Recognition of digits from brain data: A neural decoding approach

Imagined digit recognition is a significant research area within Brain–Computer Interface (BCI), involving non-invasive electroencephalography (EEG) for decoding brain activity. Imagined-digit decoding enhances communication for non-verbal individuals. This study proposes a Common Spatial Pattern-Convolutional Neural Network (CSP-CNN) framework for classifying imagined digits in both English and Bangla, enabling bilingual EEG-based BCI applications. The study analyzed the MindBigData dataset and newly developed English and Bangla imagined-digit datasets. The EEG signals were preprocessed, and CSP-extracted spatial features were fed to a CNN classifier. The proposed framework achieved a multiclass accuracy of 79.10% on the public dataset, outperforming prior results with an 8.86% improvement. The developed datasets were rigorously evaluated using subject-independent Leave-One-Subject-Out (LOSO) and GroupKFold cross-validation. Under the more stringent LOSO protocol, multiclass accuracies of 84.30% for the English dataset and 81.26% for the Bangla dataset were achieved. GroupKFold evaluation achieved accuracies of 98.70% and 95.44%, respectively, indicating effective subject-independent performance and cross-subject generalization. Additionally, a regional study revealed that the frontal lobe was the most significant area for imagined digit classification. Furthermore, explainable AI (XAI) analysis indicated that the frontal EEG channels, particularly the F4 channel, contributed prominently to the model’s decision-making process. The results demonstrate the effectiveness of the deep learning-based BCI system for bilingual imagined-digit recognition and highlight the potential of BCI systems for future passive and assistive applications.