Machine learning is integrated nowadays in many data-driven applications that attempt to model the behavior of a system. Thus, the implementation of machine-learning algorithms for medical applications is growing, enabling doctors to make decisions based on the output of the model of the system’s behavior. The upper airway is involved in a variety of disorders that lead to non-specific symptoms; thus, upper-airway obstruction is frequently unrecognized or misdiagnosed. Bronchoscopy, which is a minimally invasive procedure, and lung function (spirometry) tests, which are relatively demanding for the patient, are currently the most common methods for diagnosing respiratory diseases. In this study, a novel, non-invasive procedure is proposed in which tracheal obstruction is identified based on brain signals. Specifically, the spectral information in electroencephalogram (EEG) signals is used as an input to an ensemble learner approach based on ordinal and non-ordinal classification algorithms, where the classification problem involves identifying the degree of airway obstruction. An experiment was conducted in which four healthy subjects breathed through three-dimensional (3D) geometric models of the trachea that mimicked different obstruction rates. Multi-subject classification was carried out in which the classification model of each subject was produced by training the model on the other subjects' datasets. The main findings were as follows. Firstly, the in-house ordinal classification algorithms, which included a C4.5 and a random-forest algorithm, both based on a weighted information-gain ratio measure, yielded better classification results than their non-ordinal counterparts and other conventional classifiers. Additionally, the study showed that when integrating the two types of algorithms (ordinal and non-ordinal) into an ensemble approach, the performance was improved relative to each individual classifier. Finally, the classification accuracy is such that the proposed method of using EEG signals for the identification of the degree of tracheal obstruction by means of an ensemble approach shows promise as a supplemental clinical test.

如今，机器学习已集成在许多试图为系统行为建模的数据驱动的应用程序中。因此，用于医学应用的机器学习算法的实现正在增长，使医生能够根据系统行为模型的输出做出决策。上呼吸道涉及多种导致非特异性症状的疾病。因此，上呼吸道阻塞常常无法被识别或误诊。支气管镜检查是一种微创手术，而肺功能（肺活量测定）测试对患者的要求相对较高，目前是诊断呼吸系统疾病的最常用方法。在这项研究中，提出了一种新颖的，非侵入性的程序，其中根据脑信号识别气管阻塞。具体来说，脑电图（EEG）信号中的频谱信息用作基于有序和非有序分类算法的整体学习器方法的输入，其中分类问题涉及识别气道阻塞的程度。进行了一项实验，其中四名健康受试者通过模仿不同阻塞率的气管的三维（3D）几何模型进行了呼吸。进行多主题分类，其中通过在其他主题的数据集上训练模型来生成每个主题的分类模型。主要发现如下。首先，内部序数分类算法（包括C4.5和随机森林算法）均基于加权信息增益比度量，比非常规分类器和其他常规分类器产生更好的分类结果。此外，研究表明，将两种类型的算法（常规和非常规）集成到集成方法中时，相对于每个单独的分类器，性能都会得到改善。最后，分类的准确性使得所提出的利用脑电信号通过整体方法识别气管阻塞程度的方法显示了作为补充临床试验的希望。