睡眠纺锤波和K复合波作为暂态振荡脑电波型，可以被用于评估认知能力和相关睡眠障碍的重要特征波，也是睡眠分期的重要标志之一。目前临床上广泛使用的人工检测方法仍被认为是最准确的金标准，但对医师的专业要求较高，且不同医师间的检测结果存在较大差异，因此会占用较多的临床资源，并且现有自动化检测方法性能表现欠佳，因此本文提出了基于语义分割的使用多特征融合的特征波检测模型MuFF。根据纺锤波和K复合波的时域定义和频率范围，并且考虑到其统计特性在不同年龄段以及不同性别的人群中存在差异性，本文综合时域波形、时频域图像以及元数据并结合U-Net结构对特征波进行检测。MuFF通过对时域信号和时频域图像经过不同的编码器，将这两者特征进行拼接，然后使用元数据编码器对元数据进行编码，并加到拼接之后的特征上，然后通过轻量化注意力模块进行特征融合。本文通过交叉验证的方式获得多个MuFF模型，最终通过集成的方式将单个模型的结果汇聚得到最终输出，算法简称为MuFF-E。在公开数据集MODA上纺锤波检测MuFF-E取得了84.5%的F1值、87.4%的精确率以及82.3%的召回率，验证了在特征波检测上多特征融合的检测算法的有效性，相较于目前其他的纺锤波检测算法和K复合波检测算法有一定的提升。考虑到实际应用中睡眠纺锤波和K复合波的现有标注数据较少，并且人工标注过于消耗财力人力，模型在不同的数据集上表现差异较大，泛化水平较低。为了提高模型的泛化能力以及在小样本条件下的表现，本文将自监督学习应用到脑电表征学习上，并且提出了使用对比预测编码和时频域对比相结合的代理任务。算法将时域和时频域图像看作是不同视角的观测值，对应位置为正样本，不同位置为负样本，由此设计了时频域对比的损失函数，可以在大规模的无标注数据上对编码器进行预训练，并在少量有标注数据上进行有监督微调。实验结果显示，在20%标注数据量的条件下，在公开数据集上纺锤波检测取得了74.6%的F1值。在北京同仁医院数据集上纺锤波检测取得了72.5%的F1值，K复合波检测取得了71.6%的F1值，验证了小样本条件下自监督算法的有效性并超过了其他的自监督算法。
Sleep spindles and K-complexes, as transient oscillatory EEG waveforms, can be used to assess cognitive abilities and important features of related sleep disorders, and they are also key markers for sleep staging. Currently, the manually detected methods widely used in clinical practice are still considered the most accurate gold standard. However, these methods demand high levels of professional expertise from physicians, and there is significant variability in detection results among different physicians. Consequently, these methods consume substantial clinical resources. Moreover, the performance of existing automated detection methods is suboptimal. Therefore, this thesis proposes a feature wave detection model based on semantic segmentation and multi-feature fusion, named MuFF.Based on the time-domain definition and frequency range of spindles and K-complexes, and considering the statistical differences across different age groups and genders, this thesis integrates time-domain waveforms, time-frequency domain images, and metadata, utilizing a U-Net structure for feature wave detection. MuFF detects feature waves by encoding the time-domain signals and time-frequency domain images with different encoders, concatenating these features, and then encoding the metadata with a metadata encoder. The metadata is added to the concatenated features, followed by feature fusion through a lightweight attention module. Multiple MuFF models are obtained through cross-validation, and the final output is produced by aggregating the results of individual models using an ensemble method, abbreviated as MuFF-E. On the publicly available MODA dataset, MuFF-E achieved an F1 score of 84.5%, a precision of 87.4%, and a recall of 82.3% for spindle detection, validating the effectiveness of the multi-feature fusion detection algorithm for feature wave detection. This represents an improvement over current spindle detection algorithms and K-complex detection algorithms.Considering the limited annotated data available for sleep spindles and K-complexes in practical applications, and the significant financial and labor costs associated with manual annotation, models often exhibit significant performance variability across different datasets and generally have low generalization capabilities. To enhance the generalization ability of the model and its performance under small sample conditions, this thesis applies self-supervised learning to EEG representation learning and proposes a surrogate task combining contrastive predictive coding and time-frequency domain contrast. The algorithm treats time-domain and time-frequency domain images as observations from different perspectives, with corresponding positions as positive samples and different positions as negative samples. This design of the time-frequency domain contrast loss function allows the encoder to be pre-trained on large-scale unlabeled data and fine-tuned with a small amount of labeled data. Experimental results show that under the condition of 20% labeled data, the spindle detection achieved an F1 score of 74.6% on the public dataset. On the Beijing Tongren Hospital dataset, spindle detection achieved an F1 score of 72.5%, and K-complex detection achieved an F1 score of 71.6%, validating the effectiveness of the self-supervised algorithm under small sample conditions and surpassing other self-supervised algorithms.
