Sage Journals: Discover world-class research

Abstract

Background:

The placenta accreta spectrum (PAS) represents a significant risk factor for severe postpartum hemorrhage. Recent studies have demonstrated the safety of neuraxial anesthesia (NA) in cesarean delivery (CD) for patients with PAS.

Objectives:

To evaluate the risk of severe peripartum hemorrhage in patients with PAS who underwent CD under NA.

Design:

A multicenter retrospective cohort study.

Methods:

This study analyzed 214 patients diagnosed with PAS. Logistic regression was used to identify factors increasing the risk of severe peripartum hemorrhage. A total of six machine learning (ML) algorithms were employed for model validation.

Results:

The predictive model includes the following risk factors: age at delivery >33 years (p = 0.004), history of CD >1 (p = 0.020), preoperative HGB ⩽ 100 g/L (p = 0.013), placenta previa classification (p = 0.001), vascular lacunae within the placenta (p = 0.015), and labor duration (p = 0.026). The validation of ML algorithms revealed that the model achieved an accuracy ranging from 0.68 to 0.71, with an area under the receiver operating characteristic curve between 0.75 and 0.79. A nomogram list and web-based calculator were constructed for clinical implementation, and a risk stratification system was established based on model scores.

Conclusion:

A prenatal risk assessment model was developed to estimate the likelihood of severe peripartum hemorrhage in PAS patients undergoing CD under NA. This model may provide preliminary support for clinicians in tailoring anesthetic management strategies for potentially high-risk cases, but further studies are needed to confirm its clinical utility.

Keywords

cesarean delivery machine learning neuraxial anesthesia placenta accreta spectrum predictive model

Introduction

Placenta accreta spectrum (PAS) disorders represent severe conditions characterized by the abnormal attachment of the placenta to the uterine wall.¹ Major risk factors contributing to PAS disorders include previous cesarean delivery (CD), uterine curettage, placenta previa, and conception through in vitro fertilization.^2–4 In the past few decades, there has been a notable global increase in CD rates, rising from below 10% to over 30%.^5,6 This increase has been accompanied by a 10-fold escalation in the incidence of PAS disorders.

Pregnant individuals with PAS disorders are at an increased risk of experiencing severe complications during delivery. These complications include severe peripartum hemorrhage, disseminated intravascular coagulation (DIC), and potential organ damage, all of which can contribute to preventable maternal mortality.^2,7,8 Considering the high stakes, there is currently no consensus on the optimal anesthesia approach for PAS disorders. Neuraxial anesthesia (NA) is generally favored for routine CD due to benefits like reduced fetal exposure to anesthetics, fewer maternal airway interventions, and a more active labor experience. However, in PAS cases, anesthesiologists often lean toward general anesthesia (GA) owing to the risks of significant blood loss and hemodynamic instability associated with the sympathectomy effect of NA.⁸ Recent research, however, has shown that NA can be successfully applied in a considerable number of PAS disorder cases.^9,10 This highlights the need for larger-scale studies to thoroughly evaluate the safety and effectiveness of NA in the management of PAS disorders.

Therefore, the primary objective of this multicenter retrospective study is to identify risk factors associated with severe peripartum hemorrhage of PAS patients during CD under NA and to develop a predictive model. This model aims to assist clinicians in optimizing anesthesia management strategies, thereby improving perinatal outcomes.

Methods

Ethics

Ethical approval for this study (KYLL-202209-028) was provided by the Ethics Committee of Qilu Hospital, Shandong University, Shandong, China on October 14, 2022, which waived the requirement for written informed consent. To ensure privacy protection, all patient data were anonymized prior to analysis. We followed the TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis) when preparing this article (Supplemental Material).¹¹

General information

Data were collected on 214 pregnant patients diagnosed with PAS from four institutions: Qilu Hospital of Shandong University, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan Maternity and Child Care Hospital, and Qilu Hospital of Shandong University Dezhou Hospital. The data collection period extended from January 2010 to January 2022.

The diagnostic criteria for PAS in this study included the following: (1) difficulty or inability to deliver the placenta manually, coupled with an indiscernible cleavage plane between the placenta and myometrium; (2) significant bleeding at the placental site following challenging manual delivery; (3) histological confirmation of placental anomalies through uterine or placental specimens obtained during hysterectomy; and (4) prenatally diagnosed PAS with Doppler ultrasound or magnetic resonance imaging (MRI) that was subsequently confirmed at the time of delivery.¹² The types of PAS included in this study were accreta, increta, and percreta.

This multicenter retrospective study included patients who met the following criteria: (1) a confirmed diagnosis of PAS disorders; (2) undergoing CD with NA, which encompassed spinal, epidural, combined spinal-epidural anesthesia, and cases where NA was converted to GA intraoperatively; and (3) availability of comprehensive clinical data. The choice of anesthesia method was determined based on the patient’s medical condition, the anesthesiologist’s preference, and the obstetrician’s surgical approach.

