Inoue Ibuki, Zhao Xuyang, Komeiji Shuji, Yoshida Noboru, Sugano Hidenori, Nakajima Madoka, Tanaka Toshihisa
Annu Int Conf IEEE Eng Med Biol Soc. 2024 Jul;2024:1-4. doi: 10.1109/EMBC53108.2024.10782804.
Interictal epileptic discharge (IED) detection from electroencephalography (EEG) is an important but difficult step in the epilepsy diagnosis. To reduce the workload of doctors, some diagnostic auxiliary methods based on deep learning have been proposed. However, deep learning models often need more explainability, and even if they are explainable, their structure is usually complex. This paper presents a lightweight and explainable machine learning-based model named LightIED for detecting IEDs in EEG. The EEG data is first plotted as the image in the experiment and fed into the model for the IED detection task. Then, the Grad-CAM is used to analyze the output results and visualize the basis of inference. The detection accuracy of IEDs with the LightIED is almost equivalent to the current state-of-the-art (SoTA) model, Satelight, and higher than other Vision Transformer-based models. Moreover, the number of parameters is less than one-third compared to Satelight, the existing lightweight model. In addition, the visualizing results by Grad-CAM highlight the IEDs. Our results demonstrate that the proposed LightIED effectively detects IEDs with reasonable visualization.
从脑电图（EEG）中检测发作间期癫痫放电（IED）是癫痫诊断中的一个重要但困难的步骤。为了减轻医生的工作量，已经提出了一些基于深度学习的诊断辅助方法。然而，深度学习模型通常需要更强的可解释性，并且即使它们具有可解释性，其结构通常也很复杂。本文提出了一种基于机器学习的轻量级且可解释的模型，名为LightIED，用于检测脑电图中的发作间期癫痫放电。在实验中，脑电图数据首先被绘制为图像，然后输入到模型中进行发作间期癫痫放电检测任务。然后，使用Grad-CAM分析输出结果并可视化推理依据。LightIED对发作间期癫痫放电的检测准确率几乎与当前最先进的（SoTA）模型Satelight相当，且高于其他基于视觉Transformer的模型。此外，与现有的轻量级模型Satelight相比，其参数数量减少了三分之二以上。此外，Grad-CAM的可视化结果突出显示了发作间期癫痫放电。我们的结果表明，所提出的LightIED能够有效地检测发作间期癫痫放电，并具有合理的可视化效果。
