Mohammad Umair, Saeed Fahad
Knight Foundation School of Computing and Information Sciences, Florida International University, Miami, FL, USA.
Int Conf Distrib Comput Sens Syst Workshops. 2024 Apr-May;2024:620-626. doi: 10.1109/dcoss-iot61029.2024.00097. Epub 2024 Aug 12.
Seizures pose a significant health hazard for over 50 million individuals with epilepsy worldwide, with approximately 56% experiencing uncontrollable seizures according to the CDC. Predicting seizures is challenging even with the availability of various sensors (gyroscopes, pulse rate sensors, heart rate monitors, etc). Electroencephalography (EEG) data can directly measure the activity of the brain and has been the choice of leveraging deep learning (DL) models for seizure prediction. Despite DL models achieving over 95% accuracy on retroactive clinical-grade EEG data, this performance fails to translate in real-world settings where the accuracy goes down to 66% - which warrants further investigation. Moreover, consumer-grade wearable EEG headsets, characterized by lower data quality and a varying number of channels across brands, present additional challenges. In this paper, we estimate the robustness of DL models which are trained on clinical-grade EEG data but tested on the type of data expected from consumer-grade wearable EEG headsets. We select the previously published model SPERTL to estimate its robustness when: (1) predicting with data from less leads/channels, (2) predicting when faced with streaming data, (3) evaluating performance on imbalanced data with more interictal segments. Our results are compared against baseline results from the SPERTL model which we have re-configured to operate independently of the number of channels with an average baseline area under the curve (AUC) score of 98.56%. Our results demonstrate that though the model is surprisingly resilient to streaming and noisy data, reducing the number of channels and a higher class imbalance have a more severe degradation. The AUC across all cross-validation sets degrades only by 2% and 3% on average for noisy and streaming data, respectively. However, a performance reduction, on average, is observed by 32% when imbalance is increased with higher percentage of interictal samples, and up to 16% when using lower number of channels.
癫痫发作对全球超过5000万癫痫患者构成了重大的健康威胁，根据美国疾病控制与预防中心（CDC）的数据，约56%的患者经历过无法控制的癫痫发作。即便有各种传感器（陀螺仪、脉搏率传感器、心率监测器等），预测癫痫发作仍具有挑战性。脑电图（EEG）数据可以直接测量大脑活动，并且一直是利用深度学习（DL）模型进行癫痫发作预测的选择。尽管DL模型在追溯性临床级EEG数据上的准确率超过了95%，但这种性能在现实环境中却无法实现，在现实环境中准确率降至66%，这值得进一步研究。此外，消费级可穿戴EEG耳机具有数据质量较低且各品牌通道数量不同的特点，带来了额外的挑战。在本文中，我们评估了在临床级EEG数据上训练但在消费级可穿戴EEG耳机预期的数据类型上进行测试的DL模型的稳健性。我们选择先前发表的模型SPERTL，在以下情况下评估其稳健性：（1）使用较少导联/通道的数据进行预测；（2）面对流式数据时进行预测；（3）在具有更多发作间期片段的不平衡数据上评估性能。我们将结果与SPERTL模型的基线结果进行比较，我们重新配置了该模型使其独立于通道数量运行，平均基线曲线下面积（AUC）分数为[X]%。我们的结果表明，尽管该模型对流式数据和噪声数据具有惊人的弹性，但减少通道数量和更高的类别不平衡会导致更严重的性能下降。在所有交叉验证集中，噪声数据和流式数据的AUC平均分别仅下降2%和3%。然而，当发作间期样本百分比更高导致不平衡增加时，平均性能下降32%，使用较少通道时性能下降高达16%。 （注：原文中“98.56%”处缺失具体信息，翻译时保留原文格式以便读者理解此处信息缺失情况）