Data collection

Patient data were meticulously extracted from electronic medical records, which included a range of variables such as delivery age, gestational weeks, pregnancy history, incidence of preoperative vaginal bleeding, results of laboratory tests such as hemoglobin (HGB) and platelets, outcomes of antenatal ultrasound examinations, the volume of intraoperative blood loss, and the units of transfused packed red blood cells (PRBC).

Definitions and outcomes

In accordance with the screening criteria for severe peripartum hemorrhage as outlined by the American College of Obstetricians and Gynecologists and the Society for Maternal-Fetal Medicine, and informed by additional studies, the criterion adopted in this study for severe peripartum hemorrhage was defined as either intraoperative blood loss of 1500 mL or more or the transfusion of four or more units of PRBC.^13,14 This definition was pivotal in developing predictive models focused on significant intraoperative blood loss and transfusion instances in PAS patients undergoing CD with NA.

Statistical analysis

Median values were adopted as cutoff points for classifying continuous variables. The Chi-square test or Fisher’s exact test was applied to categorical variables for the initial screening of potential risk factors for severe peripartum hemorrhage. Variables with a p value less than 0.10 were subsequently included in a multivariate logistic regression analysis to construct the risk prediction model. This analysis yielded odds ratios, 95% confidence intervals, and p values.

Machine learning (ML) algorithms have been shown to outperform traditional regression methods in developing prognostic models, as evidenced by several recent studies.^15–17 Six commonly used ML algorithms—logistic regression, random forest, AdaBoost, decision tree, k-nearest neighbor, and naïve Bayes—were applied in this study to evaluate the performance of the developed model. We selected logistic regression for its simplicity and interpretability, random forest for handling high-dimensional data and complex interactions, AdaBoost for enhancing model performance and robustness, decision tree for its intuitive structure, k-nearest neighbor for capturing local data structures in nonlinear relationships, and naïve Bayes for its computational efficiency and effectiveness with smaller datasets.^18–23 These algorithms were chosen to provide a comprehensive comparison across different modeling approaches, ensuring robustness and reliability in our predictive model. Due to the size of the dataset, a fivefold cross-validation scheme was implemented, better suited for smaller datasets. The dataset was randomly divided into five equal parts, each maintaining a similar distribution of adverse events as the entire dataset. Four subsets were used for training and one for validation. This process was repeated 100 times. Results are presented as mean ± standard deviation. Model performance was evaluated by calculating the average and standard deviation of the receiver operating characteristic curves (ROC) and area under the curve (AUC) across multiple cross-validation rounds. The ROC curve illustrated the variation of the true-positive rate against the false-positive rate.²⁴ An AUC baseline of 0.5 was used as a reference, with higher values indicating better discrimination efficiency of the model. Sensitivity, specificity, positive predictive value, negative predictive value, and overall accuracy were also computed.

Subsequently, the model was visually presented. Employing the independent risk factors identified by the analysis, a nomogram was developed to predict the individual likelihood of severe peripartum hemorrhage in patients. A nomogram is a visual tool that simplifies complex regression results, making calculations more precise.²⁵ In addition, a user-friendly web calculator was created using the SHINY platform (available at https://shiny.posit.co) is primarily used for building interactive web apps for data science, statistics, and analytics. The calibration curve and decision curve analysis (DCA) were performed to evaluate the calibration ability and clinical utility of the model.²⁶

In addition, a risk stratification analysis was conducted to assess the likelihood of severe perinatal hemorrhage. First, total risk scores were calculated by summing individual risk factor scores from the nomogram. Risk stratification was performed by categorizing patients into low-, intermediate-, and high-risk groups based on score tertiles (33rd and 66th percentiles). Box plots were used to display the risk scores of patients in low-, intermediate-, and high-risk groups. Bar charts were used to show the total number of patients, the number of patients with adverse complications, and those without adverse complications in each risk group.

Statistical analyses were performed using IBM SPSS Statistics (version 26.0; IBM Corp., Armonk, NY, USA), R software (version 4.1.2; R Foundation for Statistical Computing, Vienna, Austria, https://www.r-project.org), and the Python machine learning library Scikit-learn (compatible with Python 3.6.4; an open-source library maintained by the community, https://scikit-learn.org).

Results

Figure 1 is the flow chart of this study. Table 1 summarizes the clinical characteristics of the 214 patients included in the study. Of the patients, 119 (55.6%) were ⩽33 years old, 32 (15.0%) had a history of more than one CD, preoperative HGB levels were below 100 g/L in 126 (58.9%) patients, and preoperative platelet levels were below 100 × 10⁹/L in 5 (2.3%) patients. Preoperative vaginal bleeding was observed in 79 (36.9%) patients. Among them, 44 (44/214, 20.6%) experienced intraoperative blood loss of ⩾1500 mL, and 60 (60/214, 28.0%) required transfusions of ⩾4 units of PRBC.

