Donos Cristian, Mîndruţă Ioana, Barborica Andrei
Physics Department, Bucharest University, Bucharest, Romania.
Department of Neurology, Bucharest University Emergency Hospital, Bucharest, Romania.
Front Neurosci. 2020 Mar 23;14:183. doi: 10.3389/fnins.2020.00183. eCollection 2020.
High-frequency oscillations >80 Hz (HFOs) have unique features distinguishing them from spikes and artifactual components that can be well-evidenced in the time-frequency representations. We introduce an unsupervised HFO detector that uses computer-vision algorithms to detect HFO landmarks on two-dimensional (2D) time-frequency maps. To validate the detector, we introduce an analytical model of the HFO based on a sinewave having a Gaussian envelope, for which analytical equations in time-frequency space can be derived, allowing us to establish a direct correspondence between common HFO detection criteria in the time domain with the ones in the frequency domain, used by the computer-vision detection algorithm. The detector identifies potential HFO events on the time-frequency representation, which are classified as true HFOs if criteria regarding the HFO's frequency, amplitude, and duration are met. The detector is validated on simulated HFOs according to the analytical model, in the presence of noise, with different signal-to-noise ratios (SNRs) ranging from -9 to 0 dB. The detector's sensitivity was 0.64 at an SNR of -9 dB, 0.98 at -6 dB, and >0.99 at -3 dB and 0 dB, while its positive prediction value was >0.95, regardless of the SNR. Using the same simulation dataset, our detector is benchmarked against four previously published HFO detectors. The F-measure, a combined metric that takes into account both sensitivity and positive prediction value, was used to compare detection algorithms. Our detector surpassed the other detectors at -6, -3, and 0 dB and had the second best F-score at -9 dB SNR after the MNI detector (0.77 vs. 0.83). The ability to detect HFOs in clinical recordings has been tested on a set of 36 intracranial electroencephalogram (EEG) channels in six patients, with 89% of the detections being validated by two independent reviewers. The results demonstrate that the unsupervised detection of HFOs based on their 2D features in time-frequency maps is feasible and has a performance comparable or better than the most used HFO detectors.
频率大于80Hz的高频振荡（HFOs）具有独特特征，使其有别于尖峰信号和人为成分，这在时频表示中能够得到充分证明。我们介绍一种无监督HFO检测器，它使用计算机视觉算法在二维（2D）时频图上检测HFO特征点。为验证该检测器，我们引入基于具有高斯包络的正弦波的HFO分析模型，据此可推导时频空间中的解析方程，从而使我们能够在时域中常见的HFO检测标准与计算机视觉检测算法所使用的频域检测标准之间建立直接对应关系。该检测器在时频表示上识别潜在的HFO事件，若满足有关HFO频率、幅度和持续时间的标准，则将其分类为真正的HFO。该检测器根据分析模型在存在噪声的情况下，针对不同信噪比（SNR）范围从-9至0dB的模拟HFO进行了验证。在SNR为-9dB时，检测器的灵敏度为0.64，在-6dB时为0.98，在-3dB和0dB时大于0.99，而其阳性预测值大于0.95，与SNR无关。使用相同的模拟数据集，我们的检测器与之前发表的四种HFO检测器进行了基准测试。F值是一种综合考虑灵敏度和阳性预测值的指标，用于比较检测算法。我们的检测器在-6dB、-3dB和0dB时超过了其他检测器，在SNR为-9dB时，其F分数仅次于MNI检测器，位居第二（0.77对0.83）。在六名患者的一组36个颅内脑电图（EEG）通道上测试了在临床记录中检测HFO的能力，89%的检测结果得到了两名独立评审员的验证。结果表明，基于时频图中HFO二维特征的无监督检测是可行的，并且其性能与最常用的HFO检测器相当或更好。
