一种可用于肝癌呼气信号鉴别的改进AdaBoost级联分类器

An Improved AdaBoost Cascade Classifier for Identifying Breath Signals of Liver Cancer

为了降低呼气检测技术在肝癌筛查中的漏诊率,本研究设计一种改进的AdaBoost级联分类器,并将其应用于鉴别健康志愿者和肝癌患者的呼气信号。首先,对训练样本进行自助划分获得一组训练子集。基于该训练子集,先后利用不同的机器学习算法,采用K折交叉训练和投票法得到多个子分类器;接着,将多个子分类器加权组合得到一个改进的AdaBoost分类器;然后,再次自助划分训练样本,以新的训练子集训练得到另一个AdaBoost分类器;最后,将两个AdaBoost分类器串联形成级联分类器。测试样本送入该级联分类器后,按照级联规则,潜在的异常样本将被反复筛查。以电子鼻采集到的120名志愿者的呼气信号的Relief优化特征集为训练样本,构建改进AdaBoost级联分类器,并对40例测试样本进行鉴别。结果表明,该级联分类器可有效区分出测试组中的肝癌患者和健康人的呼气信号,平均敏感性为93.42%,明显优于传统AdaBoost级联分类器,漏诊率显著降低。此外,该级联分类器的稳定性较好,精度的变异系数仅为3.95%。可见,改进AdaBoost级联分类器可有效提升分类器对肝癌呼气信号的检测能力,对实现基于呼气检测的肝癌无创普及性筛查技术的研究具有重要意义。

To reduce false negative rate of breath detection techniques in liver cancer screening,an improved AdaBoost cascade classifier was designed and applied to discriminate breath signals from healthy volunteers and liver cancer patients.First,a set of training subsets was obtained by self-help division of training samples.Based on the training subset,multiple sub-classifiers were successively obtained using different machine learning algorithms with K-fold cross-training and voting method.Next,multiple sub-classifiers were weighted and combined to obtain an improved AdaBoost classifier.Then,the training samples were self-subdivided and trained again with a new training subset to obtain another AdaBoost classifier.Finally,the two AdaBoost classifiers were concatenated in tandem to form a cascade classifier.After the test samples were fed into this cascade classifier,potentially anomalous samples were repeatedly screened according to the cascade rule.In this study,the relief-optimized feature set of the breath signals of 120 volunteers collected by the electronic nose(eNose)was used as the training sample to construct an improved AdaBoost cascade classifier and to discriminate the 40 test samples.The results showed that the classifier effectively distinguished the exhaled breath signals of liver cancer patients and healthy people in the test group,and the average sensitivity reached 93.42%,which was significantly better than the traditional AdaBoost cascade classifier,and the false negative rate was significantly reduced.In addition,the stability of this cascade classifier was good,and the coefficient of variation of the precision was only 3.95%.In conclusion,the improved AdaBoost cascade classifier effectively improved the classifier's discrimination accuracy of liver cancer breath signals,which was important for the study of breath-based noninvasive universal screening for liver cancers.