Figure 1.

Flowchart of the study.

Table 1.

Preoperative characteristics of patients.

Characteristics	Total (n = 214)	Adverse event (n = 67)	No adverse event (n = 147)	p Value
Age at delivery (years)				0.014
⩽33	119 (55.6)	29 (43.3)	90 (61.2)
>33	95 (44.4)	38 (56.7)	57 (38.8)
Gravidity				0.063
⩽3	119 (55.6)	31 (46.3)	88 (59.9)
>3	95 (44.4)	36 (53.7)	59 (40.1)
History of CD				0.004
⩽1	182 (85.0)	50 (74.6)	132 (89.8)
>1	32 (15.0)	17 (25.4)	15 (10.2)
History of curettage of uterine				0.168
⩽1	170 (79.4)	57 (85.1)	113 (76.9)
>1	44 (20.6)	10 (14.9)	34 (23.1)
Preoperative HGB level (g/L)				<0.001
⩽100	126 (58.9)	53 (79.1)	73 (49.7)
>100	88 (41.1)	14 (20.9)	74 (50.3)
Preoperative platelets level (10⁹/L)				0.178
⩽100	5 (2.3)	3 (4.5)	2 (1.4)
>100	209 (97.7)	64 (95.5)	145 (98.6)
Obstetric complications
Pregnancy-induced hypertension	14 (6.5)	2 (3.0)	12 (8.2)	0.234
Gestational diabetes mellitus	58 (27.1)	12 (17.9)	46 (31.3)	0.041
Placenta previa classification				<0.001
Marginal	24 (11.2)	9 (13.4)	15 (10.2)
Partial	38 (17.8)	2 (3.0)	36 (24.5)
Complete	127 (59.3)	55 (82.1)	72 (49.0)
None	25 (11.7)	1 (1.5)	24 (16.3)
Prenatal ultrasound results
Retroplacental myometrial thickness <1 mm	58 (27.1)	33 (49.3)	25 (17.0)	<0.001
Vascular lacunae within the placenta	40 (18.7)	22 (32.8)	18 (12.2)	<0.001
Hypervascularity of uterine–placental margin	64 (29.9)	25 (37.3)	39 (26.5)	0.110
Irregularity of uterine–bladder interface	19 (8.9)	10 (14.9)	9 (6.1)	0.036
Hypervascularity of the uterine serosa–bladder wall interface	17 (7.9)	9 (13.4)	8 (5.4)	0.045
Hypervascularity of the cervix	8 (3.7)	6 (9.0)	2 (1.4)	0.013
Dexamethasone practice to promote fetal lung	92 (43.0)	38 (56.7)	54 (36.7)	0.006
Preoperative vaginal bleeding	79 (36.9)	31 (46.3)	48 (32.7)	0.056
Emergency cesarean delivery	32 (15.0)	9 (13.4)	23 (15.6)	0.674
Labor duration (weeks)				0.002
24 ⩽ GW < 34	30 (14.0)	15 (22.4)	15 (10.2)
34 ⩽ GW < 37	80 (37.4)	31 (46.3)	49 (33.3)
37 ⩽ GW	104 (48.6)	21 (31.3)	83 (56.5)
PAS classification (postpartum diagnosis)				—
Accreta	68 (31.8)	13 (14.6)	55 (44.0)
Increta	128 (59.8)	66 (74.2)	62 (49.6)
Percreta	18 (8.4)	10 (11.2)	8 (6.4)
Adverse event (postpartum diagnosis)				—
Intraoperative blood loss ⩾1500 mL	—	44 (65.7)	—
PRBC transfusion ⩾4 units	—	60 (89.6)	—
Combined	—	67 (100)	—

Values are n (%).

CD, cesarean delivery; GW, gestational weeks; HGB, hemoglobin; PAS, placenta accreta spectrum; PRBC, packed red blood cells.

Model development and validation

A Chi-square test was initially used to identify potential risk factors for severe peripartum hemorrhage. Variables with p value <0.10 included age at delivery >33 years, gravidity >3, history of CD >1, preoperative HGB ⩽ 100 g/L, gestational diabetes mellitus, placenta previa classification, preoperative vaginal bleeding, retroplacental myometrial thickness <1 mm, vascular lacunae within the placenta, irregularity of uterine–bladder interface, hypervascularity of the uterine serosa–bladder wall interface, hypervascularity of the cervix, dexamethasone use for fetal lung maturation, and labor duration. Following multivariate logistic regression analysis, age at delivery >33 years (p = 0.004), history of CD >1 (p = 0.020), preoperative HGB ⩽100 g/L (p = 0.013), placenta previa classification (p = 0.001), vascular lacunae within the placenta (p = 0.015), and labor duration (p = 0.026) remained significantly associated with an increased risk of severe peripartum hemorrhage (Table 2).

