辅助生殖技术作为子痫前期的预测因子：当前临床预测模型的系统评价。

全文（开放获取） · 英文原文

【Abstract】 Assisted Reproductive Technology (ART) increases the risk of preeclampsia (PE), yet no specific prediction model tailored to the ART population exists. This review evaluates prediction models incorporating ART as a predictor, focusing on risk of bias, performance, generalizability and clinical applicability. A systematic MEDLINE search up to July 2023 identified models predicting PE that included ART as a predictor. Exclusion comprised studies on outcomes after PE onset, univariable models and validation studies of previously validated models. Risk of bias, applicability and generalizability were assessed using the Prediction model Risk Of Bias ASsessment Tool (PROBAST) and the Transparent Reporting of a multivariable prediction model for Individual Prognosis and Diagnosis (TRIPOD) guideline. Out of 15,305 records, 161 articles were assessed in full-text, yielding fifteen eligible studies. Eleven studies were added from a prior systematic review. Performance varied widely (AUC 0.58-0.99) and only few (7/26) models had been externally validated. Risk of bias was unclear in most studies (21/26), due to analytical issues. Despite ART being an established risk factor for PE, robust and generalizable models with proven external validity are lacking. Future research should prioritize development of a high quality prediction model including ART, or a model developed specifically within ART populations. 【Introduction】 Infertility affects 8–12% of couples worldwide. Assisted Reproductive Technology (ART), posing these couples an opportunity to conceive [], is becoming increasingly successful. Each year, over 3 million ART cycles are reported [], reflecting an ongoing expansion of indications. While ART has transformed the landscape of reproductive health, it has also introduced new challenges related to potential adverse obstetric outcomes. Studies indicate that pregnancies resulting from ART have a nearly doubled risk of hypertensive disorders of pregnancy, including preeclampsia (PE) [–], compared to naturally conceived pregnancies. PE, characterized by hypertension accompanied by proteinuria, maternal end-organ dysfunction, or uteroplacental dysfunction [], remains one of the leading causes of maternal mortality when left untreated or not recognized in time []. Moreover, hypertensive complications in pregnancy are associated with long-term maternal risks, including increased cardiovascular disease and decreased quality of life [–]. Additionally, severe PE increases the risk of fetal growth restriction and preterm birth due to underlying placental insufficiency [–], highlighting the need for timely identification and management, especially in ART pregnancies. Several theories have been proposed to explain the increased risk following ART. The need for ART itself may reflect underlying reproductive characteristics associated with an increased risk of PE, including advanced maternal age, nulliparity after prolonged infertility, and infertility-related diagnoses [, ]. Moreover, recent evidence suggest that specific ART related factors, such as programmed frozen embryo transfer cycles lacking a corpus luteum and its vasoactive products, and multiple gestation resulting from multiple embryo transfers, may causally contribute to PE risk []. Finally, altered maternal–fetal immune interactions, as reflected in oocyte donation, have been hypothesized to play a role in PE development [, , ]. While many prediction models for PE exist, none specifically address ART populations []. Initially, this review aimed to assess models developed exclusively for ART pregnancies, but their absence required a broader approach. Therefore, this review identifies prediction models for preeclampsia incorporating ART as a predictor and evaluates their risk of bias, performance, generalizability, applicability and reproducibility using TRIPOD and PROBAST. 【Materials and Methods】 This study was performed in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Statement guidelines []. This review is registered in PROSPERO (ID CRD420250477748). Registration was completed in 2025 after the process of study selection. No deviations from the registered protocol were made. 