Khabti Joharah, AlAhmadi Saad, Soudani Adel
College of Computer and Information Sciences (CCIS), King Saud University, Riyadh 11543, Saudi Arabia.
King Salman Center for Disability Research, Riyadh 11614, Saudi Arabia.
Sensors (Basel). 2024 Dec 19;24(24):8127. doi: 10.3390/s24248127.
One of the most promising applications for electroencephalogram (EEG)-based brain-computer interfaces (BCIs) is motor rehabilitation through motor imagery (MI) tasks. However, current MI training requires physical attendance, while remote MI training can be applied anywhere, facilitating flexible rehabilitation. Providing remote MI training raises challenges to ensuring an accurate recognition of MI tasks by healthcare providers, in addition to managing computation and communication costs. The MI tasks are recognized through EEG signal processing and classification, which can drain sensor energy due to the complexity of the data and the presence of redundant information, often influenced by subject-dependent factors. To address these challenges, we propose in this paper a multi-subject transfer-learning approach for an efficient MI training framework in remote rehabilitation within an IoT environment. For efficient implementation, we propose an IoT architecture that includes cloud/edge computing as a solution to enhance the system's efficiency and reduce the use of network resources. Furthermore, deep-learning classification with and without channel selection is applied in the cloud, while multi-subject transfer-learning classification is utilized at the edge node. Various transfer-learning strategies, including different epochs, freezing layers, and data divisions, were employed to improve accuracy and efficiency. To validate this framework, we used the BCI IV 2a dataset, focusing on subjects 7, 8, and 9 as targets. The results demonstrated that our approach significantly enhanced the average accuracy in both multi-subject and single-subject transfer-learning classification. In three-subject transfer-learning classification, the FCNNA model achieved up to 79.77% accuracy without channel selection and 76.90% with channel selection. For two-subject and single-subject transfer learning, the application of transfer learning improved the average accuracy by up to 6.55% and 12.19%, respectively, compared to classification without transfer learning. This framework offers a promising solution for remote MI rehabilitation, providing both accurate task recognition and efficient resource usage.
基于脑电图（EEG）的脑机接口（BCI）最有前景的应用之一是通过运动想象（MI）任务进行运动康复。然而，当前的MI训练需要患者亲自到场，而远程MI训练可以在任何地方进行，便于灵活康复。提供远程MI训练除了要管理计算和通信成本外，还对确保医疗保健人员准确识别MI任务提出了挑战。MI任务是通过EEG信号处理和分类来识别的，由于数据的复杂性和冗余信息的存在，这可能会消耗传感器能量，而这些往往受到个体相关因素的影响。为了应对这些挑战，我们在本文中提出了一种多主体迁移学习方法，用于物联网环境下远程康复中的高效MI训练框架。为了高效实现，我们提出了一种物联网架构，该架构包括云/边缘计算，作为提高系统效率和减少网络资源使用的解决方案。此外，在云端应用了带通道选择和不带通道选择的深度学习分类，而在边缘节点则采用了多主体迁移学习分类。采用了各种迁移学习策略，包括不同的轮次、冻结层和数据划分，以提高准确性和效率。为了验证这个框架，我们使用了BCI IV 2a数据集，将受试者7、8和9作为目标。结果表明，我们的方法在多主体和单主体迁移学习分类中均显著提高了平均准确率。在三主体迁移学习分类中，FCNNA模型在不进行通道选择时准确率高达79.77%，在进行通道选择时为76.90%。对于双主体和单主体迁移学习，与不使用迁移学习的分类相比，迁移学习的应用分别将平均准确率提高了6.55%和12.19%。这个框架为远程MI康复提供了一个很有前景的解决方案，既能提供准确的任务识别，又能高效使用资源。
