Pengembangan Model LightGBM untuk Klasifikasi IED dari Sinyal EEG

Purnama, Naufal (2024) Pengembangan Model LightGBM untuk Klasifikasi IED dari Sinyal EEG. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025201195-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Epilepsi merupakan gangguan neurologis yang ditandai oleh kejang tidak terduga yang dapat berulang. Ketidakpastian serangan ini dapat mempengaruhi kegiatan sehari-hari, hubungan interpersonal, dan kesejahteraan emosional. Tenaga medis menggunakan berbagai teknik untuk memprediksi serangan, salah satunya menginterpretasi elektroensefalografi (EEG) secara manual. Dengan kemajuan di bidang pembelajaran mesin, prosedur ini telah dicoba diotomatisasi. Fürbass et al. telah mengembangkan model deep learning untuk mendeteksi interictal epileptiform discharge (IED). Riset tersebut memanfaatkan dataset rekaman EEG dari Aarhus University Hospital. Algoritma yang dikembangkan mencatat sensitivitas 89%, spesifisitas 70%, dan akurasi 80%.
Tugas akhir ini bertujuan untuk menggunakan dataset rekaman EEG Aarhus University Hospital untuk mengembangkan model LightGBM. Model ini akan dilatih untuk mengenali pola Interictal Epileptiform Discharge (IED) epileptik dan nonepileptik. Terdapat dua model LightGBM yang dilatih. Model baseline dilatih pada time series EEG murni yang di-flatten. Sementara itu, model kedua akan dilatih pada data EEG yang telah diekstraksi fitur. Ekstraksi fitur meliputi Rekaman EEG yang telah dibersihkan akan diekstraksi fitur. Fitur yang diekstraksi mencakup fitur daya dari subband frekuensi yang didapatkan dari metode Welch. Selain itu, dihitung fitur statistikal dari koefisien yang diperoleh dari Discrete Wavelet Transform (DWT) menggunakan wavelet Daubechies-4 dengan tingkat dekomposisi lima. Model baseline yang dilatih mencatat akurasi 70%, sensitivitas 78%, spesifisitas 62%, presisi 68,39%, dan skor-F1 72,18%. Sementara itu, model ekstraksi fitur mencatat akurasi 84%, sensitivitas 82%, spesifisitas 86%, presisi 86,50%, dan skor F1 83,41%. Hasil ini menunjukkan pengaruh ekstraksi fitur terhadap hasil akhir pelatihan model.
========================================================================================================================
Epilepsy is a neurological disorder characterized by recurrent, unpredictable seizures. The unpredictability of these seizures can influence daily activities, relationships, and emotional well-being. Medical professionals utilize a variety of techniques to predict seizures, an example being the manual interpretation of electroencephalography (EEG) recordings. With the advancements made in machine learning, research has been made to develop an automatic epilepsy detector. Fürbass et al. have developed a automatic IED-detection algorithm using deep learning. The described algorithm had a sensitivity of 89%, a specificity of 70%, and an overall accuracy of 80%.
This final project aims to advance seizure prediction by developing a LightGBM model leveraging the Aarhus University Hospital dataset of EEG recordings. The model will be trained to recognize patterns within epileptic and non-epileptic interictal epileptiform discharges (IED). Two models were trained. A baseline model was trained on the flattened multichannel EEG time series. The second model was trained on features extracted from the raw EEG data. These features include the power measure of frequency subbands of interest (alpha, beta, delta, and gamma) obtained from the Welch Method. The other features were statistical features from the coefficients obtained from discrete wavelet transform. The baseline model achieved 70% accuracy, 78% sensitivity, 62% specificity, 68,39% precision, and an F1-score of 72,18%. On the other hand, the second model trained on extracted features achieved 84% accuracy, 82% sensitivity, 86% specificity, 86,50% precision, and an F1-score of 83,41%. These results show the significant impact of feature engineering and extraction on training LightGBM to classify epileptic IEDs.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	elektroensefalografi, electroencephalography, interictal epileptiform discharge, lightgbm
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Naufal Adli Purnama
Date Deposited:	06 Aug 2024 00:42
Last Modified:	06 Aug 2024 00:42
URI:	http://repository.its.ac.id/id/eprint/111524

Actions (login required)

View Item