Back to search

Institutional Repository of Vilnius University

Automatic eeg sleep classification /

Abstract

dc:description

SVM algorithm was used to classify data in 3 scenarios: exact sleep stage, asleep or not, Sleep Stage 1 or Wake. The choice of splitting data into 30 and 60 second chunks, did not yield any significant results. Firstly, it was possible to achieve a 78.14% accuracy of classifying the exact sleep stage, however, it is important to not forget that the appearance of each stage in the dataset highly varied. The two-stage approach was looking promising with the initial results producing a 96.46% accuracy, but due to a high chance of favoritism (100% specificity, 0% sensitivity) it wasn’t reliable. After balancing the dataset to contain the same quantity of each stage, the accuracy dropped to 82.35%, however, sensitivity and specificity leveled out with the difference between them of 0.36%. Lastly, the highest focus was made on the two categories approach which was expected to yield good results being highly accurate and useful. After the initial phase, the accuracy was at 87.54%, sensitivity at 82.33%, and specificity at 90.38%. During the data balancing stage, results increased to 90.09%, 86.42%, and 93.75% respectively. Finally, after finetuning the regularization parameter the outcome is as follows: accuracy - 92.51%, sensitivity - 88.61%, specificity - 96.41%. Furthermore, 3 other algorithms were analysed in order to determine if they can improve current results. Firstly, kNN was tested - intial results showing 90.84% accuracy, however after finetuning the core parameters it was possible to increase it to 91.07%. Another simple algorithm AdaBoost performed the worst - 90.32% accuracy. Lastly, Neural Net has been run with multiple hidden layer configurations and varying regularization parameter, this has proved to be the best performing algorithm topping with 95.25% accuracy and only a 1.11% difference between sensitivity and specificity. By adding statistics of the previous epoch it was possible to increase accuracy of Neural Net and SVM to 95.69% and 93.16% respectively, while kNN didn't show improvements and AdaBoost had suffered due to memory limitations. Using data augmentation by adding more statistical features (skewness and kurtosis) instead, improved only the SVM algorithm sightly. Convolutional Neural Network algorithm performed not as well as the other algorithms: 85.9% accuracy with 91.74% sensitivity and 79.63% specificity. However, with differently shaped data, this method could be expected to perform better. Lastly, an application was developed to test the algorithms analyzed, which can be used with edf files with the same format as in this project. It enables the user to select their own data files, algorithm, and produces a graph of classification results.

Degree

thesis:*
Grantor dc:publisher
Institutional Repository of Vilnius University
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Čiulčinskas, Kristijonas,

Subjects

dc:subject × 1

Rights

dc:rights
Statement dc:rights
  • info:eu-repo/semantics/openAccess
Language dc:language
eng

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
Vilnius University
Base URL
epublications.vu.lt/oai
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Čiulčinskas, Kristijonas,. Automatic eeg sleep classification /. Institutional Repository of Vilnius University, 2025. https://repository.vu.lt/VU:ELABAETD239024332&prefLang=en_US