Application of Machine Learning Algorithms to Thiopurine Monitoring in IBD

描述（由申请人提供）：在日常医疗护理中，以血液计数、血液化学和其他生物标志物的形式收集大量数据。尽管投入巨大，但对这些数据的解释却少得可怜。在医学之外，从识别信用卡欺诈到推荐图书购买等各种应用中，机器学习技术推动了一场大型数据集分析的革命。尽管生物信息学在美国国立卫生研究院路线图倡议中占有突出地位，但这些显著的进步对医疗保健几乎没有影响。本提案的广泛、长期目标是优化、实施、测试和在全国范围内分发机器学习算法，这些算法将利用大型数据集中的模式来提高医学诊断和预后的准确性。我们建议使用硫嘌呤治疗炎症性肠病期间的免疫抑制监测作为示范案例。较低的治疗指数使得优化硫嘌呤剂量对炎症性肠病很重要，血清代谢物测定的益处有限。我们的初步数据表明，机器学习算法可用于大幅提高预后准确性并降低监测硫嘌呤使用的成本。该建议的中心假设是，血液计数和血液化学中存在与硫嘌呤药物有效免疫抑制相关的模式，可用于指导药物剂量。这一假设的基本原理基于两个观察结果。首先，我们的初步数据表明，机器学习可以通过分析实验室数据来识别免疫系统激活的重大变化。第二，已发表的数据表明，准确性较低的硫嘌呤代谢物试验在指导硫嘌呤剂量调整方面是合理有效的。本申请提出了一套改进的硫嘌呤监测算法的优化、实现和测试，并通过国家癌症研究所支持的LIDDEx（实验室信息数字数据交换）架构在全国范围内交付优化算法。本提案的具体目的是：(1)利用纵向临床数据和新颖的数学方法，以肠道炎症的客观证据为金标准，改进现有的硫嘌呤治疗临床反应的算法；(2)前瞻性检验硫嘌呤监测算法是否能准确分类免疫抑制的IBD患者和硫嘌呤药物不依从的IBD患者，以及这些算法是否能前瞻性指导患者的硫嘌呤给药；(3)使用LIDDEx网格架构在web服务器上实现这些修订后的算法，使全国临床使用，并在弗吉尼亚州安娜堡IBD诊所对该实现进行现场测试。拟议的研究将直接影响整个美国的患者护理，并通过展示该信息学架构的有效性，促进生物信息学在临床护理中的进一步创新和应用。

DESCRIPTION (provided by applicant): During routine medical care, enormous amounts of data are collected in the form of blood counts, blood chemistries, and other biomarkers. Despite this huge investment, remarkably little effort is applied to the interpretation of this data. Outside of medicine, a revolution in the analysis of large datasets has been driven by machine learning techniques in diverse applications ranging from identifying credit card fraud to making recommendations for book purchases. Despite the prominence of bioinformatics in the NIH Roadmap Initiative, these remarkable advances have had little impact on medical care. The broad, long-term objective of this proposal is to optimize, implement, test, and nationally distribute machine learning algorithms which will utilize patterns in large datasets to improve diagnostic and prognostic accuracy in medicine. We propose to use the monitoring of immune suppression during thiopurine therapy for inflammatory bowel disease as a demonstration case. A low therapeutic index makes it important to optimize thiopurine dosage for inflammatory bowel disease, and assays for serum metabolites are of limited benefit. Our preliminary data show that machine learning algorithms can be used to substantially improve prognostic accuracy and reduce costs in monitoring thiopurine use. The central hypothesis of this proposal is that there are patterns in the blood counts and blood chemistries associated with effective immune suppression by thiopurine medications which can be used to guide medication dosing. The rationale for this hypothesis is based on two observations. First, our preliminary data demonstrates that machine learning can identify significant changes in immune system activation through analysis of laboratory data. Second, published data suggests that the less accurate thiopurine metabolite tests are reasonably effective in guiding dose adjustment of thiopurines. This application proposes the optimization, implementation and testing of a improved set of thiopurine monitoring algorithms, and the nationwide delivery of the optimized algorithms through the National Cancer Institute-supported LIDDEx (Laboratory Information Digital Data Exchange) architecture. The specific aims of this proposal are to: (1) use longitudinal clinical data and novel mathematical methods to improve the existing algorithm for clinical response to thiopurine therapy, using objective evidence of bowel inflammation as the gold standard; (2) prospectively test whether the thiopurine monitoring algorithms can accurately classify IBD patients who are immunosuppressed and patients who are non-adherent to thiopurine medications, and whether these algorithms can prospectively guide dosing of thiopurines in patients; and (3) implement these revised algorithms on a web server using the LIDDEx grid architecture to enable nationwide clinical use, and field test this implementation in the Ann Arbor VA IBD clinic. The proposed studies will directly impact patient care throughout the United States, and by demonstrating the effectiveness of this informatics architecture, spur further innovation and application of bioinformatics to clinical care.