A Novel and Modern Calculator to Predict Vaginal Birth after Cesarean Delivery

Abstract

Objective

Counseling patients who are considering a trial of labor after cesarean (TOLAC) is a challenging task given the risks and benefits of either approach. While calculators exist to give patients an idea of their likelihood of having a successful vaginal birth after cesarean (VBAC), their validity is limited by outdated mathematical methods used to develop them. Most importantly, current VBAC calculators only offer insight into the chance of successful VBAC, without any ability to predict the risk of adverse outcomes relevant to both the patient and neonate. The objective of this study is to develop a prediction model for individualized risks and benefits of a TOLAC using modern mathematical techniques.

Study Design

This was a secondary analysis of the Cesarean Registry database, the same database used in developing the Maternal–Fetal Medicine Units (MFMU) VBAC calculator. The primary outcome was the prediction of the success of VBAC. Secondary outcomes were the prediction of uterine rupture, maternal complications, and neonatal complications. Inclusion criteria were term, singleton gestation, and cephalic presentation pregnancies with one prior low transverse cesarean delivery (CD). Exclusion criteria included intrauterine fetal demise, planned cesarean, and prior myomectomy. Univariate comparisons identified variables that were independently associated with VBAC. An optimal decision tree was used to create a prediction model. A test set was withheld for validation. A risk calculator tool was developed for the prediction of successful VBAC and adverse perinatal outcomes. Adverse maternal outcomes: uterine dehiscence, hysterectomy, postpartum hemorrhage, endometritis, intensive care unit admission, thromboembolic event, readmission, and organ injury. Adverse neonatal outcomes: hypoxic-ischemic encephalopathy, respiratory distress, seizures, apnea, respirator use, death, and cord blood pH < 7.1.

Results

The study population included 73,262 deliveries of which 12,942 patients met inclusion and exclusion criteria. After removing cases for the test set, the included patients were 8,078 patients, of which 5,970 people had a successful VBAC (73.9%). Parity, number of years since prior CD, prepregnancy body mass index (BMI), delivery BMI, maternal age, and previous VBAC were associated with successful VBAC. A risk predictor calculator was created, and a receiver operator characteristic curve was developed with an area under the curve of 0.72. The tool was also developed to identify a person's risk of uterine rupture, composite maternal morbidity, and neonatal morbidity.

Conclusion

VBAC for patients with term, cephalic, singleton gestation was associated with several variables. This advanced calculator tool will facilitate shared decision-making about the value of a TOLAC regarding the personalized risks of maternal and neonatal morbidity. By using more advanced mathematical models, this tool allows providers to predict not only the likelihood of successful VBAC but also the risk of maternal and neonatal morbidity involved in attempting VBAC.

Key Points

• Current VBAC calculators are limited by the mathematical methods used to make them.

• This novel calculator uses more advanced machine-learning methods than previous calculators.

• This VBAC calculator predicts both the chance of success VBAC and the risk of morbidity.

• The more modern VBAC calculator gives providers more information to use when counseling patients.

 

MissingLinkText

Keywords

cesarean delivery - machine learning - risk prediction - VBAC

The rate of cesarean deliveries (CD) in the United States has risen sharply over the last three decades.[1] While there are many factors driving this increase, one of the largest contributors is the hesitancy of many patients to attempt a vaginal birth after cesarean (VBAC). Providers and patients may be hesitant to attempt VBAC for a multitude of reasons, most notably due to the increased risk of uterine rupture or dehiscence; a catastrophic complication of VBAC that carries high levels of maternal morbidity.[2] [3] [4] Increased efforts to prevent primary CD's in low-risk patients will hopefully lead to fewer repeat CD's in future decades; however, many patients today are likely being counseled to undergo repeat CD. Risks of repeated CD's are well described with the most significant risks being an increase in abnormal placentation with subsequent pregnancies and a higher risk of intraoperative complications with each successive CD.[5] In all cases of patients undergoing a trial of labor after cesarean (TOLAC), providers are currently unable to offer comprehensive personalized risk/benefit counseling of attempting TOLAC to aid patient decision-making. Previous retrospective studies have identified factors associated with a decreased likelihood of successful VBAC, such as high body mass index (BMI), no prior spontaneous vaginal delivery, and “fetal distress” as an indication for CD.[2] Additionally, the well-known calculator from the Maternal–Fetal Medicine Units Network (MFMU) was developed to give providers and their patients a predictive tool to use in attempting to predict the likelihood of successful VBAC. This calculator remains widely used by clinicians but with substantial limitations. External validations of this calculator have shown it has a moderate degree of accuracy in predicting successful VBAC when the calculated success rate is high (i.e., above 65%), but is less accurate when the predicted success rate for VBAC is lower than 35%, with many women who had a low predicted success rate going on to have a successful VBAC.[6] One major limitation of the calculator that has come into question in recent years is its use of race in the prediction of successful VBAC,[7] which has been subsequently removed in an updated version.[8] Race and ethnicity are both used as predictive variables in the original MFMU VBAC calculator, both of which have been questioned in validation studies that demonstrate their limited predictive value in predicting VBAC.[9] While the use of race and ethnicity has since been removed from the calculator, the MFMU VBAC calculator is still limited in its utility given its age, use of simpler, multivariate regression models to predict variables associated with successful VBAC, and inability to take more variables into account during calculations.[9]