Table 2.

Multivariate logistic regression analysis.

Characteristics	Multivariate
Characteristics	OR (95% CI)	p Value
Age at delivery >33 years (vs ⩽33)	2.97 (1.42–6.22)	0.004
History of CD >1 (vs ⩽1)	3.23 (1.20–8.69)	0.020
Preoperative HGB level ⩽100 g/L (vs >100)	0.18 (0.05–0.69)	0.013
Placenta previa classification		0.001
Marginal	21.20 (1.95–230.80)	0.012
Partial	1.70 (0.12–23.61)	0.692
Complete	25.66 (2.69–244.69)	0.005
None	Reference
Vascular lacunae within the placenta (vs no)	2.92 (1.24–6.91)	0.015
Labor duration (weeks)		0.026
24 ⩽ GW < 34	1.79 (0.39–8.12)	0.452
34 ⩽ GW < 37	0.40 (0.11–1.43)	0.159
37 ⩽ GW	Reference

CD, cesarean delivery; CI, confidence intervals; GW, gestational weeks; HGB, hemoglobin; OR, odds ratio.

Six ML algorithms were utilized for the validation of the model. The outcomes of the fivefold cross-validation applied to the cohort are detailed in Table 3. These ML algorithms revealed that the model achieved an accuracy ranging from 0.68 to 0.71, with the AUC between 0.75 and 0.79. Figure 2 displays the calibration curve and the DCA curve of the model. The calibration curve closely aligns with the ideal curve, indicating good calibration accuracy, and the DCA curve suggests a certain level of clinical benefit.

Table 3.

Fivefold cross-validation of adverse event prediction model based on ML algorithm.

ML algorithm	AUC	Sensitivity	Specificity	PPV	NPV	Overall accuracy
Logistic regression	0.79 ± 0.05	0.70 ± 0.05	0.67 ± 0.09	0.50 ± 0.08	0.83 ± 0.03	0.68 ± 0.07
Random forest	0.78 ± 0.05	0.76 ± 0.11	0.65 ± 0.08	0.50 ± 0.06	0.86 ± 0.05	0.69 ± 0.05
AdaBoost	0.79 ± 0.04	0.70 ± 0.05	0.71 ± 0.07	0.53 ± 0.06	0.84 ± 0.03	0.71 ± 0.05
Decision tree	0.75 ± 0.04	0.81 ± 0.08	0.65 ± 0.06	0.52 ± 0.03	0.88 ± 0.04	0.70 ± 0.03
K-nearest neighbor	0.76 ± 0.02	0.73 ± 0.12	0.71 ± 0.06	0.53 ± 0.02	0.86 ± 0.05	0.71 ± 0.02
Naïve Bayes	0.79 ± 0.06	0.76 ± 0.10	0.67 ± 0.18	0.54 ± 0.08	0.87 ± 0.03	0.70 ± 0.10

Values are mean ± standard deviation.

AUC, area under the receiver operating characteristic curve; ML, machine learning; NPV, negative predictive value; PPV, positive predictive value.

Figure 2.

(a) Calibration curve and (b) DCA curve of the model.

Model visualization

The nomogram, constructed based on regression coefficients from logistic regression, is presented in Figure 3(a). An illustrative example of the model’s application involves a 34-year-old pregnant woman with PAS at 37 weeks of gestation, planning to undergo CD under NA. Her preoperative HGB level is 90 g/L, and she has a history of one previous CD, with an ultrasound indicating partial placenta previa with vascular lacunae. According to the nomogram, her risk scores for each clinical characteristic (age, gestational weeks, HGB level, history of CD, placenta previa type, and placental lacunae) cumulatively result in a total risk score of 164, corresponding to an approximate 30% risk of severe peripartum hemorrhage.

Figure 3.

(a) Nomogram and (b) the web interface using SHINY of the predictive model.

A user-friendly web application has been developed using SHINY (https://predictingperioperativeadverseeventsinpaspatients.shinyapps.io/RZUI/) to facilitate the use of this predictive model by researchers and clinicians. The web interface (Figure 3(b)) enables users to input a PAS patient’s risk factors and estimate the probability of severe peripartum hemorrhage.

Risk stratification

For risk stratification, the total score of each patient, derived from the nomogram, was categorized into three groups (Table 4): ⩽128 points, indicating a low-risk severe peripartum hemorrhage rate of 5.1%; 128–186 points, suggesting an intermediate-risk rate of 30.9%; and >186 points, denoting a high-risk rate of 63.0%. Figure 4 illustrates the distribution of these risk scores and the counts of individuals with and without severe peripartum hemorrhage in each subgroup.

