PCCNN: A CNN classification model integrating EEG time-frequency features for stroke classification

• Published in: Cognitive robotics Vol. 5; pp. 211 - 225
• Main Authors: Wang, Teng, Li, Fenglian, Yang, Jia, Jia, Wenhui, Hu, Fengyun
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 2025; KeAi Communications Co. Ltd
• Subjects: Convolutional neural network; Discrete wavelet transform; EEG; Empirical mode decomposition; Stroke; Stroke; Empirical mode decomposition; Discrete wavelet transform; EEG; Convolutional neural network
• ISSN: 2667-2413; 2667-2413
• DOI: 10.1016/j.cogr.2025.05.002

Stroke classification is crucial for timely diagnosis and treatment, as it helps differentiate between hemorrhagic and ischemic strokes, which require distinct clinical interventions. This paper proposes a stroke classification method using multi-channel electroencephalography (EEG) data. Unlike single-channel data or simple multi-channel concatenation, our method processes EEG data as a channel matrix, significantly improving classification performance. We employ two complementary feature extraction techniques: discrete wavelet transform (DWT) and empirical mode decomposition (EMD). DWT extracts multi-scale wavelet coefficients from stroke-related frequency bands, while EMD decomposes EEG signals into intrinsic mode functions (IMFs), representing narrowband oscillation components. To enhance feature quality, we propose a hybrid selection method that integrates four metrics—information entropy, power spectral density (PSD) distance, statistical significance, and maximum information coefficient (MIC)—to comprehensively evaluate IMFs. This method accounts for both the intrinsic information content of EEG signals and the inter-class differences between hemorrhagic and ischemic stroke subjects. Furthermore, this paper designs a pyramid cascade convolutional neural network (PCCNN) model with multi-branch independent learning and hierarchical fusion. Each DWT and EMD feature is processed by an independent one-dimensional convolutional neural networks (1D-CNN) branch for targeted extraction. A pyramid fusion mechanism integrates branch outputs into a fused feature vector, enabling the feature interaction through a top-level fusion CNN. Experimental results demonstrate that the proposed method, which integrates channel matrix processing, high-quality DWT and EMD feature selection, and multi-branch feature fusion, significantly outperforms single-feature methods. The fusion feature achieves a classification accuracy of 99.48 %, effectively distinguishing EEG data of hemorrhagic and ischemic stroke.