基于最小分类误差准则的呼吸音分类技术

Respiratory Sound Classification Approach Based on Minimum Classification Error

从大量呼吸音样本中归纳综合出肺部病理特征的科学表示,实现自动化、定量化的呼吸音分类,是现代医疗信息化技术的重要研究内容之一.提出了一种基于最小分类误差(minimum classification error,MCE)准则的呼吸音分类方法,建立呼吸音类别的分类误差损失函数,采用广义概率下降法(generalized probabilistic decent,GPD)估计得到呼吸音的隐马尔科夫模型(hidden Markov model,HMM)参数,以增强不同类型呼吸音模型的区分能力.实验结果表明,与传统的最大似然(maximum likelihood,ML)法相比,基于MCE准则求解的HMM模型,具有更好的分类效果,提高了识别准确率,客观证明了基于MCE准则的呼吸音分类技术的有效性.

Unlike the traditional auscultation, automatic respiratory sound classification technology summarizes the scientific descriptions of pathological features from a large number of respiratory sound samples.And it serves as an automatic and quantitative auscultation tool to diagnose abnormalities and disorders in the lung. A classification procedure based on minimum classification error （MCE） approach using hidden Markov models （HMM） is proposed in this paper.The parameters of HMM are estimated by loss functions between different models of normal sounds and abnormal sounds, which aim to distinguished healthy subjects and patients. The experiment results show that the proposed HMM-MCE approach obtains higher classfication accuracy in comparison with the traditional HMM-ML method.