Table 4.

Risk score and risk stratification.

Risk group	Score	Cases	Adverse event	No adverse event	Adverse event rate, %
Low risk	⩽128	78	4	74	5.1
Intermediate risk	128–186	71	22	49	30.9
High risk	>186	65	41	24	63.0

Figure 4.

Risk score and risk stratification.

Discussion

This study successfully developed a risk assessment model to predict severe peripartum hemorrhage during CD with NA in patients with PAS. Significant predictors identified include age at delivery over 33 years, more than one previous CD, preoperative HGB level of 100 g/L or less, placenta previa classification, presence of vascular lacunae in the placenta, and gestational weeks. A nomogram was created for convenient clinical application, aiding clinicians in selecting appropriate anesthesia prior to CD, and anticipating severe peripartum hemorrhage in PAS patients.

Series studies have consistently demonstrated that NA can effectively be employed in a majority of CD complicated by PAS.^9,10,27 Notably, NA showcases comparable outcomes to GA concerning blood loss, transfusion rates, improved neonatal outcomes, and reduced respiratory complications.²⁷ Nonetheless, instances of NA converting to GA range from 10% to 45%, often driven by factors such as the need for hysterectomy due to anticipated extensive blood loss, significant resuscitation requirements, airway safeguarding, and acidosis management.^10,27,28 Literature reports also cite intensive care unit (ICU) admission rates, major intraoperative transfusion rates, and postoperative complication rates of 4% to 50%, 25% to 50%, and 11% to 27%, respectively.^10,29 Given the absence of anesthesia management guidelines for PAS patients and the limited studies specifically evaluating NA risk in PAS patients, our study retrospectively analyzed PAS patients treated with NA across multiple centers. We then developed a predictive model for severe peripartum hemorrhage.

Placenta previa classification emerged as an independent predictor of severe peripartum hemorrhage within our study. In our cohort, 88.3% were diagnosed with placenta previa, which is characterized by the placenta developing within the lower uterine segment and graded based on the relationship and/or distance between the lower placental edge and the internal cervical orifice.³⁰ Placenta previa not only acts as a risk factor for PAS but also exhibits an association with an elevated risk of severe maternal outcomes among patients diagnosed with PAS. A retrospective study involving 3793 patients with PAS, utilizing data from the US National Inpatient Sample database of the Agency for Healthcare Research and Quality, discovered that placenta previa was independently linked to an increased risk of severe maternal and surgical morbidities, including hysterectomy, severe peripartum hemorrhage, blood product transfusion, shock, DIC, or other coagulopathies.²⁸ Furthermore, the classification of placenta previa bears significance. Complete or major placenta previa, where the placenta covers the entire internal cervical orifice, has been found to correlate with heightened maternal morbidity compared to incomplete placenta previa, where the placenta abuts or only partially covers the internal cervical orifice.³¹ As demonstrated in a retrospective cohort study, this condition leads to increased surgical time, blood loss, transfusion rates, and the incidence of peripartum hysterectomy.³²

In this study, ultrasound measurements were pivotal in assessing the condition of patients with PAS. The categorization of vascular lacunae in the placenta into four grades, as proposed by Finberg and Williams, was based on their number, size, and shape. Their research indicated a direct relationship between higher lacunar grades and an increased incidence of adherent placenta.³³ This finding aligns with our previous retrospective case–control study, which also established a positive correlation between the number of lacunae and maternal complications.^34,35 In a related study, Yang et al., who examined 51 women with placenta previa, observed higher rates of massive transfusions, ICU admissions, and DIC cases in patients with lacunae compared to those without.³⁴

MRI has shown superior accuracy over ultrasound in determining the depth of invasion, the extent of the condition, and the placenta’s relationship with adjacent organs, particularly in cases involving the posterior wall of the placenta. This accuracy is crucial for customizing surgical approaches and anticipating perioperative complications. However, the higher cost and complexity of MRI imaging principles have limited its widespread use. Notably, MRI is not typically employed as a screening method but rather for suspected cases of PAS in specific studies. For instance, one study reported that MRI helped rule out 14 of 16 cases that had false-positive ultrasound diagnoses.³⁶ The ongoing quest for more accessible and accurate diagnostic methods to determine the severity of PAS is crucial for improving models that predict severe peripartum hemorrhage in PAS patients.

