Explainable Cognitive Task Classification in Pediatric EEG Using CPCC-Based Functional Connectivity Images

Objective and interpretable monitoring of cognitive states in children and adolescents is important for educational and mental–health applications, but existing EEG-based approaches often rely on hand-crafted features or opaque deep models. We propose a functional-connectivity framework that converts Complex Pearson Correlation Coefficient (CPCC) matrices into images and classifies them with a convolutional neural network (CNN). Using the HBN-EEG dataset, we analyse 3,200 EEG sessions from 598 participants (5–21 years) performing ten tasks including resting state, naturalistic movie watching, visual detection, and sequential learning. CPCC is computed in six frequency bands to obtain two complementary connectivity indices, absCPCC and imCPCC, which are fed to an ImageNet-pretrained ResNet-18 under a subject-independent split. Our CPCC-based model substantially outperforms four raw-EEG baselines and single-index CPCC variants, achieving 67.8% accuracy and 59.5% macro F1 on the test set. Grad-CAM applied to connectivity images highlights task-specific networks consistent with developmental neurophysiology, such as posterior alpha at rest and fronto–parietal theta/beta during visual and sequential tasks. These results show that CPCC-based connectivity imaging with explainable CNNs is a promising approach for decoding pediatric cognitive states and identifying network-level markers relevant to developmental neuroscience and neurodevelopmental disorders.