EEG-Based Assessment of Suicidality Risk: An Integrated Framework with Self-Adaptive Chaotic Cuckoo Search and AttentionBiSqueezeNet

This study proposes a novel integrated framework for electroencephalogram (EEG)-based suicide risk prediction designed to overcome key challenges, including signal noise, high dimensionality, and model interpretability. Our framework incorporates a robust preprocessing pipeline combining a modified Finite Impulse Response (FIR) filter and Fast Independent Component Analysis (Fast-ICA) for artifact removal. Then, a comprehensive suite of features spanning time, frequency, time-frequency, and functional connectivity domains is extracted, while dimensionality reduction is performed using a modified t-Distributed Stochastic Neighbor Embedding (t-SNE) technique. At the core of the framework is AttentionBiSqueezeNet, a hybrid Deep Learning (DL) architecture that combines Convolutional Neural Network (CNN) with an attention mechanism, a Bidirectional Long Short-Term Memory (Bi-LSTM), and a SqueezeNet backbone. To further enhance the capture of complex temporal dependencies, the activation functions of the Bi-LSTM are optimized using a Self-Adaptive Chaotic Cuckoo Search (SA-CCS) algorithm. The model was evaluated on a publicly available EEG dataset comprising 2,758 subjects and achieved an accuracy of 98.40%, precision of 98.87%, and sensitivity of 97.61%, outperforming all baseline methods. These results demonstrate that the proposed framework provides a robust, interpretable, and high-performing solution for early suicide risk detection using EEG data.