The optimal timing for delivery in suspected PAS cases during pregnancy remains a subject of debate. Various medical institutions offer slightly different guidelines, generally recommending planned deliveries between 34 and 36 weeks or 36 and 38 weeks of gestational age.^29,37–39 Decision tree methodology analysis, comparing delivery strategies for patients with ultrasound-detected placenta previa and PAS between 34 and 39 weeks, suggests that delivery at 34 weeks, following corticosteroid administration for lung maturation, may be the preferred timing strategy in all scenarios.⁴⁰ This analysis also indicates that, in cases with a low risk of antenatal bleeding, delivery at 37 weeks could be more suitable. Consistent with previous research, our model supports the notion that the optimal delivery timing is between 34 and 37 weeks, with the risk of hemorrhage increasing if delivery occurs after 37 weeks. The increasing prevalence of PAS cases accompanied by placenta previa elevates the likelihood of prepartum hemorrhage as gestational age advances.^41,42

ML, a subset of artificial intelligence, is gaining prominence in the medical field due to its effectiveness as a computer algorithm that learns from existing data to make accurate predictions, such as forecasting patient disease outcomes.^15,43 ML algorithms are particularly valuable when the research goal is to construct a model that precisely anticipates outcomes. These algorithms excel at uncovering complex and hidden relationships between variables and outcomes, offering versatile and clinically applicable tools.⁴⁴ In this study, ML algorithms were utilized to refine and validate high-risk factors within the model, with the employment of six ML algorithms demonstrating commendable overall accuracy.

In the classification of PAS patients undergoing CD with NA based on risk levels, it was observed that patients with low or medium risk had relatively lower chances of experiencing significant bleeding and requiring blood transfusions during CD. In these cases, NA could be a viable option for intraoperative anesthesia. However, for patients classified as high risk, GA is recommended to ensure more effective intraoperative hemodynamic management.

Currently, several studies have established predictive models for peripartum hemorrhage in patients with PAS.^44–48 Bourgioti et al.⁴⁷ found that an MRI score can be used to predict adverse maternal events, and the presence of six or more PAS-related MRI signs can predict massive peripartum hemorrhage. However, it should be noted that the cost of MRI may limit its routine application in prenatal examinations. Some ultrasound markers are closely related to massive peripartum hemorrhage, such as the number of lacunae, which is included as a high-risk factor in our established predictive model.⁴⁶ In addition, Shazly SA et al. predicted the risk of massive peripartum hemorrhage using the Placenta Accreta Risk-Antepartum (PAR-A) score, which consists of 18 items covering patient demographics, obstetric information, gynecologic history, disease characteristics, and planned management. The PAR-A score achieved an AUC of 0.85 in predicting massive blood loss.⁴⁸ Shazly et al.⁴⁴ also established a model to predict massive peripartum hemorrhage in PAS patients using ML algorithms, with the AUC for the ML antepartum prediction model being 0.84 for massive blood loss. It is worth noting that our article focused on PAS patients undergoing CD under NA. The model established in our study excluded the impact of anesthesia methods on blood loss and transfusion volume and included patient demographics, gynecologic history, ultrasound characteristics, and clinical features for a comprehensive assessment of patients. We anticipate more multicenter, prospective studies to validate the effectiveness of our model in the future. Such validation is crucial for its subsequent integration and effective application within clinical settings.

The study presents several strengths, notably introducing the first predictive model for PAS patients undergoing CD with NA. The prenatal assessment model demonstrates impressive accuracy, making it a practical tool for assessing severe peripartum hemorrhage in PAS patients. Furthermore, such predictive models provide anesthesiologists with valuable insights, aiding them in making informed anesthesia choices that could potentially enhance pregnancy outcomes in clinical practice. However, the study has some limitations. Its retrospective design relies on pre-recorded data, which can introduce selection and recall biases, affecting the validity of the findings. Variability in practices across different institutions can also impact the generalizability of our results, as hospitals may have distinct protocols for managing PAS patients, leading to inconsistencies in outcomes. In addition, due to the absence of pathological evidence in uterus-preserving procedures, most PAS diagnoses are based on obstetrician expertise, which may result in misdiagnosis or missed diagnoses, particularly in complex cases.

Conclusion

In this study, we established and validated a novel prenatal prediction model for evaluating the risk of severe peripartum hemorrhage in patients with PAS who underwent CD under NA. This predictive tool demonstrates potential value in aiding clinicians with risk stratification and optimization of perioperative management strategies. However, considering the inherent limitations of its retrospective design, further validation through rigorous prospective multicenter studies with larger cohorts is warranted before its widespread clinical implementation.

Supplemental Material

sj-docx-1-reh-10.1177_26334941251317644 – Supplemental material for Development of a predictive model for severe peripartum hemorrhage in placenta accreta spectrum cases under neuraxial anesthesia: a multicenter retrospective analysis

Supplemental material, sj-docx-1-reh-10.1177_26334941251317644 for Development of a predictive model for severe peripartum hemorrhage in placenta accreta spectrum cases under neuraxial anesthesia: a multicenter retrospective analysis by Yanan Li, Liang Li, Xiao Song, Fanqing Meng, Meiling Zhang, Yarong Li and Ran Chu in Therapeutic Advances in Reproductive Health

Footnotes

Acknowledgements

None.

