A Hybrid Deep Learning Ensemble Framework for EEG-Based Depression Detection

EEG, due to its noninvasive nature, low cost, and high temporal resolution, is an ideal modality for monitoring brain electrical activity. EEG patterns in depressed patients exhibit distinct characteristics, such as frontal alpha asymmetry and increased theta power in the frontal central regions. Earlier EEG-based studies relied heavily on handcrafted feature extraction methods, including entropy measures and frequency band power analysis. However, the emergence of deep learning has enabled end-to-end learning of hierarchical and complex representations directly from raw EEG data. The main objective of this research is to propose a hybrid deep learning ensemble framework that leverages the complementary strengths of CNN, ResNet-۱D, and TCN architectures to analyze raw EEG signals without manual feature engineering. The proposed method aims to maximize depression classification performance while improving model generalization. Unlike many previous studies that primarily focus on improving classification accuracy using a single deep learning architecture, this work emphasizes robustness and generalization. By integrating complementary deep learning models within a stacked ensemble framework, the proposed approach aims to reduce performance variance and improve reliability, which is particularly important in clinical screening applications. The key innovation of this work lies in combining CNN, ResNet-۱D, and TCN architectures within a unified ensemble framework. Due to its dilated convolutional structure, TCN can effectively model long-range temporal dependencies and often outperforms recurrent architectures such as LSTM in time-series analysis. This three-model ensemble enables simultaneous modeling of local patterns (CNN), deep hierarchical features (ResNet), and long-term temporal dependencies (TCN), resulting in a robust depression detection framework. Unlike prior studies that primarily focus on improving classification accuracy through increasingly complex single architectures, this work emphasizes robustness and generalization by integrating complementary deep learning models. The novelty of the proposed framework lies not in proposing a new neural architecture, but in the systematic combination of CNN, ResNet-۱D, and TCN within a stacked ensemble to reduce variance and enhance reliability in EEG-based depression screening.