Automated lung sound analysis to improve the clinical diagnosis of pulmonary tuberculosis in children

项目摘要 儿童结核病（TB）死亡的大多数是未开始治疗的儿童， 迫切需要防止延误诊断。然而，儿童结核病的诊断具有挑战性， 基于乳腺癌的检测是侵入性的，在儿童中的检出率较低，并且胸部X线检查（CXR）不是常规检查 在高负担的环境中，需要受过训练的读者。因此，世界卫生组织（WHO） 认可结核病治疗决策算法，以减少儿童及时开始结核病治疗的延误 向较低级别的医疗机构就诊。然而，TB治疗决策算法具有不确定的准确性 现有的数据表明，在没有CXR的情况下，它们可能具有较差的特异性。人工智能 应用于使用数字听诊器（Lung AI）收集的肺音的算法具有增强 在接受结核病评估的儿童中检测下呼吸道疾病。的总体目标 拟议的项目是评估Lung AI是否可以提高结核病治疗决策算法的准确性 儿童结核病的治疗我们假设，肺AI可以提高特异性，同时保持高灵敏度， 结核病治疗决策算法。为了评估这一假设，我们将利用1）现有的良好表征的 评估新型儿科结核病诊断的队列; 2）数字听诊器和 肺结核AI算法; 3）在高结核负担环境中评估数字健康工具的经验。在aim中 1，我们将使用现有数据检查新的和当前的结核病治疗决策算法的准确性 从乌干达、南非和冈比亚三个正在进行的儿童结核病诊断队列中收集。在 与此同时，我们将前瞻性地招募乌干达有结核病症状的儿童，进行全面的结核病评估， 根据NIH共识定义对结核病状态进行分类，并使用无线数字听诊器记录肺音。 在目标2中，我们将使用这些肺部声音来训练肺部AI算法，使用深度学习模型来识别较低的 参照标准化肺音定义和放射学CXR读数，儿童气道异常。 我们将在一组独立的结核病症状儿童和健康儿童的测试中评估这些模型。在 目标3，我们将创建一个肺AI模型，以检测儿童中微生物确认或临床诊断的结核病， 并将其准确性与1）性能最佳的TB治疗决策算法和2）机器学习算法进行比较。 在独立测试集中，结合肺部AI和临床变量的模型。实现这些目标将 确定结核病治疗决策算法的实用性，同时展示一种简单， 在医疗点提供经济实惠的数字医疗解决方案，以支持结核病和其他疾病的早期诊断和治疗 在较低级别的卫生设施中，儿童的呼吸道疾病。

PROJECT SUMMARY The majority of deaths from childhood tuberculosis (TB) are in children not initiated on treatment, highlighting the urgent need to prevent delays in diagnosis. However, the diagnosis of childhood TB is challenging, as sputum-based testing is invasive and has low yield in children, and chest X-ray (CXR) is not routinely available in high burden settings and requires trained readers. The World Health Organization (WHO) has therefore endorsed TB treatment decision algorithms to reduce delays in timely TB treatment initiation among children presenting to lower-level health facilities. However, TB treatment decision algorithms have uncertain accuracy and available data suggest they are likely to have poor specificity in the absence of CXR. Artificial-intelligence algorithms applied to lung sounds collected using a digital stethoscope (Lung AI) have the potential to enhance detection of lower respiratory tract disease in children being evaluated for TB. The overall objective of the proposed project is to evaluate whether Lung AI can improve the accuracy of TB treatment decision algorithms for childhood TB. We hypothesize that Lung AI can improve specificity while maintaining the high sensitivity of TB treatment decision algorithms. To assess this hypothesis, we will leverage 1) existing well-characterized cohorts to evaluate novel pediatric TB diagnostics; 2) expertise in digital stethoscopes and the development of Lung AI algorithms for TB; and 3) experience in assessing digital health tools in high TB burden settings. In Aim 1, we will examine the accuracy of new and current TB treatment decision algorithms using existing data collected from three ongoing childhood TB diagnostic cohorts in Uganda, South Africa, and the Gambia. At the same time, we will prospectively enroll children with TB symptoms in Uganda, perform a complete TB evaluation to classify TB status per NIH consensus definitions, and record lung sounds using a wireless digital stethoscope. In Aim 2, we will use these lung sounds to train Lung AI algorithms using deep learning models to identify lower airway abnormalities in children in reference to standardized lung sound definitions and radiology CXR reads. We will evaluate these models in an independent test set of children with TB symptoms and healthy children. In Aim 3, we will create a Lung AI model to detect microbiologically-confirmed or clinically-diagnosed TB in children, and compare its accuracy to 1) the best-performing TB treatment decision algorithm, and 2) a machine learning model that combines Lung AI and clinical variables, in the independent test set. Completion of these aims will determine the utility of TB treatment decision algorithms, while demonstrating the potential of a simple, affordable, digital health solution at the point-of-care to support the early diagnosis and treatment of TB and other respiratory diseases in children at lower-level health facilities.