Background

Inter-ictal state is a period between convolutions (seizures). Expert neurologist looks for inter-ictal activity within this period to support the diagnosis of epilepsy. The focus of this work is to automate the process of identification of inter-ictal activity from EEG and to distinguish it from the activity of a controlled patient. Also, we have worked on differentiating between different epileptic states. This work uses the Benchmark Bonn dataset and novel patient data collected from Max Hospital, Saket. Five groups are considered from Bonn database with the first four groups, with one case each and group five divided into ten cases and data collected from hospital is also considered. This study explores four cases under group 5, reporting of which is not available in the literature for bonn dataset and also reports the results obtained from data collected at Max Hospital. Two scenarios for Group 5 are presented under the first, the complete signal of length 23.6 s is converted into scalograms and in next scenario the complete signal is broken into segments of 2 s to make a comparative study with real time database.

New method

In this work, Continuous Wavelet Transform (CWT) is used to convert the signals to scalograms. Scalograms are further, resized to reduce computation. A fifteen-layer novel Convolution neural network architecture is applied to these scalograms to classify these into their respective classes. Also present results obtained on Bonn data with two second segments and show a comparative performance with the real time data.

Comparison and results

State of the art methods are compared with our novel methodology in context to the five out of fifteen cases considered under different groups. The performance evaluation parameters considered are accuracy, specificity, sensitivity, and F1 score. This study establishes a clear comparison and outperforms the state-of-the-art methods in context with performance measures considered. We have achieved a classification accuracy of 99.83 %, 100, 100 %, 99.833 % for cases AB-C, AB-D, CD-A, CD-B. Also, our method has outperformed pre-existing methods, achieving classification accuracy of 99.8 %, 100 %, 99.8 %, 99.75 %, and 99.75 % for AB-E, CD-E, ABCD-E, AB-CD-E, and AB-CD. For the real time data collected at Max Hospital, Saket and Benchmark dataset the model provided an accuracy of 91.7 %.

Conclusion

The novel Convolution neural network architecture deployed in this work has outperformed the existing methods for the cases previously considered by researchers, which used the same benchmark dataset. Also, for the problem of identification of inter-ictal discharges, the proposed model has shown excellent performance concerning all the performance metrics with the data collected from clinical setup. Further, this study establishes the efficiency and usability of scalogram for identification of different epileptic states.

中文翻译：

---

使用尺度图实现卷积神经网络识别癫痫活动

背景

发作间期是卷积（癫痫发作）之间的时期。专家神经学家在此期间寻找发作间期活动以支持癫痫的诊断。这项工作的重点是自动化从 EEG 识别发作间期活动的过程，并将其与受控患者的活动区分开来。此外，我们还致力于区分不同的癫痫状态。这项工作使用 Benchmark Bonn 数据集和从 Saket 的 Max Hospital 收集的新患者数据。从波恩数据库中考虑了五组，前四组，每组 1 例，第 5 组分为 10 例，还考虑了从医院收集的数据。本研究探讨了第 5 组下的四个案例，波恩数据集的文献中没有报告，并且还报告了从 Max Hospital 收集的数据中获得的结果。第 5 组的两个场景呈现在第一个场景中，将长度为 23.6 s 的完整信号转换为尺度图，在下一个场景中，将完整信号分解为 2 s 的片段，与实时数据库进行比较研究。

新方法

在这项工作中，连续小波变换 (CWT) 用于将信号转换为尺度图。缩放图进一步调整大小以减少计算量。将十五层新颖的卷积神经网络架构应用于这些尺度图，以将它们分类到各自的类别中。还展示了在波恩数据上获得的具有两个秒段的结果，并显示了与实时数据的比较性能。

比较和结果

将最先进的方法与我们在不同组下考虑的十五个案例中的五个在上下文中的新方法进行比较。考虑的性能评估参数是准确性、特异性、敏感性和 F1 分数。本研究建立了一个清晰的比较，并在考虑性能指标的情况下优于最先进的方法。对于案例 AB-C、AB-D、CD-A、CD-B，我们实现了 99.83%、100、100%、99.833% 的分类准确率。此外，我们的方法优于现有方法，对 AB-E、CD-E、ABCD-E、AB-CD-E 和A B C D。对于在 Max Hospital、Saket 和 Benchmark 数据集中收集的实时数据，该模型提供了 91.7% 的准确度。

结论

在这项工作中部署的新型卷积神经网络架构在研究人员之前考虑的使用相同基准数据集的情况下优于现有方法。此外，对于识别发作间期放电的问题，所提出的模型在所有性能指标和从临床设置中收集的数据方面都表现出出色的性能。此外，本研究确定了分级图在识别不同癫痫状态方面的效率和可用性。