Declarations

ORCID iDs

Yanan Li

Liang Li

Xiao Song

Fanqing Meng

Meiling Zhang

Yarong Li

Ran Chu

Supplemental material

Supplemental material for this article is available online.

References

Jauniaux

Bhide

Kennedy

, et al. FIGO consensus guidelines on placenta accreta spectrum disorders: prenatal diagnosis and screening. Int J Gynaecol Obstet 2018; 140: 274–280.

Einerson

Weiniger

CF.

Placenta accreta spectrum disorder: updates on anesthetic and surgical management strategies. Int J Obstet Anesth 2021; 46: 102975.

Salmanian

Fox

Arian

, et al. In vitro fertilization as an independent risk factor for placenta accreta spectrum. Am J Obstet Gynecol 2020; 223: 568.e1–568.e5.

Nieto

Echavarría

Carvajal

, et al. Placenta accreta: importance of a multidisciplinary approach in the Colombian hospital setting. J Matern Fetal Neonatal Med 2020; 33: 1321–1329.

Kocherginsky

Hibbard

JU.

Abnormal placentation: twenty-year analysis. Am J Obstet Gynecol 2005; 192: 1458–1461.

Solheim

Esakoff

Little

, et al. The effect of cesarean delivery rates on the future incidence of placenta previa, placenta accreta, and maternal mortality. J Matern Fetal Neonatal Med 2011; 24: 1341–1346.

Collins

Alemdar

van Beekhuizen

, et al. Evidence-based guidelines for the management of abnormally invasive placenta: recommendations from the International Society for Abnormally Invasive Placenta. Am J Obstet Gynecol 2019; 220: 511–526.

Warrick

Rollins

MD.

Peripartum anesthesia considerations for placenta accreta. Clin Obstet Gynecol 2018; 61: 808–827.

Taylor

Russell

Anaesthesia for abnormally invasive placenta: a single-institution case series. Int J Obstet Anesth 2017; 30: 10–15.

10.

Markley

Farber

Perlman

, et al. Neuraxial anesthesia during cesarean delivery for placenta previa with suspected morbidly adherent placenta: a retrospective analysis. Anesth Analg 2018; 127: 930–938.

11.

Collins

Moons

KGM

Dhiman

, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024; 385: e078378.

12.

Kayem

Deneux-Tharaux

Sentilhes

PACCRETA: clinical situations at high risk of placenta ACCRETA/percreta: impact of diagnostic methods and management on maternal morbidity. Acta Obstet Gynecol Scand 2013; 92: 476–482.

13.

Seligman

Ramachandran

Hegde

, et al. Obstetric interventions and maternal morbidity among women who experience severe postpartum hemorrhage during cesarean delivery. Int J Obstet Anesth 2017; 31: 27–36.

14.

Kilpatrick

Ecker

JL.

Severe maternal morbidity: screening and review. Am J Obstet Gynecol 2016; 215: B17–B22.

15.

Goldstein

Navar

Carter

RE.

Moving beyond regression techniques in cardiovascular risk prediction: applying machine learning to address analytic challenges. Eur Heart J 2017; 38: 1805–1814.

16.

Hae

Kang

Kim

, et al. Machine learning assessment of myocardial ischemia using angiography: development and retrospective validation. PLoS Med 2018; 15: e1002693.

17.

Krittanawong

Zhang

Wang

, et al. Artificial intelligence in precision cardiovascular medicine. J Am Coll Cardiol 2017; 69: 2657–2664.

18.

Park

HA.

An introduction to logistic regression: from basic concepts to interpretation with particular attention to nursing domain. J Korean Acad Nurs 2013; 43: 154–164.

19.

Svetnik

Liaw

Tong

, et al. Random forest: a classification and regression tool for compound classification and QSAR modeling. J Chem Inf Comput Sci 2003; 43: 1947–1958.

20.

Hothorn

Boosting—an unusual yet attractive optimiser. Methods Inf Med 2014; 53: 417–418.

21.

Speybroeck

Classification and regression trees. Int J Public Health 2012; 57: 243–246.

22.

Wang

A fast exact k-nearest neighbors algorithm for high dimensional search using k-means clustering and triangle inequality. Proc Int Jt Conf Neural Netw 2012; 43: 2351–2358.

23.

Zhang

Naïve Bayes classification in R. Ann Transl Med 2016; 4: 241.

24.

Steyerberg

Vickers

Cook

, et al. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology 2010; 21: 128–138.

25.

Park

SY.

Nomogram: an analogue tool to deliver digital knowledge. J Thorac Cardiovasc Surg 2018; 155: 1793.

