of the psg as a metric for pathology identification is cited by Bliwise et al. [1]. The pathological case investigated in this study was pre-Parkinsonian disease. This case is of interest because, to the author's knowledge, studies to investigate quantitative features to describe the human sleep cycles of pre- Parkinsonian disease patients, to date of this writing, have not been published. In this study a total of 67 quantitative features were investigated in the characterization of the human sleep cycles for adult pre- Parkinsonian patients and normal subject psgs. Adult normal human sleep may contain time durations of at least 6.5 hours [2]. According to international sleep scoring standards, a minimum of four biological channels are required in a psg recording [3]. Computation of all 67 features over such a large data set for multiple patients/subjects poses computational efficiency issues especially when attempts are made to incorporate these automated methods in a clinical environment. To alleviate the computational burden of processing all 67 features, in this study, intelligent feature selection techniques were incorporated to establish optimal feature sub-sets that best characterized the human sleep cycles. Feature sub-sets for characterization of psg data for adult pre- Parkinsonian patients and normal control subjects were obtained using the sequential forward and backward feature selection algorithms and k-Nearest Neighbor (k-NN) classification. An investigation of these feature selection techniques toward the characterization of adult pre-Parkinsonian patients and normal control subject psgs are provided in this paper.