Klasifikacija EEG posnetkov na normalne in posnetke bolnikov s Parkinsonovo boleznijo z uporabo globokega učenja

Kramar, Sebastjan (avtor), Smrdel, Aleš (mentor)

Parkinsonova bolezen je ena najpogostejših nevrodegenerativnih bolezni, katere potek se lahko omili, če je odkrita dovolj zgodaj, kar pa je v praksi pogosto težko. Zato v okviru diplome raziskujemo možnost razvoja modela za klasifikacijo posnetkov EEG na posnetke oseb s Parkinsonovo boleznijo in posnetke zdravih oseb. Pri razvoju modela pa se poslužujemo globokega učenja. Pri študiji smo uporabili EEG posnetke iz množice 50 subjektov, od tega 25 bolnikov in 25 kontrolnih oseb. EEG posnetke sprva predprocesiramo, nato jih z uporabo nevronskih mrež klasificiramo. Modele smo osnovali na podlagi eno-dimenzionalnih konvolucijskih nevronskih mrež z mehanizmom pozornosti. Pri treniranju modela smo uporabili 84 % podatkovne množice, preostalih 16 % pa je bilo uporabljeno za preverjanje. Model uspešno klasificira segmente EEG posnetkov bolnikov s Parkinsonovo boleznijo z natančnostjo 0,839, kar je konkurenčno s trenutnimi modeli. Sam sistem smo nato še nadgradili s preprostim modelom uvrščanja posnetkov na podlagi njihovih segmentov v posnetke bolnikov s Parkinsonovo boleznijo ter posnetke zdravih oseb. Nadgradnja je uvrščala posnetke z natančnostjo 1.

• Jezik: Slovenski jezik
• Ključne besede: EEG, Parkinsonova bolezen, klasifikacija, nevronska mreža, konvolucija, pozornost
• Vrsta gradiva: Diplomsko delo/naloga
• Tipologija: 2.11 - Diplomsko delo
• Organizacija: FRI - Fakulteta za računalništvo in informatiko
• Leto izida: 2023
• PID: 20.500.12556/RUL-149197
• COBISS.SI-ID: 164843011
• Datum objave v RUL: 05.09.2023

Sekundarni jezik

• Jezik: Angleški jezik
• Naslov: Classification of EEG recordings into normal and recordings of patients with Parkinson’s disease using deep learning

Izvleček

Parkinson's disease is one of the most prevalent neurodegenerative diseases. Its progression can be slowed if detected early enough, which is often challenging in practice. Therefore, within the scope of this diploma work, we are investigating the capability of classifying EEG recordings of patients with Parkinson's disease and healthy individuals into their respective groups with the help of deep learning. In this study, we used EEG recordings from a set of 50 subjects, of which 25 are patients and 25 are healthy controls. We first preprocess the EEG recordings, then classify them using neural networks. Our models are based on one-dimensional convolutional neural networks with an attention mechanism. We trained the models on 84% of the dataset, with the remaining 16% used for validation. The model successfully classifies segments of EEG recordings of patients with Parkinson's disease with an accuracy of 0.839, which is competitive with current models. We further enhanced the system with a simple model for classifying the recordings based on their segments, into either recordings of patients with Parkinson's disease or recordings of healthy individuals. The enhanced model classified the recordings with a perfect accuracy score of 1.

• Ključne besede: EEG, Parkinson's disease, classification, neural network, convolution, attention