【Search Strategy and Study Selection】 A systematic literature search was conducted in MEDLINE, covering the period from January 2017 – July 31 st 2023, to update the search from De Kat et al. (2019) []. While De Kat et al. identified general prediction models for PE and selected only models with AUC ≥ 0.6, this review specifically focuses on models that included ART. The search strategy included free-text terms as ‘prediction model’ and ‘preeclampsia’. The complete search string is provided in Supplement 1. Two researchers (GL and EELOL) independently performed the literature search. Results were exported into a citation management software (Endnote), and articles were subsequently screened by both researchers. The screening process involved two stages: first, titles and abstracts were reviewed; second, full-text articles were analyzed. Any discrepancies between the researchers were resolved through consensus meetings. Additionally, the full-text of the studies included in the review by De Kat et al. [] were analyzed and incorporated into the final study selection if the prediction models presented included ART as predictor. 【Eligibility Criteria】 We included studies that developed (development only) or validated (validation only), or both developed and validated a prediction model for PE, with the definition as used by the authors and with ART as a predictor in the final model. If a validation study of a previously developed model within the inclusion period was available, we included both the development and corresponding validation study. ART was defined as any fertility-related treatment in which hormonal stimulation takes place []. For the current review, studies that reported mode of conception rather than explicitly identifying ART as a predictor were also included. Studies published in English were eligible for inclusion. In the first stage, all prediction models for PE were considered. In the second stage, we determined whether ART was included in the final model. Timing of use of the prediction model was preconceptionally or in early pregnancy. Exclusion criteria were studies focusing on outcome after PE onset, such as publications on the fullPIERS model []. We also excluded studies that focused solely on single variables, such as biomarkers [], without formulating a prediction model. Furthermore, studies testing the performance of previously externally validated models, such as those from the Fetal Medicine Foundation, were excluded unless they involved adding at least one new variable or extended the original model. Systematic reviews, case reports, guidelines and letters were also excluded. 【Risk of Bias Assessment】 The risk of bias assessment for this systematic review was performed using the Prediction model Risk Of Bias ASsessment Tool (PROBAST) [, ], a framework specifically designed for evaluating prediction models. PROBAST involves risk of bias assessment across four domains: participants, predictors, outcome and analysis. In addition, it assesses applicability of prediction models. PROBAST does not address confounding, as it is not relevant in prediction studies. Notably, De Kat et al. [] employed the Quality in Prognosis Studies (QUIPS) tool for their quality assessment. Although QUIPS evaluates six domains, it was not designed for scoring prediction models. Accordingly, the studies included from De Kat et al. [] were also assessed on bias using PROBAST. Risk of bias was rated low (+), unclear (?) or high (-). The overall rating reflects an aggregate assessment across all domains. For example, a study rated as “?” in the Analysis domain was still classified as overall “ + ” when the remaining domains were all rated “ + ”. Conversely, a study with “?” in the Analysis domain was rated overall “-” when the other domains were consistently “-”. For the PROBAST tool, studies were classified having a low risk of bias if appropriate data sources, such as cohort studies were used, and included participants who corresponded to unselected participants representative of the target population. Predictors were required to have clear and consistent definitions, assessed uniformly across all patients. Continuous and categorical predictors had to be handled appropriately, by avoiding data-driven categorization and properly addressing non-linearity. Outcome information should not have been used during predictor assessment, and all predictors should have been available at the time the model was intended to be used. Outcomes were determined objectively, following established guidelines or prior publications on the topic and were assessed uniformly across all participants without knowledge of predictor information. The time interval between predictor assessment and outcome determination needed to be appropriate to ensure the recording of relevant and representative outcomes. Studies were also required to have an adequate sample size, with at least 20 events per variable and continuous variables could not be categorized into two or more groups when included in the model. Missing data for predictors or outcomes had to be appropriately handled, either through explicit reporting of minimal missing data or the application of multiple imputation. Predictor selection based solely on univariable analysis was considered inappropriate. Complexities