26.

Van Calster

Nieboer

Vergouwe

, et al. A calibration hierarchy for risk models was defined: from utopia to empirical data. J Clin Epidemiol 2016; 74: 167–176.

27.

Nguyen-Lu

Carvalho

Kingdom

, et al. Mode of anesthesia and clinical outcomes of patients undergoing Cesarean delivery for invasive placentation: a retrospective cohort study of 50 consecutive cases. Can J Anaesth 2016; 63: 1233–1244.

28.

Han

Guo

Yang

, et al. Association of placenta previa with severe maternal morbidity among patients with placenta accreta spectrum disorder. JAMA Netw Open 2022; 5: e2228002.

29.

Shamshirsaz

Fox

Salmanian

, et al. Maternal morbidity in patients with morbidly adherent placenta treated with and without a standardized multidisciplinary approach. Am J Obstet Gynecol 2015; 212: 218.e1–9.

30.

Jauniaux

Alfirevic

Bhide

, et al. Placenta praevia and placenta accreta: diagnosis and management: Green-top Guideline No. 27a. BJOG 2019; 126: e1–e48.

31.

Daskalakis

Simou

Zacharakis

, et al. Impact of placenta previa on obstetric outcome. Int J Gynaecol Obstet 2011; 114: 238–241.

32.

Wortman

Schaefer

McIntire

, et al. Complete placenta previa: ultrasound biometry and surgical outcomes. AJP Rep 2018; 8: e74–e78.

33.

Finberg

Williams

JW.

Placenta accreta: prospective sonographic diagnosis in patients with placenta previa and prior cesarean section. J Ultrasound Med 1992; 11: 333–343.

34.

Yang

Lim

Kim

, et al. Sonographic findings of placental lacunae and the prediction of adherent placenta in women with placenta previa totalis and prior Cesarean section. Ultrasound Obstet Gynecol 2006; 28: 178–182.

35.

Chu

Wang

, et al. Role of abdominal aortic balloon placement in planned conservative management of placenta previa with placenta increta or percreta. Front Med (Lausanne) 2021; 8: 767748.

36.

Riteau

Tassin

Chambon

, et al. Accuracy of ultrasonography and magnetic resonance imaging in the diagnosis of placenta accreta. PLoS One 2014; 9: e94866.

37.

Al-Khan

Gupta

Illsley

, et al. Maternal and fetal outcomes in placenta accreta after institution of team-managed care. Reprod Sci 2014; 21: 761–771.

38.

Camuzcuoglu

Vural

Hilali

, et al. Surgical management of 58 patients with placenta praevia percreta. Wien Klin Wochenschr 2016; 128: 360–366.

39.

Rac

MWF

Wells

Twickler

, et al. Placenta accreta and vaginal bleeding according to gestational age at delivery. Obstet Gynecol 2015; 125: 808–813.

40.

Robinson

Grobman

WA.

Effectiveness of timing strategies for delivery of individuals with placenta previa and accreta. Obstet Gynecol 2010; 116: 835–842.

41.

Warshak

Ramos

Eskander

, et al. Effect of predelivery diagnosis in 99 consecutive cases of placenta accreta. Obstet Gynecol 2010; 115: 65–69.

42.

Wright

Pri-Paz

Herzog

, et al. Predictors of massive blood loss in women with placenta accreta. Am J Obstet Gynecol 2011; 205: 38.e1–6.

43.

Darcy

Louie

Roberts

LW.

Machine learning and the profession of medicine. JAMA 2016; 315: 551–552.

44.

Shazly

Hortu

Shih

, et al. Prediction of clinical outcomes in women with placenta accreta spectrum using machine learning models: an international multicenter study. J Matern Fetal Neonatal Med 2022; 35: 6644–6653.

45.

Shazly

Hortu

Shih

, et al. Prediction of success of uterus-preserving management in women with placenta accreta spectrum (CON-PAS score): a multicenter international study. Int J Gynaecol Obstet 2021; 154: 304–311.

46.

Hussein

Momtaz

Elsheikhah

, et al. The role of ultrasound in prediction of intra-operative blood loss in cases of placenta accreta spectrum disorders. Arch Gynecol Obstet 2020; 302: 1143–1150.

47.

Bourgioti

Zafeiropoulou

Fotopoulos

, et al. MRI prognosticators for adverse maternal and neonatal clinical outcome in patients at high risk for placenta accreta spectrum (PAS) disorders. J Magn Reson Imaging 2019; 50: 602–618.

48.

Shazly

Anan

Makukhina

, et al. Placenta accreta risk-antepartum score in predicting clinical outcomes of placenta accreta spectrum: a multicenter validation study. Int J Gynaecol Obstet 2022; 158: 424–431.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.09 MB