DECODING REACHING MOVEMENT DIRECTION: AN EEG STUDY WITH DEEP LEARNING AND EXPLANATION TECHNIQUES

Decoding the direction of arm reaching movements from non-invasive neural recordings represents a central step toward more intuitive and effective brain–computer interfaces (BCIs). Conventional machine-learning approaches have been widely applied to this problem; on the contrary, the use of deep learning approaches remains relatively underexplored. The aim of this study is to investigate the capability of a convolutional neural network (CNN) to decode reaching movement direction from the electroencephalographic signals (EEG) recorded during a delayed center-out reaching paradigm and to gain insights into the neural features driving the network’s decisions using explainability techniques. EEG were acquired from twenty healthy participants and processed using the EEGNet architecture to classify reaching targets under three levels of directional resolution: fine-grained (five targets), coarse-grained (three targets), and binary proximity-based (two targets) conditions. The explainability techniques integrated with the CNN (DeepLIFT and occlusion analyses) allowed a data-driven characterization of the spatio-temporal patterns underlying decoding performance. The proposed framework performed above chance in each classification problem, with an AUC (area-under-the-ROC curve) of 0.76±0.06, 0.82±0.07 and 0.88±0.04 for the five-, three-, and two-target classifications, respectively. Explainability analyses suggested that movement direction information was mainly represented during the movement preparation phase, with prominent contributions from centro-parietal and parieto-occipital regions. These findings were supported by complementary EEG analyses based on event-related spectral perturbations. Overall, our results suggest that CNN-based EEG decoding can reliably capture information about reaching movements and may offer valuable insights both for neuroscience investigation and future design of non-invasive BCI systems.