DSpace: Utilizing Data Science to Predict and Improve Health Outcomes in Pediatric HIV

摘要 代谢综合征（MetS）在撒哈拉以南非洲（SSA）感染艾滋病毒的儿童中迅速增加。 根据我们的初步数据，每30名16至19岁感染艾滋病毒的儿童中就有1名是 被诊断为代谢综合征此外，除了代谢综合征，结核病（TB）仍然是发病的主要原因， 艾滋病毒感染儿童的死亡率。此外，感染艾滋病毒的儿童患结核病的风险是其他儿童的30倍， 与未感染艾滋病毒的儿童相比，死亡风险明显更高。临床上，艾滋病毒感染者中的结核病 儿童表现出广泛的异质性（潜伏性结核病或活动性结核病[很可能、明确或可能]）， 对诊断提出了重大挑战。儿童结核病的少菌性意味着只有少量 一部分具有相容临床表现的儿童可以进行细菌学确认。已经 一直在努力开发数据科学工具，以解决患者分类和风险分层， MetS和提高结核病的诊断在成年西方人口。然而，这些技术并没有 非洲是艾滋病和结核病感染者负担最重的地区， 非传染性疾病的负担正在迅速增加。 此外，MetS是糖尿病（DM）早期发展的已知风险因素， 成年期心血管疾病（CVD）。不幸的是，干预措施（无论是药物还是非药物） 改善长期代谢障碍（MetS）儿童的代谢危险因素 不能完全预防或逆转CVD或DM并发症，这可能是当前时机的结果 这些干预措施是在代谢危险因素已经存在多年之后实施的。因此，在本发明中， 确定MetS的纵向风险变得势在必行。 同样，多组学数据的可用性为研究宿主遗传学提供了宝贵的机会 撒哈拉以南非洲儿童结核病的研究，以推动开发高度敏感的结核病诊断算法， 非常需要。因此，本应用程序的首要目标是利用数据科学方法， 将大时间电子健康记录（EHR）与多组学数据集成以预测和改善健康 非洲感染艾滋病毒儿童的结果。这项回顾性、描述性纵向研究将利用 来自贝勒国际儿科艾滋病倡议（BIPAI）的约118，000名艾滋病毒感染儿童的现有数据 在乌干达、博茨瓦纳和斯威士兰的项目。在目标1中，我们将使用机器学习来识别信息 纵向EHR和基因组数据中的特征来预测HIV感染儿童的MetS。我们亦会 制定与基于度鲁特韦的联合用药相关的MetS发生的综合风险评分 抗逆转录病毒疗法本提案的目标2将侧重于使用可解释的机器学习来揭示 多组学数据中的分子签名以及时间EHR中的特征性特征， HIV感染儿童结核病诊断的预测模型能力。这一努力将转化为 为结核病的诊断和未来发展制定临床相关的综合风险评分， 验证非痰TB诊断生物标志物。该应用程序提供了一个模型方法 框架，可应用于艾滋病毒感染儿童的多模态数据，并提高我们的理解 如何有效地利用人工智能来针对个性化或公共卫生干预措施， 在非洲艾滋病毒连续护理的整个范围内取得成果。

Abstract Metabolic Syndrome (MetS) is rapidly increasing in children infected with HIV in sub-Saharan Africa (SSA). According to our preliminary data, 1 in 30 children infected with HIV between the age of 16 and 19 are diagnosed with MetS. In addition, to MetS, tuberculosis (TB) remains a leading cause of morbidity and mortality among HIV-infected children. Moreover, children with HIV have a 30-fold risk of developing TB and a significantly higher risk of death compared to non-HIV-infected children. Clinically, TB in HIV-infected children manifests with extensive heterogeneity (latent TB or active TB [probable, definite, or possible]), which poses a significant diagnostic challenge. The paucibacillary nature of pediatric TB means that only a small fraction of children with a compatible clinical presentation can be bacteriologically confirmed. There have been various efforts to develop data science tools to address patient classification and risk stratification of MetS and improve the diagnosis of TB in adult Western populations. However, these technologies have not been deployed and evaluated in Africa, which bears the biggest burden of people infected with HIV and TB and where the burden of non-communicable diseases is growing rapidly. Furthermore, MetS is a known risk factor for the early development of diabetes mellitus (DM) and cardiovascular disease (CVD) in adulthood. Unfortunately, interventions (either pharmacological or non- pharmacological) that improve metabolic risk factors for children with long-term metabolic impairment (MetS) do not completely prevent or reverse CVD or DM complications, which may be the result of the current timing of interventions which are implemented after metabolic risk factors have been present for many years. Thus, the determination of the longitudinal risk of MetS becomes imperative. Similarly, the availability of multi-omics data presents a valuable opportunity to investigate the host genetics of TB disease in SSA children to advance the development of highly sensitive TB diagnostic algorithms that are much needed. Therefore, the overarching goal of this application is to utilize data science approaches to integrate large temporal electronic health records (EHR) with multi-omics data to predict and improve health outcomes of HIV-infected children in Africa. This retrospective, descriptive longitudinal study will leverage existing data on ~118,000 HIV-infected children from the Baylor International Pediatric AIDS Initiative (BIPAI) programs in Uganda, Botswana and Eswatini. In Aim 1, we will use machine learning to identify informative features within longitudinal EHRs and genomic data to predict MetS in HIV-infected children. We shall also develop composite risk scores for the development of MetS associated with dolutegravir-based combination antiretroviral therapy. Aim 2 of this proposal will focus on the use of explainable machine learning to uncover molecular signatures in multi-omics data as well as characteristic features in temporal EHR that improve the power of predictive models for the diagnosis of TB in HIV-infected children. This effort will translate into developing clinically relevant composite risk scores for the diagnosis of TB and the future development and validation of non-sputum TB diagnostic biomarkers. This application provides a model methodological framework that can be applied to multimodal data in HIV-infected children and improves our understanding of how to effectively use artificial intelligence to target personalized or public health interventions that improve outcomes across the entire spectrum of the HIV continuum care in Africa.