Comparison of Sleep State Classification Performance Using Random Forests, Hidden Markov Models, and Non-homogeneous Hidden Markov Models

• Author / Creator

  Mathieu Chalifour

• In this work, the CF00N polysomnograph data of 75 patients, with ranging severeties of Obstructive Sleep Apnea (OSA), is presented and analyzed in terms of sleep state classification. The pre-processing and cleaning of each polysomnograph recording were performed in R (R Core Team, 2019) using independent component analysis. Then three sets (Renyi Entropy, Moments of Discrete Wavelet coefficient (DWC), Moments of Non-EEG signals) of statistical features, 40 in total, were extracted from the time, frequency, and time-frequency domain of every 30second epoch using the C4M1 and non-EEG channels. A random forest feature selection was then performed selecting nine multivariate normal features to be used for comparison of classification performance against all three feature sets. Classification performance of sleep states for each patient’s epochs was analyzed first using random forests with a 10-fold cross validation and then a leave-one-patientout-cross-validation (LOPOCV). In the 10-fold cross validation, the mean (standard deviation) accuracy of the four feature sets was found to be 73.34%(0.07), for Renyi features, 75.26%(0.07), for DWC features, 60.64%(0.1), for non-EEG features, and76.85%(0.06) for the final features. In the LOPOCV, the mean performance measures of the random forest was found to decrease and the variance increase each feature set when the testing and training data did not share epochs from the samepatient. The mean accuracy results for the LOPOCV were 67.1%(7), for Renyi features, 74.5%(5.3), for DWC, 33.6%(6.3), for non-EEG features, and 72.2%(6.4) for the final features. The classification performance in the LOPOCV was furtheranalyzed using a 2-way MANOVA, which found no significant difference between the means of classification performance measures for the patient age and OSA group combinations.Then, using the Renyi entropy and final features of each patient’s epochs, hidden Markov models (HMMs) and non-homogeneous hidden Markov models (NHMMs) were fitted using 500 random starting points. The HMMs and NHMMs were fitted via the R library depmixs4 (Visser & Speekenbrink, 2010). The mean classification accuracy using the Renyi features was 67.1%(7.9) for HMM and 68.9%(7.8) for the NHMM and, using the final features, 65.3%(8.7) for HMM and 67.6%(9.1) for NHMM. Again, 2-way MANOVA was employed, with the only significant difference found between the mean performance measures of the age and OSA groups for the NHMM that used the final features. Furthermore, the comparison between HMMs and NHMMs that used the Renyi features found that on average the NHMM accuracy was between 0.5% and 3.1% higher than HMM. When the HMMs and NHMMs used the final features, the NHMM accuracy was on average between 0.4% and 4.2% higher than HMM. A comparison of classification accuracy for the random forest LOPOCV versus the HMM and NHMM found that, when using the Renyi features, the HMM and NHMM typically performed better than the random forest and, when using the final features the random forest performed better than the HMM and NHMM. The analysis of this thesis demonstrates that although the random forest, HMM and NHMM can be successful at classifying sleep states, the HMM and NHMM are superior,since the random forest lacks a model for the dependence structure between sleep states. The modelling of sleep state transitions captured by the HMM and NHMM can provide sleep experts with further insight to the underlying dynamics of sleep.

• Subjects / Keywords

  • Sleep State classification
  • HMM
  • Non-homogeneous hidden Markov model
  • Hidden Markov Model
  • Random Forest
  • NHMM

• Graduation date

  Fall 2020

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-41bp-tp04

• License

  Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.