Torma Szabolcs, Szegletes Luca
Department of Automation and Applied Informatics, Faculty of Electrical Engineering and Informatics, Budapest University of Technology and Economics, Műegyetem rkp. 3., H-1111 Budapest, Hungary.
J Neural Eng. 2025 Jan 7;22(1). doi: 10.1088/1741-2552/ada0e4.
. The development of deep learning models for electroencephalography (EEG) signal processing is often constrained by the limited availability of high-quality data. Data augmentation techniques are among the solutions to overcome these challenges, and deep neural generative models, with their data synthesis capabilities, are potential candidates.. The current work investigates enhanced diffusion probabilistic models (DPM) and sampling methods for brain signal generation and data augmentation. We used implicit sampling and progressive distillation to shorten the inference and subsequently analyzed the effects of these methods on the generated data. To assess the feasibility of our method, four classification models were trained and evaluated in an inter-subject setting on datasets augmented with synthetic signals.Our analysis of generative metrics and statistical evaluations, including subject- and group-level tests, showed that our DPMs could generate visual evoked potentials and motor imagery signals. Distilled, single-step DPMs were trained on two publicly available datasets and were used to synthesize relatively high-quality EEG samples. The performance of the classifiers was improved by the application of the synthesized signals. The present work demonstrates that DPMs are capable of augmenting data with high fidelity and improving the diversity of EEG signals. Although samples can be generated in a single step, there is a significant trade-off between the data quality and sampling steps.The findings and results of this study demonstrate the promising capability of diffusion models for EEG synthesis, which marks progress toward an efficient and generalizable augmentation method for various EEG decoding tasks.
用于脑电图（EEG）信号处理的深度学习模型的发展常常受到高质量数据有限可用性的限制。数据增强技术是克服这些挑战的解决方案之一，而具有数据合成能力的深度神经生成模型是潜在的候选方法。当前的工作研究了用于脑信号生成和数据增强的增强扩散概率模型（DPM）和采样方法。我们使用隐式采样和渐进蒸馏来缩短推理过程，随后分析了这些方法对生成数据的影响。为了评估我们方法的可行性，在一个跨受试者的设置中，在使用合成信号增强的数据集上训练和评估了四个分类模型。我们对生成指标和统计评估的分析，包括受试者和组水平的测试，表明我们的DPM可以生成视觉诱发电位和运动想象信号。经过蒸馏的单步DPM在两个公开可用的数据集上进行了训练，并用于合成相对高质量的EEG样本。通过应用合成信号提高了分类器的性能。目前的工作表明，DPM能够以高保真度增强数据并提高EEG信号的多样性。尽管可以一步生成样本，但在数据质量和采样步骤之间存在显著的权衡。本研究的发现和结果证明了扩散模型在EEG合成方面具有广阔前景的能力，这标志着朝着一种适用于各种EEG解码任务的高效且通用的增强方法迈进了一步。
