Interpretable seizure detection with signal temporal logic neural network
Interpretable seizure detection with signal temporal logic neural network
复制标题
DOI：
10.1016/j.bspc.2022.103998
复制
发表时间：
2022-09
期刊：
影响因子：
--
通讯作者：
Ruixuan Yan;A. Julius
中科院分区：
文献类型：
--
作者：
Ruixuan Yan;A. Julius
在这项工作中，我们开发了一种新的神经符号模型，用于使用多视图数据表示的自动癫痫发作检测。首先，使用多视图特征提取技术提取光谱和线长度特征。接下来，一个信号时序逻辑神经网络（石），结合神经网络和时序逻辑的好处，构建癫痫发作和非癫痫发作数据进行分类。STONE是以这样一种方式设计的，即每个神经元都有一个符号表示，对应于加权信号时序逻辑（wSTL）公式中的一个分量。与传统的STL推理算法相比，STONE是端到端可微的，因此可以通过反向传播来完成学习。此外，STONE提高了癫痫发作检测模型的可解释性，因为STONE的结果是一个可解释和人类可读的wSTL公式。重要的是，wSTL公式揭示了癫痫发作背后的推理，作为EEG信号演变的描述。STONE在两个流行的EEG数据库上进行了测试，并证明与现有的最先进的模型相比，在准确性，灵敏度和特异性方面实现了有前途的检测性能。此外，STONE可以提供人类可读的公式作为癫痫发作特征的描述，并且该公式也是可视化的，以便于分类器的解释，这是现有癫痫发作检测方法中缺失的属性。
In this work, we develop a novel neuro-symbolic model for automated seizure detection using multi-views of data representation. Firstly, the spectral and line length features are extracted using a multi-view feature extraction technique. Next, a signal temporal logic neural network (STONE) that combines the benefits of neural networks and temporal logics is constructed to classify the seizure and nonseizure data. STONE is designed in such a way that each neuron has a symbolic representation corresponding to a component in a weighted signal temporal logic (wSTL) formula. Compared with traditional STL inference algorithms, STONE is end-to-end differentiable such that the learning can be accomplished through back-propagation. In addition, STONE improves the interpretability of seizure detection models as the outcome of STONE is a wSTL formula that is interpretable and human-readable. Importantly, the wSTL formula reveals the reasoning behind seizure as a description of the evolution of EEG signals. STONE is tested on two popular EEG databases and demonstrated to achieve promising detection performance in terms of accuracy, sensitivity, and specificity when compared with existing state-of-the-art models. Furthermore, STONE can provide a human-readable formula as a description of the seizure characteristics, and the formula is also visualizable for easy interpretation of the classifier, which is a missing property in existing seizure detection methods.