@masterthesis{RandeniKoralalage, type = {Bachelor Thesis}, author = {Randeni Koralalage, Lochana Chathuranatha}, title = {Implementation of the Brain Imaging Data Structure (BIDS) with EEG Data for Machine Learning-Based applications in Schizophrenia patients}, url = {http://nbn-resolving.de/urn:nbn:de:hbz:1383-opus4-20887}, school = {Hochschule Rhein-Waal}, pages = {66}, abstract = {Schizophrenia, a complex mental disorder characterized by delusions, hallucinations, and cognitive impairments, remains challenging to diagnose accurately due to the reliance on subjective clinical evaluations. This study aimed to develop a robust machine learning model for classifying schizophrenia patients using neurophysiological markers derived from Electroencephalography (EEG) data. The work explored multiple machine-learning architectures, including traditional Support Vector Machines (SVM) and deep learning models such as Long Short-Term Memory (LSTM) and CNN-LSTM hybrids. EEG data was collected from schizophrenia patients and healthy controls, with preprocessing conducted using the Brain Imaging Data Structure (BIDS) framework to ensure consistency and standardization. Given the varying quality and structure of the EEG recordings—such as missing channels and differing recording lengths—the study employed several preprocessing techniques, including channel interpolation, zero-padding, and limiting data to the first three minutes of resting-state recordings for standardization. Data imbalance was addressed through the application of the Synthetic Minority Over-sampling Technique (SMOTE) to ensure balanced representation of both schizophrenia and control groups during model training. The findings showed that the SVM model reached an accuracy of 88.8\% in patient classification, whereas the LSTM model, after addressing data imbalances through SMOTE, achieved an improved classification accuracy of 99.3\%. However, challenges arose when attempting to predict schizophrenia severity levels using MMN task data and a CNN-LSTM hybrid model. Despite efforts to standardize the number of epochs across subjects, the model achieved a more modest accuracy of 66.7\%. This lower accuracy was largely attributed to the limited and unbalanced dataset, as some subjects had fewer than the target number of epochs. Further challenges included BIDS conversion issues due to discrepancies in file naming conventions between .sdt and .set files, which required considerable time to resolve. The successful creation of a tailored BIDS conversion script ensured the data could be standardized and used effectively for analysis. In conclusion, while the classification models achieved high accuracy, particularly the LSTM-based approach, predicting schizophrenia severity levels remains a challenge due to the limitations of the available dataset. Future research will focus on gathering larger and more balanced datasets, refining preprocessing methods, and optimizing model architectures to improve prediction accuracy. Ultimately, the goal is to apply these models in clinical settings to facilitate early diagnosis and intervention, providing timely medical care to schizophrenia patients before their symptoms worsen.}, language = {en} }