An Ensemble Deep Learning Framework with Adaptive Windowing for EEGBased Brain Disorder Detection

Electroencephalogram (EEG)-based signal analysis serves as an important tool for evaluating brain dynamics and diagnosing neurological conditions, yet traditional interpretation remains difficult due to noise, variability, and the dependence on specialist expertise. Such challenges underscore the importance of developing automated diagnostic approaches that are more consistent and clinically dependable. This thesis introduces a multi-stage computational framework designed to improve the detection of abnormal EEG activity through refined feature selection, deep neural modelling, and ensemble decision strategies. The first stage of the framework applies a windowing method, which adaptively segments each EEG recording into its most informative temporal interval, based on Generalised Morse Wavelet (GMW) analysis, reducing irrelevant fluctuations and strengthening signal relevance. From these refined windows, five alternative signal representations max pooling, min pooling, average pooling, and two decimation schemes are generated to capture different temporal and amplitude characteristics. Each version is processed using an 18-layer CNN–LSTM model capable of learning combined spatial and temporal patterns in the EEG. To address the shortcomings of individual models, particularly their variable sensitivity (the ability of a model to correctly detect abnormal cases), a meta-classification layer is introduced. Three ensemble techniques: Support Vector Machines, Linear Regression, and Random Forest are used to integrate the outputs of the five CNN–LSTM branches. Among these, the Random Forest ensemble delivered the strongest performance, reaching 81.8% accuracy and a sensitivity of 92.86% on the TUH Abnormal EEG Corpus, demonstrating a substantial reduction in missed abnormal events compared with existing approaches. The results confirm that combining wavelet-based windowing, diverse feature extraction, and ensemble fusion can significantly enhance the robustness and clinical reliability of automated EEG analysis. The proposed framework provides a foundation for future systems capable of real-time monitoring and broader diagnostic functionality, with potential extensions including multi-class classification, transformer-based learning, explainability techniques, and cross-dataset validation.