in the data, such as non-linearity or interactions, had to be identified and addressed, and model performance was required to be evaluated using metrics like calibration and discrimination. Internal validation techniques, including bootstrapping or cross-validation, were expected, with adjustments for optimism or overfitting. Finally, predictors and their assigned weights in the final model needed to align with the results of reported multivariable analysis. Validation could not proceed if any predictor data required by the model are missing, as this compromises the assessment of the model's performance and reliability. Applicability was rated as low (+), unclear (?) or high (-) concern regarding the relevance of the model to the review question. Judgements were based on whether the included participants and setting matched the review question, whether the definition, assessment or timing of predictors in the model match the review question, and whether the outcome, its definition, timing or determination match the review question. Detailed overview of domain specific assessments and derivation of the overall ratings are provided in Supplement 2. Risk of bias and applicability were independently assessed by GL and EELOL. Any discrepancies in scoring were discussed with a third observer (MLvdH) in a consensus meeting. 【Assessment of Reporting Quality】 The quality of reporting in the prediction models was assessed using the TRIPOD statement [], which includes a checklist of 22 items and 15 subitems focused on key reporting aspects of title/abstract, introduction, methods, results and discussion. GL and EELOL assessed each study, marking each item as ‘yes’ if it was sufficiently reported or ‘no’ if it was not. Any discrepancies in scoring were discussed with a third observer (MLvdH) in a consensus meeting. A reporting quality score was calculated based on the number of items that were met. Although this score is not a validated measure for prediction models, it has been used in other studies to summarize reporting quality [–]. 【Data Extraction】 GL extracted data from all selected articles, using the Checklist for Appraisal and data extraction for systematic Reviews on prediction Modelling Studies []. Data extraction comprised study design, participants, predictors and outcome, the timepoint at which the prediction model is intended to be applied, model development methods, model performance and validation statistics, and means for clinical application. In studies with multiple models, only those with ART in the final predictor selection were extracted. Additionally, the impact of ART and its significance to the model's accuracy and predictive power were assessed. Data were synthesized narratively and presented in tables. 【Model Performance】 As model performance measures, metrics of discrimination and calibration were extracted. Discrimination measures the model’s ability to distinguish between patients who experienced the outcome and those who did not, typically assessed using the concordance statistic (c-statistic), which represents the area under the receiver operating characteristic (ROC) curve. A c-statistic of 1 indicates perfect discrimination, while a value of 0.5 indicates no better than random guessing. Calibration assesses the agreement between predicted probabilities and observed outcome frequencies, often visualized with a calibration plot that compares observed frequencies against predicted probabilities []. In case c-statistics and calibration slopes were not provided, detection rates (DR) and false positive rates (FR) were reported. 【Results】 【Study Selection】 The study selection process is illustrated according to the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) flowchart [] (Fig. ). The initial search yielded 15,305 records. After removing duplicates, 14,682 articles advanced to the first screening stage, where titles and abstract were assessed. This resulted in the exclusion of 14,521 articles, leaving 161 for full-text assessment in the second stage. Some studies proposed a model using Bayesian analysis [] with prior risks based on ‘prior risk factors’ or ‘maternal factors’, though these factors were not always specified [–]. Following correspondence with the authors, studies in which ART was not considered a prior ‘maternal factor’, or where authors did not respond, were excluded [–]. Ultimately, 16 studies from this search met the eligibility criteria [–]. Preferred reporting items for systematic reviews and meta-analyses flowchart Additionally, full-text screening of the 70 studies included in the research by De Kat et al. [] identified 10 extra studies with ART defined as predictor [–]. Consequently, a total of 26 articles were included in the current review. 