Classificação de Imagética Motora Para Controle de Órtese de Mão Com Eeg de Baixa Densidade E Aprendizado Profundo.

Assistive orthoses are essential tools in motor rehabilitation; however, their clinical adoption remains limited by high costs, low portability, and the complexity of the required instrumentation. This paper evaluates a non-invasive Brain-Computer Interface (BCI) pipeline based on motor imagery (MI) for controlling a 3D-printed hand orthosis. The control module processes the predicted class in real-time, converting signals into actuation commands while implementing a minimum confidence threshold to mitigate spurious activations. The system was evaluated using the PhysioNet EEG Motor Movement/Imagery public dataset, comparing five Deep Learning architectures: CNN1D, CNN2D, EEGNet, DeepConvNet, and ShallowFBCSPNet. Two acquisition scenarios were analyzed: a 64-channel configuration and a reduced 3-channel setup. In the 64-channel scenario, DeepConvNet and ShallowFBCSPNet outperformed the other models. In the reduced electrode setup, CNN2D and ShallowFBCSPNet achieved the best performance. These results demonstrate the feasibility of motor imagery classification using only three channels, significantly reducing instrumental complexity and favoring integration with low-channel hardware and physical actuation systems. Unlike traditional approaches that rely on high-density EEG and focus solely on signal analysis, this work integrates a streamlined decoding system into a functional 3D-printed orthosis, enhancing portability for activities of daily living.