AI enabled biomarker identification from exhaled breath condensates for early detection of secondary infection in COPD patients

目前对呼吸道病原体的诊断可能需要侵入性采样，这使得常规筛查劳动密集、令人不快且对患者不切实际。由于牛结核病感染可以从呼出气中存在的生物标志物中诊断，我们希望进一步开发这种方法来诊断慢性阻塞性肺疾病（COPD）患者的感染。慢性阻塞性肺病的特征是长期的呼吸问题，继发性感染可能会加剧这种问题。全世界近12%的人患有COPD，每年导致300万人死亡。使用和开发先进的成像和机器学习技术来分析质谱（MS）数据，我们的目标是确定特定的呼出生物标志物，这些生物标志物将在症状发作之前诊断继发性感染。这将允许积极治疗，以防止加重和住院。这些呼出的生物标记物的检测系统最终可以被开发出来，以实现在护理点的非侵入性快速监测。成像和机器学习技术能否在COPD患者呼出气样本的MS数据中识别继发性感染的诊断生物标志物？2.确定的生物标志物是否可用于检测/预测正在发生的感染或病情加重？3.不同技术鉴定的生物标志物之间的相关性如何？使用公开可用的数据集将使学生熟悉相关数据集所涉及的计算方法，并测试初始方法在收集临床样本时将已知分组分离为潜在解决方案开发的一部分的能力。我们将对现有的COPD患者队列进行采样。将在感染期间（感染）和恢复期（正常）采集呼吸样本，以识别感染特异性生物标志物。将通过MS分析样品，以鉴别存在的分子及其浓度。将涵盖整个质量范围的MS数据转换为热图，描绘样品的相对质荷比（m/z）强度分布。然后，将使用复杂的图像处理技术对这些预处理数据进行分析。最初的MS数据也将被清理并准备用于基于机器学习的方法进行交叉验证。第二阶段将应用最先进的模式识别方法，如DeepCluster和机器学习算法，包括支持向量机，随机森林和人工神经网络处理MS数据，以将COPD患者分类为感染或未感染。然后，学生将以此为起点，开发更先进的算法，以提高生物标志物识别的准确性，从而预测可能的感染。在这个完善阶段，SLS学者将贡献他们的专业知识，以确保确定的生物标志物是生物相关的和有意义的。这些数据还将通过机器学习方法进行分析，以提供成像算法结果的计算机验证，并为识别推定的感染生物标志物提供进一步的信心。在第三阶段，将使用未知的临床样本验证第二阶段预测的生物标志物。与临床合作者合作，将通过MS分析盲态样本，并通过开发的算法使用数据，根据存在的生物标志物将每个样本分配为正常或感染。一旦通过计算方法分配，样本将被揭盲，以评估这种方法的成功，并允许进一步细化。这项工作将有助于开发有效的先进图像处理和机器学习算法，用于临床环境。它还可以提供医疗支持系统的进步，允许对感染进行积极的临床管理，以预防COPD加重并减少住院治疗。

Current diagnosis of respiratory pathogens can require invasive sampling which makes routine screening labour intensive, unpleasant and impractical for patients. As bovine TB infections can be diagnosed from biomarkers present in exhaled breath, we wish to further develop this approach for the diagnosis of infections in chronic obstructive pulmonary disease (COPD) patients. COPD is characterised by longterm breathing problems which can be exacerbated by secondary infections. Worldwide nearly 12% of people have COPD which causes 3 million deaths annually. Using and developing advanced imaging and machine learning techniques to analyse mass spectrometry (MS) data, we aim to identify specific exhaled biomarkers that will diagnose secondary infections prior to the onset of symptoms. This would allow proactive treatment to prevent exacerbations and hospitalisation. Detection systems for these exhaled biomarkers could ultimately be developed to enable noninvasive rapid monitoring at the point of care.The proposed work will address three questions:1. Can imaging and machine learning techniques identify diagnostic biomarkers for secondary infections in MS data of exhaled breath samples from COPD patients?2. Can identified biomarkers be used to detect/predict developing infections or exacerbations?3. How well do biomarkers identified by different techniques correlate?Use of publicly available datasets will allow the student to become familiar with the computational methods involved on related data sets and to test initial methods for their ability to segregate known groupings as part of the development of potential solutions while clinical samples are collected. We will sample an existing cohort of COPD patients to which we have access. Breath samples will be collected during infections (infected) and on recovery (normal) to enable identification of infection specific biomarkers. Samples will be analysed by MS to identify the molecules present and their concentrations. The data from the MS covering the whole mass range will be converted to a heat map depicting the relative mass-tocharge (m/z) intensity distribution of the sample. This preprocessed data will then be analysed using sophisticated image processing techniques. The initial MS data will also be cleaned and prepared for a machine learning based approach for cross validation.Stage two will apply state-of-the-art pattern recognition methods such as DeepCluster and machine learning algorithms including support vector machine, random forests, and artificial neural networks to the processed MS data to classify COPD patients as infected or not. The student will then use this as a starting point to develop more advanced algorithms to improve the accuracy of biomarker identification to enable prediction of possible infections. During this refinement phase, the SLS academics will contribute their expertise to ensure identified biomarkers are biologically relevant and meaningful. The data will also be analysed by a machine learning approach to provide in silico validation of the results from the imaging algorithms and provide further confidence to the identification of putative biomarkers of infection.In stage three, predicted biomarkers from stage 2 will be validated using unknown clinical samples. Working with clinical collaborators, blinded samples will be analysed by MS and the data used by the developed algorithms to assign each as normal or infected based on the biomarkers present. Once assigned by the computational approach, samples will be unblinded to assess the success of this approach, and allow further refinement.This work will contribute to the development of effective advanced image processing and machine learning algorithms for use in clinical settings. It may also provide advances in healthcare support systems that allow proactive clinicalmanagement of infections to prevent COPD exacerbations and reduce hospitalisations.