EEG-BASED MOTOR IMAGERY PREDICTION FOR STROKE REHABILITATION USING QUINE-MCCLUSKEY-LATENT ATTENTION NETWORK WITH DANDELION OPTIMIZER

Stroke is a leading neurological disorder caused by the rupture or blockage of cerebral blood vessels, leading to brain cell damage and impaired motor functions. Electroencephalography (EEG) Motor Imagery (MI) Prediction utilizes EEG a non-invasive technique with high temporal resolution to record brain activity during imagined movements. By analysing these motor imagery signals, predictive models can assist in monitoring and enhancing motor recovery, making EEG a promising tool in stroke rehabilitation. To address the need for accurate and interpretable MI prediction, the study presented an EEG-based motor imagery prediction framework using a Quine-McCluskey-Latent Attention Network integrated with Dandelion Optimizer (QMc-LAN-DaO). Initially, raw EEG signals are acquired from Brain-Computer Interface Competition IV-2a (BCI Competition IV-2a) dataset, undergo pre-processing using the Cubature Kalman Filter (CKF) for noise reduction, and are then normalized within [8-24 Hz] frequency band to preserve motor-related rhythms essential for motor imagery analysis. For feature extraction, the Clifford Fourier Mellin Transform (CFMT) is employed to obtain robust, scale- and rotation-invariant spatiotemporal features essential for distinguishing motor intentions. Then prediction using Quine-McCluskey Latent Attention Network (QMc-LAN) combines the strengths of Quine-McCluskey Binary Classifier (QMcBC), known for its logical feature minimization, with the Latent Attention Network (LAN), which captures task-relevant hidden representations. This hybrid approach enhances prediction accuracy by efficiently selecting and focusing on the most discriminative EEG features for motor imagery prediction. To further optimize performance, the Dandelion Optimizer (DaO) fine-tunes the loss function by adaptively balancing feature noise and repeatability error. The proposed QM-LAN with DaO delivers high accuracy of 99.94% and interpretability, making it a valuable tool for MI-based stroke rehabilitation, neurofeedback, and BCI applications.