A case study in time series classification using machine learning and deep learning

Abstract

This thesis focuses on the analysis and classification of electroencephalogram (EEG) data to discover patterns related to genetic predisposition to alcoholism. Classification of EEG signals plays a crucial role in the diagnosis of neurological conditions and the development of brain-computer interface (BCI) systems. The primary aim of this study is to develop and evaluate a classification pipeline based on time-frequency representations of EEG data, particularly spectrogram images, using both traditional machine learning and deep learning models. This study also seeks to compare the performance of various approaches and discuss the practical implications of classification accuracy in medical applications. The study used various pre-processing methods including denoising via Singular Spectrum Analysis (SSA) and spectral transformations such as Short Time Fourier Transform (STFT) to improve signal quality and representation. Both classical machine learning algorithms (e.g., Logistic Regression, Support Vector Machines and boosting algorithms) and deep learning architectures (e.g., CNN, LSTM and Bi-LSTM) were evaluated. Various optimization techniques have been applied throughout the modeling process to improve the performance of these models. In general, it is aimed to determine the most effective algorithms to achieve high classification accuracy on EEG data. Experimental findings show that deep learning models such as CNN and Bi-LSTM provide superior performance in EEG signal classification compared to traditional machine learning algorithms. These models effectively capture the complex temporal, spatial and nonlinear dynamics of EEG data that are critical for accurate interpretation. Among classical machine learning techniques, boosting algorithms have demonstrated competitive results, although they generally do not perform as well as deep learning approaches. Furthermore, the study highlights the impact of sampling rates on model accuracy and shows that optimal frequency resolution is important to maximize classification performance. This research contributes to the understanding of EEG signal processing and provides a comprehensive framework for EEG-based classification tasks with potential applications in the field of clinical diagnosis.