阿尔茨海默病和轻度认知障碍患者脑电特征分析与识别研究

Analysis and identification of electroencephalogram features in patients with Alzheimer’s disease and mild cognitive impairment

目的分析阿尔茨海默病(AD)和轻度认知障碍(MCI)患者脑电特征, 并结合特征进行分类预测。方法选取天津医科大学总医院神经内科就诊患者135例为研究对象, 其中AD患者34例, MCI患者67例, 健康对照(HC)34例, 分别采集静息状态下脑电信号并进行预处理。提取多频段尺度的相对功率谱密度特征和样本熵特征, 比较3组被试脑电特征的全脑差异, 然后再细分脑区、单个导联深入分析。融合上述2种特征, 通过支持向量机(SVM)对AD、MCI和HC进行分类预测。结果额叶区域δ相对功率谱密度高于其他区域, 枕叶和颞叶区域表现出相对更低的分布占比。θ频段相对功率谱密度在各个脑区的大小分布情况较为平均。α频段相对功率谱密度较高的导联均集中于枕叶。β频段的相对功率谱密度较高的导联则主要集中在顶叶和颞叶。除中央叶外, 其余各个脑区及全脑, AD组的δ频段相对功率谱密度值均高于MCI组和HC组(均P<0.05、0.01)。AD组全脑及各个脑区的θ频段相对功率谱密度高于MCI组和HC组(均P<0.001)。AD组的α频段相对功率谱密度仅在颞叶低于其他组(均P<0.05)。AD组β频段的相对功率谱密度在全脑及各脑区高于其他组(P<0.05、0.01、0.001)。AD组与HC组中央叶的C3导联的δ频段相对功率谱密度差异比较有统计学意义(P<0.05)。AD组、MCI组和HC组颞叶γ频段相对功率谱密度均高于其他区域。AD组T3导联处γ频段相对功率谱密度明显低于T4导联处。AD组和MCI组的全脑平均及各个脑区平均样本熵均低于HC组(均P<0.05)。AD组C3导联处的样本熵低于MCI组(P<0.05)。相对功率谱密度、样本熵以及融合2种特征的实际数据分类评价指标(准确率、精确率、召回率和F1分数)与重排数据差异均有统计学意义(均P<0.001);当分类特征中融合了相对功率谱密度特征及样本熵特征时, 分类预测效果最好, 准确率达80%, 精确率达78%, 召回率为78%, F1分数为79%。结论相对功率谱密度与样本熵分析能够从不同角度(线性与非线性)揭示AD、MCI患者脑电活动的异常, 在分类预测时结合相对功率谱密度和样本熵特征, 能够提高分类效果。

Objective To analyze the electroencephalogram(EEG)features of patients with Alzheimer’s disease(AD)and mild cognitive impairment(MCI),and to combine the characteristics for classification and prediction.Methods One hundred and thirty-five patients attending the Department of Neurology at the General Hospital of Tianjin Medical University were enrolled,including 34 patients with AD,67 patients with MCI,and 34 healthy control(HC).The electroencephalogram signals of these patients in the resting state were collected and preprocessed.Relative power spectral density features and sample entropy features on a multi-band scale were extracted to compare the whole-brain differences in electroencephalogram features among the 3 groups of subjects,and then subdivided into brain regions and individual leads for in-depth analysis. The above two features were fused to classify and predict AD, MCI, and HC by support vector machine (SVM). Results The frontal regions had higher δ relative power spectral densities than the other regions, and the occipital and temporal regions showed relatively lower distributions. θ-Band relative power spectral densities had a more even distribution of sizes across brain regions. α-Band relative power spectral densities were concentrated in the occipital lobe, while β-band relative power spectral densities were mainly concentrated in the parietal and temporal lobes. Except for the central lobe, the δ-band relative power spectral densities of the AD group were higher than those of the MCI group (P < 0.05) and HC group (P < 0.01) in all brain regions and the whole brain. θbandrelative power spectral densities of the AD group were higher than those of the MCI gourp (P < 0.001) and HC group (P < 0.001) in the whole brain and in all brain regions. α-Band relative power spectral densities of the AD group were lower than those of the other groups only in the temporal lobe (all P < 0.05). The relative power spectral density of the β-band in the AD group was higher than that of the other groups in the whole brain and in all brain regions (P < 0.05, 0.01, 0.001). The difference in the relative power spectral density of the δ-band in the C3 lead in the central lobe of the AD and HC groups was statistically significant (P < 0.05). The relative power spectral density of the γ-band in the temporal lobe was higher than that in the other regions of the AD group, the MCI group, and the HC group. The relative power spectral density of the γ-band in the T3 lead in the AD group was significantly lower than that in the T4 lead. The average entropy of samples in the whole brain and in each brain region was lower than that in the HC group in the AD and MCI groups (all P < 0.05). The entropy of the samples at lead C3 in the AD group was lower than that in the MCI group (P < 0.05). The differences between the relative power spectral density, sample entropy, and the actual data classification evaluation indexes (accuracy rate, precision rate, recall rate, and F1 score) that fused the two features, and the rearranged data were all statistically significant (all P < 0.001). When the relative power spectral density feature and the sample entropy feature were fused in the classification features, the best classification prediction was achieved, with an accuracy rate of 80%, a precision rate of 78%, a recall rate of 78%, and the F1 score of 79%. Conclusions Relative power spectral density and sample entropy analysis can reveal the abnormalities of electroencephalogram activities of AD and MCI patients from different perspectives (linear and nonlinear), and the combination of these two features in classification prediction can improve the classification effect.