【Study Characteristics, Sample Size and Modelling Method】 Studies were published between 2009 and 2023. Eleven studies were conducted across multiple centers [, , , , , , , , , , ]. The number of participants ranged from 94 [] to 141.299 [], with in total 625.001 participants of whom 20.142 developed PE. None of the studies performed a sample size calculation prior to model development. The majority of studies was conducted in Europe [, , , , , –, –], six studies in Asia [, , , , , ] and one in North America []. Key elements of the included studies are summarized in Tables and , which distinguish between Bayesian and frequentist models. In prediction model development and validation, frequentist methods estimate parameters based on observations only, assuming true parameters are fixed but unknown. These methods use metrics like confidence intervals and p-values. In contrast, Bayesian methods incorporate prior knowledge, updating beliefs as new data become available, resulting in probability distributions for model parameters []. Summary of the key elements of the studies using Bayes’ theorem combining the a-priori risk from maternal factors with biomarkers Poon [] 2009 UK GH‰; DBP ≥ 90 mmHg in ≥ 4 h, GA ≥ 20; PE‰; GH‰ with proteinuria ≥ 300 mg/24 h or ≥ 2 + on dipstick Development only Multivariate logistic regression Early PE AUC = 0.7 Late PE AUC = 0.80 Wright [] 2012 UK PE < 34wks Maternal factors Maternal factors, UtA-PI Maternal factors, MAP Combined testing PE < 37wks Maternal factors Maternal factors, UtA-PI Maternal factors, MAP Combined testing PE < 42wks Maternal factors Maternal factors, UtA-PI Maternal factors, MAP Combined testing Development only Competing risk PE < 34wks (FPR 10%) DR = 50.5 DR = 75.2 DR = 72.9 DR = 89.7 PE < 37wks (FPR 10%) DR = 43.3 DR = 55.1 DR = 59.3 DR = 71.5 PE < 42wks (FPR 10%) DR = 40.3 DR = 42.2 DR = 53.5 DR = 56.6 Khalil [] 2012 UK Cohort multicenter Maternal factors Maternal factors, vascular derived risk Maternal factors, vascular derived risk, UtA-PI, PAPP-A Development only Multivariate logistic regression AUC = 0.80 AUC = 0.84 AUC = 0.85 Park [] 2013 UK GH¥, PE¥ (early/late); > 140/90 2 × with ≥ 300 mg/24 h or ≥ 30 mg protein/mmol creatinin Early PE AUC = 0.93 Late PE AUC 0.76 Skråstad [] 2015 Norway PE¶; (SBP ≥ 140 mmHg and/or DBP ≥ 90 mmHg in > 1x ≥ 4–6 h, GA ≥ 20, with proteinuria (≥ 300 mg/24 h or ≥ 2 + on dipstick) Preterm PE¶; < GA 370 FMF Maternal characteristics, MoM-values of the mean of left and right UtAPI, MAP PAPP-A and PlGF; Perkin Elmer (PREDICTOR) Prior risk (based on BMI, ethnicity, parity, family history of preeclampsia, chronic hypertension and MAP) Posterior risk (based on prior risk and MoMs of MAP, UtAPI, PlGF and PAPPA) FMF [] Maternal characteristics, MoM-values of the mean of left and right UtAPI, MAP PAPP-A and PlGF; PREDICTOR [] ‘Maternal characteristics’, MAP, UtA-PI, PlGF, PAPP-A Validation only [, ] Multivariate logistic regression PE < GA42: AUC = 0.77 PE < GA37: AUC = 0.94 Gestational age at delivery with PE PE; BP ≥ 140/90, ≥ 2x, > GA20 with proteinuria ≥ 300 mg/24 h or ≥ 2 + on dipstick MF [] MF, MAP MF, UtA-PI MF, PAPP-A MF, PlGF MF, MAP, UtA-PI MF, MAP, PAPP-A MF, MAP, PlGF MF, UtA-PI, PAPP-A MF, UtA-PI, PlGF MF, PlGF, PAPP-A MF, MAP, UtA-PI, PAPP-A MF, MAP, PAPP-A, PlGF MF, MAP, UtA-PI, PlGF MF, UtA-PI, PAPP-A, PlGF MF, MAP, UtA-PI, PAPP-A, PlGF Validation only [] Multivariable Gaussian distribution 5 × cross validation AUC = 0.80 AUC = 0.845 AUC = 0.841 AUC = 0.822 AUC = 0.872 AUC = 0.876 AUC = 0.860 AUC = 0.896 AUC = 0.851 AUC = 0.884 AUC = 0.873 AUC = 0.884 AUC = 0.897 AUC = 0.906 AUC = 0.885 AUC = 0.907 PE GH¥, PE¥ (early/late); BP ≥ 140/90, ≥ 2x, > GA20 with proteinuria ≥ 300 mg/24 h or ≥ 2 + on dipstick Cohort multicenter MF [] MF, MAP MF, UtA-PI MF, PAPP-A MF, MAP, UtA-PI MF, MAP, PAPP-A MF, MAP, PlGF MF, UtA-PI, PAPP-A MF, UtA-PI, PlGF MF, PlGF, PAPP-A MF, MAP, UtI-PI, PAPP-A MF, MAP, PAPP-A, PlGF MF, MAP, UtA-PI, PlGF MF, UtA-PI, PAPP-A, PlGF MF, MAP, UtA-PI, PAPP-A, PlGF AUC = 0.9071 AUC = 0.9309 AUC = 0.8546 AUC = 0.9506 AUC = 0.9667 AUC = 0.9133 AUC = 0.9674 AUC = 0.9339 AUC = 0.9772 AUC = 0.9510 AUC = 0.9644 AUC = 0.9672 AUC = 0.9870 AUC = 0.9769 AUC = 0.9865 Maymon [] 2017 Israël Major fetal anomaly, nuchal translucency ≥ 3.5 mm, fetal reduction of high order multiplicity, chromosomal abnormalities, any maternal disease that could affect fetal growth, treatment with LMWH due to thrombophilia AUC = 0.981 DR = 75% for 10% FPR Francisco [] 2017 UK GA of delivery with PE¥ (BP ≥ 140/90, ≥ 2x, > GA20 with proteinuria ≥ 300 mg/24 h or ≥ 2 + on dipstick) Development only Competing risk model/Gaussian distribution Risk cut-off 1/10: Singleton SPR: 0.8 DR PE < 37: 16.1% DR PE > 37: 12.7% Francisco [] 2017 UK GA of delivery with PE¥ (BP ≥ 140/