Personalized Postpartum Hemorrhage Prediction Using Machine Learning And Polygenic Risk Scores

摘要 产后出血，定义为分娩后24小时内估计至少失血1000毫升，是 导致严重孕产妇发病率和死亡率的主要原因。每年，产后出血并发2-3% 占所有怀孕人数的1/4，占全球产妇死亡人数的14万人。在美国，还有 显著的种族差异：与黑人女性相比，黑人女性与出血相关的死亡风险高出五倍 非黑人女性。虽然临床产后出血风险预测工具已经开发出来，但它们未能 识别高达40%的病例；因此，目前临床上还没有广泛采用的基于证据的预测工具 练习一下。因此，迫切需要一种高效、准确、个性化的产后出血风险预测工具 需要的。最近，机器学习方法被越来越多地用于开发准确的预测 与传统统计方法相比具有更好性能的模型，并发现新的 预报器，几乎没有预先指定的内容。此外，可解释的机器学习方法允许 决策透明，减少偏见。通过这种方式，机器学习模型可能会带来更多 比目前现有的工具更准确地预测产后出血。此外，由于高达18%的 产后出血的风险是家族性的，许多临床风险因素与产后相关 出血具有良好的多基因架构，使用多基因风险工具可能会进一步增强 产后出血风险预测。与美国国立卫生研究院提高产妇安全的倡议目标保持一致 和结果，我们建议开发一种高保真算法，结合临床和遗传因素， 以更准确地预测孕妇产后出血的风险。我们将利用我们的富豪 患者数据库和最先进的计算工具：(1)开发改进的算法对患者进行分层 产后出血风险，重点放在透明度和减少偏见上，以及(2)描述贡献 遗传因素对产后出血风险的影响。总体而言，这个项目将提高我们准确预测的能力 产后出血高危患者与新的预测因素的研究，临床之间的相互作用 和遗传贡献者，以及机器学习和多基因风险评分的新应用 结果。最终，我们的目标是在临床实践中验证和实施这些工具，从而极大地 加强预防孕产妇发病率和死亡率的能力。通过实现这些目标，我将制定一项 建立我的研究轨迹并过渡到作为一名医生的独立所必需的特定技能- 利用转换计算方法预测和改善产科不良结局的科学家。

ABSTRACT Postpartum hemorrhage, defined as estimated blood loss of at least 1000 mL within 24 hours of delivery, is the leading cause for severe maternal morbidity and mortality. Annually, postpartum hemorrhage complicates 2-3% of all pregnancies and accounts for 140,000 maternal deaths globally. In the United States, there are also significant racial disparities: Black women have a five-fold higher risk of hemorrhage-related death compared to non-Black women. While clinical postpartum hemorrhage risk prediction tools have been developed, they fail to identify up to 40% of cases; as a result, no evidence-based prediction tool is currently widely adopted in clinical practice. Thus, an efficient, precise, and personalized postpartum hemorrhage risk prediction tool is urgently needed. Recently, machine learning approaches have been increasingly used to develop accurate predictive models with superior performance compared to the traditional statistical approaches and to discover new predictors, with little prior pre-specification. Moreover, the explainable machine learning methods allow for transparent decision making and reduction of bias. In this way, machine learning models may lead to more accurate postpartum hemorrhage prediction than currently existing tools. In addition, since up to 18% of postpartum hemorrhage risk is familial and many of the clinical risk factors associated with postpartum hemorrhage have a well-established polygenic architecture, using polygenic risk tools may further enhance postpartum hemorrhage risk prediction. In line with the NIH IMPROVE initiative goals to improve maternal safety and outcomes, we propose here to develop a high-fidelity algorithm, combining both clinical and genetic factors, to more accurately predict the risk of postpartum hemorrhage in pregnant individuals. We will leverage our rich patient database and state-of-the-art computational tools to: (1) develop an improved algorithm to stratify patient postpartum hemorrhage risk with a focus on transparency and bias reduction, and (2) delineate the contribution of the genetics to postpartum hemorrhage risk. Overall, this project will advance our ability to precisely predict patients at risk for postpartum hemorrhage, with the investigation of novel predictors, interaction between clinical and genetic contributors, and novel application of both machine learning and polygenic risk scores to these outcomes. Ultimately, we aim to validate and implement these tools in clinical practice, leading to greatly enhanced ability to prevent maternal morbidity and mortality. By completion of these aims, I will develop a specific skill set essential for establishing my research trajectory and transition to independence as a physician- scientist utilizing translational computational approaches to predict and improve adverse obstetric outcomes.