Spatiotemporal sLORETA for Interpretable EEG Source Imaging and Binary Motor Classification

Abstract

Reliable reconstruction of cortical activity from electroencephalography (EEG) is fundamental to brain–computer interfaces, motor assessment, and neurorehabilitation. However, the ill-posed nature of the EEG inverse problem, combined with noise and nonstationary artifacts, often produces spatially diffuse source estimates that limit interpretability and downstream learning performance. In this work, we present a spatiotemporal extension of sLORETA that incorporates adaptive temporal regularization and anatomically informed spatial smoothness to improve source focality and stability during dynamic motor tasks. The proposed approach is evaluated on a public reach-and-grasp EEG dataset comprising 15 subjects performing palmar and lateral grasp movements. Quantitative analysis demonstrates consistent improvements over classical sLORETA, reducing spatial entropy from 0.96 to 0.89, spatial dispersion from 78.3 mm to 73.4 mm, and half-mass radius from 69.5 mm to 62.1 mm. Subject-level analyses confirm robust gains across all participants. To assess functional relevance for learning-based decoding, binary classification is performed using sourcespace features. The proposed method achieves mean accuracies of 91.6% ± 1.6% for lateral versus rest and 89.8% ± 2.1% for palmar versus rest, with Cohen’s κ exceeding 0.79 in both cases. These results demonstrate that spatiotemporal source modeling enhances both neurophysiological interpretability and discriminative power, providing a strong foundation for deep learning–based motor decoding and rehabilitation-oriented EEG applications.