Data_Sheet_1_Prediction of Delivery Within 7 Days After Diagnosis of Early Onset Preeclampsia Using Machine-Learning Models.docx

BackgroundEarly onset preeclampsia (eoPE) is a hypertensive disorder of pregnancy with endothelial dysfunction manifested before 34 weeks where expectant management is usually attempted. However, the timing of hospitalization, corticosteroids, and delivery remain a challenge. We aim to develop a prediction model using machine-learning tools for the need for delivery within 7 days of diagnosis (model D) and the risk of developing hemolysis, elevated liver enzymes, and low platelets (HELLP) syndrome or abruptio placentae (model HA).
Materials and MethodsA retrospective cohort of singleton pregnancies with eoPE and attempted expectant management between 2014 and 2020. A Mono-objective Genetic Algorithm based on supervised classification models was implemented to develop D and HA models. Maternal basal characteristics and data gathered during eoPE diagnosis: gestational age, blood pressure, platelets, creatinine, transaminases, angiogenesis biomarkers (soluble fms-like tyrosine kinase-1, placental growth factor), and ultrasound data were pooled for analysis. The most relevant variables were selected by bio-inspired algorithms. We developed basal models that solely included demographic characteristics of the patient (D1, HA1), and advanced models adding information available at diagnosis of eoPE (D2, HA2).
ResultsWe evaluated 215 eoPE cases and 47.9% required delivery within 7 days. The median time-to-delivery was 8 days. Basal models were better predicted by K-nearest-neighbor in D1, which had a diagnostic precision of 0.68 ± 0.09, with 63.6% sensitivity (Sn), 71.4% specificity (Sp), 70% positive predictive value (PPV), and 65.2% negative predictive value (NPV) using 13 variables and HA1 of 0.77 ± 0.09, 60.4% Sn, 80% Sp, 50% PPV, and 87.9% NPV. Models at diagnosis were better developed by support vector machine (SVM) using 18 variables, where D2’s precision improved to 0.79 ± 0.05 with 77.3% Sn, 80.1% Sp, 81.5% PPV, and 76.2% NPV, and HA2 had a precision of 0.79 ± 0.08 with 66.7% Sn, 82.8% Sp, 51.6% PPV, and 90.3% NPV.
ConclusionAt the time of diagnosis of eoPE, SVM with evolutionary feature selection process provides good predictive information of the need for delivery within 7 days and development of HELLP/abruptio placentae, using maternal characteristics and markers that can be obtained routinely. This information could be of value when assessing hospitalization and timing of antenatal corticosteroid administration.

研究背景
早发性子痫前期（early onset preeclampsia, eoPE）是一类妊娠相关高血压疾病，以孕34周前出现内皮功能障碍为特征，临床通常采取期待治疗策略。但住院时机、糖皮质激素使用时机及分娩时机的选择仍是临床亟待解决的难题。本研究旨在借助机器学习工具构建两类预测模型：模型D用于预测确诊后7天内需分娩的风险，模型HA用于预测发生溶血、肝酶升高、血小板减少综合征（hemolysis, elevated liver enzymes, low platelets syndrome, HELLP）或胎盘早剥（abruptio placentae）的风险。

材料与方法
本研究纳入2014至2020年间确诊为早发性子痫前期、接受期待治疗的单胎妊娠患者作为回顾性队列。本研究基于监督分类模型，采用单目标遗传算法构建模型D与模型HA。收集患者的基础临床特征及早发性子痫前期确诊时的相关检测数据，包括孕周、血压、血小板计数、肌酐、转氨酶、血管生成生物标志物（可溶性fms样酪氨酸激酶-1、胎盘生长因子）及超声检查数据，用于后续分析。通过仿生算法筛选出最具相关性的变量。本研究构建了仅纳入患者人口统计学特征的基础模型（D1、HA1），以及加入早发性子痫前期确诊时检测信息的进阶模型（D2、HA2）。

结果
本研究共纳入215例早发性子痫前期患者，其中47.9%的患者需在确诊后7天内分娩，患者的中位分娩间隔时间为8天。基础模型中，模型D1采用K近邻（K-nearest-neighbor）算法构建时预测效果最优，其诊断精度为0.68±0.09，灵敏度（sensitivity, Sn）为63.6%、特异度（specificity, Sp）为71.4%、阳性预测值（positive predictive value, PPV）为70%、阴性预测值（negative predictive value, NPV）为65.2%，共纳入13个变量；模型HA1的精度为0.77±0.09，灵敏度60.4%、特异度80%、阳性预测值50%、阴性预测值87.9%。确诊时构建的模型则以支持向量机（support vector machine, SVM）算法效果最佳，共纳入18个变量：模型D2的精度提升至0.79±0.05，灵敏度77.3%、特异度80.1%、阳性预测值81.5%、阴性预测值76.2%；模型HA2的精度为0.79±0.08，灵敏度66.7%、特异度82.8%、阳性预测值51.6%、阴性预测值90.3%。

结论
在早发性子痫前期确诊时，结合可常规获取的母体特征与生物标志物，采用进化特征筛选流程的支持向量机模型可较好预测7天内需分娩的需求及HELLP综合征/胎盘早剥的发生风险。该预测结果可为住院时机评估及产前糖皮质激素给药时机的临床决策提供参考价值。