Department of Pediatric Neurosurgery, Severance Children's Hospital, Yonsei University College of Medicine, 50-1 Yonsei-ro, Seodaemun-gu, Seoul, 03722, Korea.
Institute of Integrated Technology (IIT), Gwangju Institute of Science and Technology (GIST), Cheomdan-gwagiro 123, Buk-gu, Gwangju, 61005, South Korea.
Childs Nerv Syst. 2021 Jul;37(7):2239-2244. doi: 10.1007/s00381-020-05011-9. Epub 2021 May 3.
Seizures are one of the most common emergencies in the neonatal intensive care unit (NICU). They are identified through visual inspection of electroencephalography (EEG) reports and treated by neurophysiologic experts. To support clinical seizure detection, several feature-based automatic neonatal seizure detection algorithms have been proposed. However, as they were unsuitable for clinical application due to their low accuracy, we developed a new seizure detection algorithm using machine learning for single-channel EEG to overcome these limitations.
The dataset applied in our algorithm contains EEG recordings from human neonates. A 19-channel EEG system recorded the brain waves of 79 term neonates admitted to the NICU at the Helsinki University Hospital. From these datasets, we selected six patients with conformational seizure annotations for the pilot study and allocated four and two patients for our training and testing datasets, respectively. The presence of seizures in the EEGs was annotated independently by three experts through visual interpretation. We divided the data into epochs of 5 s each and further defined a seizure block to label the annotations from each expert recorded every second. Subsequently, to create a balanced dataset, any data point with a non-seizure label was moved to the training and test dataset.
The developed principal component feature-extracted machine learning algorithm used 62.5% of the relative time (only 5 s for decision) of the baseline, reaching an area under the ROC curve score of 0.91. The effect of diversified parameters was meticulously examined, and 100 principal components were extracted to optimize the model performance.
Our machine learning-based seizure detection algorithm exhibited the potential for clinical application in NICUs, general wards, and at home and proved its convenience by requiring only a single channel for implementation.
癫痫发作是新生儿重症监护病房（NICU）中最常见的急症之一。它们通过脑电图（EEG）报告的视觉检查来识别，并由神经生理专家进行治疗。为了支持临床癫痫发作检测，已经提出了几种基于特征的自动新生儿癫痫发作检测算法。然而，由于其准确性低，它们不适合临床应用，因此我们开发了一种使用机器学习的新的单通道 EEG 癫痫发作检测算法，以克服这些限制。
我们的算法应用的数据集包含人类新生儿的脑电图记录。一个 19 通道 EEG 系统记录了赫尔辛基大学医院 NICU 收治的 79 名足月新生儿的脑波。从这些数据集中，我们选择了 6 名具有构象性癫痫发作注释的患者进行试点研究，并分别将 4 名和 2 名患者分配用于我们的训练和测试数据集。脑电图中癫痫发作的存在由三位专家通过视觉解释进行独立注释。我们将数据分为每个 5 秒的时间段，并进一步定义一个癫痫发作块，以标记每位专家每秒记录的注释。随后，为了创建一个平衡的数据集，任何具有非癫痫发作标签的数据点都被移动到训练和测试数据集。
开发的主成分特征提取机器学习算法使用基线的相对时间的 62.5%（仅用于决策的 5 秒），达到了 0.91 的 ROC 曲线下面积得分。仔细检查了多样化参数的效果，并提取了 100 个主成分来优化模型性能。
我们的基于机器学习的癫痫发作检测算法具有在 NICU、普通病房以及在家中临床应用的潜力，并通过仅需单通道实施证明了其便利性。
