CAREER: Personalized Maternal Care Decision Support System for Underserved Populations

妇女在分娩和怀孕期间的死亡率，即产妇死亡率，被认为是衡量人口健康、妇女在社会中的健康状况以及医疗保健系统本身的整体健康状况的重要指标。然而，在过去的20年里，美国的孕产妇死亡率出现了令人担忧的上升，导致美国的孕产妇死亡率在发达国家中最高。先兆子痫是一种与高血压有关的妊娠并发症。每年，美国8%-10%的孕妇患有先兆子痫，除非在怀孕早期发现并进行治疗，否则可能导致产妇和/或新生儿死亡。确定先兆子痫风险较高的妇女仍然是一个挑战，因为几个因素，特别是年龄、种族和孕前疾病史，可能有助于这种情况的发展。该项目将建立创新的技术，使计算机能够了解和预测妇女在怀孕期间患先兆子痫的可能性，特别是在少数族裔群体中。建立这样一个系统需要大量的数据，从人口统计到个人健康记录，以训练计算机预测先兆子痫。该项目的主要创新之处在于，它能够从临床数据中学习，这些数据往往不完美，记录缺失或不完整，有关子痫前期病例的信息可能很少，并且在预测先兆子痫的风险时，能够对包括美洲原住民在内的不同种族群体的亚群保持公平。该项目中开发的技术还将有可能帮助建立有助于早期发现其他疾病的工具。本项目致力于开发新的基于机器学习(ML)的临床决策辅助工具，用于子痫前期(PE)的早期检测。该项目的主要新颖性在于它能够有效地解决从PE数据集学习的几个具体问题，这些问题如果得不到解决，将继续阻碍基于ML的PE早期检测的临床实施：(挑战一)PE数据集经常面临固有的类别不平衡；(挑战II)为PE早期检测构建可靠的ML模型需要挖掘大型和多样化的数据集，如电子健康记录，这对现有ML模型的可扩展性构成了重大挑战；和(挑战III)PE对某些种族群体的影响不成比例，特别是美国印第安人/美洲原住民妇女，由于这些差异，将这种ML模型的公平性变成了一个伦理问题，并对采用ML进行疾病检测构成了挑战。为了应对这些挑战，研究者将(1)开发一类新的无参数分类器来有效地解决由类不平衡引起的偏差，从而消除计算代价昂贵的超参数调整的需要，这是针对类不平衡的代价敏感学习模型的常见问题；(2)通过将学习任务描述为一个顺序决策过程，指导分类中的数据采样，开发一种新的可扩展的分类方法，用于从大规模PE数据集中学习；(3)基于平衡公平性和准确性的可处理表优化模型以及新的性能公平度量来开发一类公平分类器，以同时衡量不平衡数据的公平性和准确性。研究人员进一步研究了在可处理海量数据的新型可扩展框架内将公平ML模型应用于在线学习环境。建议的基于ML的PE检测模型的成功实施将增强对先兆子痫高危孕妇的识别，同时减少相关孕产妇健康管理系统中的种族偏见。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The rate of women dying in childbirth and pregnancy, maternal mortality, is recognized as a crucial indicator of population health, the status of women's health in the society, and the overall health of the healthcare system itself. However, the US has experienced a worrying increase in maternal mortality over the last two decades, resulting in the US reaching the highest rate among developed countries. Preeclampsia is a pregnancy complication related to high blood pressure. Each year, preeclampsia afflicts 8-10% of US pregnancies and can lead to maternal and/or neonatal death unless it is detected and treated in early stages of the pregnancy. It remains a challenge to identify women at higher risk of preeclampsia, as several factors, notably age, race, and the history of pre-pregnancy diseases, can contribute to developing the condition. This project will build innovative technologies to allow computers to understand and predict the likelihood of a woman developing preeclampsia during pregnancy, particularly among women from minority racial groups. Building such a system requires massive data, ranging from demographic to individual health records, to train the computers to predict preeclampsia. The main novelty of this project is in its capacity to learn from clinical data that are often imperfect, suffering from missing or incomplete records with possibly very little information on preeclampsia cases, and to remain fair toward subpopulations of various racial groups, including Native Americans, when predicting the risk of preeclampsia. The technologies developed in this project will also have the potential to help build tools that can help in early detection of other diseases. This project investigates developing novel machine learning (ML)-based clinical decision aid tools for early detection of preeclampsia (PE). The main novelty of this project is in its capacity to effectively address several issues specific to learning from PE datasets that, if not addressed, continue to impede the clinical implementation of ML-based early detection of PE: (Challenge I) PE datasets often face inherent class imbalance; (Challenge II) Constructing reliable ML models for early PE detection necessitates mining large and diverse datasets, such as electronic health records, posing a significant challenge to the scalability of existing ML models; and (Challenge III) PE disproportionately affects certain racial groups, notably American Indian/Native American women, turning the fairness of such ML models into an ethical concern due to these disparities, and posing a challenge in adopting ML for disease detection. In response to these challenges, the investigator will (1) develop a new class of parameter-free classifiers to effectively address the bias resulting from class imbalance, thus eliminating the need for computationally expensive hyperparameter tuning, a common issue with cost-sensitive learning models for class imbalance; (2) develop a novel scalable classification method for learning from large-scale PE datasets through formulating the learning task as a sequential decision-making process, guiding data sampling in classification; and (3) develop a class of fair classifiers based on tractable optimization models that balance fairness and accuracy as well as novel performance-fairness metrics to simultaneously measure fairness and accuracy for imbalanced data. The investigator further studies adapting the fair ML model for online learning settings within a novel scalable framework that can handle massive data. Successful implementation of the proposed ML-based PE detection models will enhance identification of pregnant women at a high risk of preeclampsia, while reducing racial biases in relevant maternal health management systems.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.