Department of Statistics, University of Missouri, Columbia, Missouri, United States of America.
Department of Electrical Engineering & Computer Science, Center for Nano/Micro Systems & Nanotechnology, University of Missouri, Columbia, Missouri, United States of America.
PLoS One. 2022 Oct 25;17(10):e0275658. doi: 10.1371/journal.pone.0275658. eCollection 2022.
Tuberculosis is one of the top ten causes of death globally and the leading cause of death from a single infectious agent. Eradicating the Tuberculosis epidemic by 2030 is one of the top United Nations Sustainable Development Goals. Early diagnosis is essential to achieving this goal because it improves individual prognosis and reduces transmission rates of asymptomatic infected. We aim to support this goal by developing rapid and sensitive diagnostics using machine learning algorithms to minimize the need for expert intervention.
A single molecule fluorescence immunosorbent assay was used to detect Tuberculosis biomarker lipoarabinomannan from a set of twenty clinical patient samples and a control set of spiked human urine. Tuberculosis status was separately confirmed by GeneXpert MTB/RIF and cell culture. Two machine learning algorithms, an automatic and a semiautomatic model, were developed and trained by the calibrated lipoarabinomannan titration assay data and then tested against the ground truth patient data. The semiautomatic model differed from the automatic model by an expert review step in the former, which calibrated the lower threshold to determine single molecules from background noise. The semiautomatic model was found to provide 88.89% clinical sensitivity, while the automatic model resulted in 77.78% clinical sensitivity.
The semiautomatic model outperformed the automatic model in clinical sensitivity as a result of the expert intervention applied during calibration and both models vastly outperformed manual expert counting in terms of time-to-detection and completion of analysis. Meanwhile, the clinical sensitivity of the automatic model could be improved significantly with a larger training dataset. In short, semiautomatic, and automatic Gaussian Mixture Models have a place in supporting rapid detection of Tuberculosis in resource-limited settings without sacrificing clinical sensitivity.
结核病是全球十大死因之一，也是单一传染病病原体导致的主要死因。到 2030 年消除结核病流行是联合国可持续发展目标之一。早期诊断对于实现这一目标至关重要，因为它可以改善个体预后并降低无症状感染者的传播率。我们旨在通过使用机器学习算法开发快速而敏感的诊断方法来支持这一目标，从而最大限度地减少对专家干预的需求。
我们使用单分子荧光免疫吸附测定法从 20 例临床患者样本和一组人工添加的人尿对照样本中检测结核病生物标志物脂阿拉伯甘露聚糖。结核病状态分别通过 GeneXpert MTB/RIF 和细胞培养来确认。我们开发了两种机器学习算法，一种是自动模型，另一种是半自动模型，它们通过经过校准的脂阿拉伯甘露聚糖滴定测定数据进行训练，然后对真实患者数据进行测试。半自动模型与自动模型的不同之处在于前者有一个专家审查步骤，该步骤校准了较低的阈值，以确定来自背景噪声的单个分子。结果发现，半自动模型的临床灵敏度为 88.89%，而自动模型的临床灵敏度为 77.78%。
半自动模型通过在校准过程中应用专家干预，在临床灵敏度方面优于自动模型，而这两种模型在检测时间和分析完成时间方面都大大优于手动专家计数。同时，通过使用更大的训练数据集，自动模型的临床灵敏度可以得到显著提高。总之，半自动和自动高斯混合模型在不牺牲临床灵敏度的情况下，在资源有限的环境中支持快速检测结核病具有一定的地位。
