Use of a machine learning framework to predict severe maternal morbidity

严重孕产妇死亡率（SMM）在美国呈上升趋势。这种发病率是 伴随着分娩并发症和不良妊娠结局，并可能有长期的- 长期对女性健康的影响。到目前为止，只有少数研究检查了风险因素， 对于SMM在美国，甚至更少的人认为交付护理的网站可能 影响SMM的发生。这项研究将测试机器学习技术的使用， 开发预测女性罹患SMM风险的模型。我们将使用基于人口的 数据来自与美国医院协会链接的马里兰州数据库的一个家庭 2010-2014年年度调查数据。我们的主要分析样本将包括 2010-2014年期间马里兰州医院的所有分娩住院。两个SMM结果 将采取措施：疾病控制和预防中心（CDC）的SMM 算法，以及包括CDC SMM算法中的任何代码的复合度量， 在分娩住院期间入住ICU和/或输血。分别为每个 在两个结果测量中，我们将首先开发多阶段最小绝对收缩， 选择算子（LASSO）模型来预测SMM，然后采用多重加性 回归树最大化LASSO模型的预测能力。接下来，我们将拟合Logit SMM回归模型调整LASSO选择的预测变量，并比较 LASSO和Logit模型的性能使用标准指标，如灵敏度，特异性， 受试者操作者特征曲线下的面积。该提案有几个方面， 创新传统的分析工具并不适合捕捉大数据的全部价值。在 相比之下，机器学习技术不受预设统计假设的约束， 期望做出更高准确度的预测。这项试点研究的成功 这将为研究机器学习帮助临床护理的潜力开辟新的途径。 产科是可以通过预测产妇风险而从其使用中受益匪浅的领域之一 早期和优化途径，为妇女及其新生儿带来最佳结果。 确定SMM的关键预测因素有助于确定健康差异、优势和 这将有助于消除产科护理中的薄弱环节，并防止对孕产妇和新生儿产生不良后果。

Severe maternal morbidity (SMM) is on the rise in the United States. Such morbidity is accompanied by delivery complications and adverse pregnancy outcomes, and can have long- term health consequences for women. To date, only a handful of studies examined risk factors for SMM in the United States, and even fewer considered the site of delivery care as potentially influencing SMM occurrence. This study will test the use of machine learning techniques to develop models for predicting women’s risk of experiencing SMM. We will use population-based data from a family of Maryland state databases linked with American Hospital Association Annual Survey data for the 2010-2014 period. Our primary analytic sample will be comprised of all delivery hospitalizations in Maryland hospitals during 2010-2014. Two SMM outcome measures will be employed: Centers for Disease Control and Prevention (CDC)’s SMM algorithm, and a composite measure that includes any of the codes in the CDC SMM algorithm, ICU admission and/or blood transfusion during the delivery hospitalization. Separately for each of the two outcome measures, we will first develop multi-stage least absolute shrinkage and selection operator (LASSO) models to predict SMM and then employ Multiple Additive Regression Tree to maximize the predictive ability of the LASSO models. Next, we will fit Logit regression models for SMM adjusting for LASSO-selected predictor variables and compare LASSO and Logit models’ performance using standard metrics such as sensitivity, specificity, area under the curve of receiver operator characteristic. The proposal has several areas of innovation. Classical analytics tools are not well suited to capture the full value of large data. In contrast, machine learning techniques are unconstrained by preset statistical assumptions and expected to make predictions with higher degrees of accuracy. The success of this pilot study will open up new avenues of study into the potential for machine learning to aid clinical care. Obstetrics is one of the areas that can greatly benefit from its use by predicting maternal risks early and optimizing pathways to the best possible outcomes for women and their newborns. Identifying key predictors of SMM can serve to ascertain health disparities, strengths and weaknesses in obstetric care, and prevent adverse maternal and neonatal outcomes.