Mohammadi Mahdi, Al-Azab Fadwa, Raahemi Bijan, Richards Gregory, Jaworska Natalia, Smith Dylan, de la Salle Sara, Blier Pierre, Knott Verner
Knowledge Discovery and Data mining Lab (KDD), University of Ottawa, Ottawa, ON, Canada.
Department of Psychiatry, McGill University, Montreal, QC, Canada.
BMC Med Inform Decis Mak. 2015 Dec 23;15:108. doi: 10.1186/s12911-015-0227-6.
Quantitative electroencephalogram (EEG) is one neuroimaging technique that has been shown to differentiate patients with major depressive disorder (MDD) and non-depressed healthy volunteers (HV) at the group-level, but its diagnostic potential for detecting differences at the individual level has yet to be realized. Quantitative EEGs produce complex data sets derived from digitally analyzed electrical activity at different frequency bands, at multiple electrode locations, and under different vigilance (eyes open vs. closed) states, resulting in potential feature patterns which may be diagnostically useful, but detectable only with advanced mathematical models.
This paper uses a data mining methodology for classifying EEGs of 53 MDD patients and 43 HVs. This included: (a) pre-processing the data, including cleaning and normalization, applying Linear Discriminant Analysis (LDA) to map the features into a new feature space; and applying Genetic Algorithm (GA) to identify the most significant features; (b) building predictive models using the Decision Tree (DT) algorithm to discover rules and hidden patterns based on the reduced and mapped features; and (c) evaluating the models based on the accuracy and false positive values on the EEG data of MDD and HV participants. Two categories of experiments were performed. The first experiment analyzed each frequency band individually, while the second experiment analyzed the bands together.
Application of LDA and GA markedly reduced the total number of utilized features by ≥ 50 % and, with all frequency bands analyzed together, the model showed average classification accuracy (MDD vs. HV) of 80 %. The best results from model testing with additional test EEG recordings from 9 MDD patients and 35 HV individuals demonstrated an accuracy of 80 % and showed an average sensitivity of 70 %, a specificity of 76 %, and a positive (PPV) and negative predictive value (NPV) of 74 and 75 %, respectively.
These initial findings suggest that the proposed automated EEG analytical approach could be a useful adjunctive diagnostic approach in clinical practice.
定量脑电图（EEG）是一种神经影像学技术，已被证明在群体水平上能够区分重度抑郁症（MDD）患者和非抑郁健康志愿者（HV），但其在个体水平上检测差异的诊断潜力尚未得到充分发挥。定量脑电图产生的复杂数据集来自于对不同频段、多个电极位置以及不同警觉状态（睁眼与闭眼）下的电活动进行数字分析，从而形成潜在的特征模式，这些模式可能具有诊断价值，但只有通过先进的数学模型才能检测到。
本文采用数据挖掘方法对53例MDD患者和43例HV的脑电图进行分类。这包括：（a）对数据进行预处理，包括清理和归一化，应用线性判别分析（LDA）将特征映射到新的特征空间；应用遗传算法（GA）识别最重要的特征；（b）使用决策树（DT）算法构建预测模型，基于减少和映射后的特征发现规则和隐藏模式；（c）根据MDD和HV参与者脑电图数据的准确性和假阳性值对模型进行评估。进行了两类实验。第一个实验分别分析每个频段，而第二个实验则一起分析这些频段。
LDA和GA的应用显著减少了所利用特征的总数≥50%，并且在所有频段一起分析时，模型显示出平均分类准确率（MDD与HV）为80%。使用来自9例MDD患者和35例HV个体的额外测试脑电图记录进行模型测试的最佳结果显示准确率为80%，平均灵敏度为70%，特异性为76%，阳性预测值（PPV）和阴性预测值（NPV）分别为74%和75%。
这些初步发现表明，所提出的自动脑电图分析方法可能是临床实践中一种有用的辅助诊断方法。