Current tools used to counsel patients on the successful likelihood of VBAC are limited. No current calculator exists to counsel patients on the individualized risks of VBAC. There is little ability for providers to accurately counsel patients regarding their individual risk of success in addition to their risk of uterine rupture and risk of neonatal complications by delivery mode. Without knowledge of the patient's personalized risk, and risk to the neonate, there is little utility in any VBAC calculator that only provides patients with the likelihood of success, regardless of its accuracy. Along with these practical limitations, older calculators also have substantial mathematical limitations. While the multivariate regression nomograms used by previous iterations of the VBAC calculator were once the mathematical standard for prediction models in health care, advanced mathematical regression techniques that utilize machine learning and artificial intelligence have given clinicians the ability to construct more intuitive, accurate, and transparent predictive mathematical models. The use of optimal decision trees (ODT) has increased in the development of predictive models in medicine for their superior accuracy and usability.[10] [11] [12]

The objective of this study was to create a novel VBAC calculator that improves upon previous calculators by utilizing more advanced machine learning algorithms. Additionally, a secondary objective was to develop a calculator that provides individualized risks and benefits of attempting VBAC by providing patients with their percent likelihood of uterine rupture, maternal morbidity, neonatal morbidity, and also the chance of successful VBAC. The goal of this calculator will ultimately be to provide patients and providers with a tool with excellent predictive power that will enhance shared decision-making in patients considering a TOLAC.

Materials and Methods

This was a retrospective cohort study of patients in the MFMU Cesarean Registry database. This database includes patients who received care between 1999 and 2002 at 19 academic medical centers in the United States, all of which were within the NICHD MFMU hospital network.[13] This database was chosen purposefully to utilize the same data as the MFMU VBAC calculator. The primary outcome of the study was a prediction of VBAC. The secondary outcomes of the analysis were a prediction of uterine rupture, a composite of maternal complications, and a composite of neonatal complications. Maternal complications included uterine dehiscence, hysterectomy, postpartum hemorrhage, endometritis, intensive care unit admission, thromboembolic events, readmission, and organ injury. Neonatal complications included hypoxic-ischemic encephalopathy, respiratory distress, seizures, apnea, respirator use, death, and cord blood pH < 7.1. Inclusion criteria for the analysis included term, singletons, cephalic presentation, and one prior low transverse CD. Exclusion criteria were intrauterine fetal demise, planned cesarean, or prior myomectomy. Patient demographic variables were chosen using univariate regression, which identified variables that were independently associated with successful VBAC. The study population included 73,262 deliveries, of which 12,942 patients met inclusion and exclusion criteria. After removing cases for the test set, the included patients were 8,078 patients, of which 5,970 people had a successful VBAC (73.9%; [Fig. 1]).

Fig. 1 A flow chart depicting the selection of patients included in the study.

With these variables identified, an ODT was utilized to develop a predictive model for our analysis. Unlike other artificial intelligence-driven machine learning models, ODTs are nonparametric. This means that variables identified as being independently associated with the primary outcome aren't assigned numerical values that weight their individual predictive value on this outcome. Contrary to parametric models, the ODT generates multiple models and aims to optimize the highest out-of-sample performance, specifically focusing on the area under the curve during test set validation. This difference is significant in multiple ways. First, it allows the ODT to be fit to a wider range of patterns in data than simpler linear regression models, such as multivariate models. Second, the lack of numerical coefficients assigned to each variable means that there are no p-values, confidence intervals, or odds ratios. Instead, the OCT is the single best model that can fit the entire dataset without any assumption of linear relationships between variables and outcomes. Also unique to OCTs is that the branches are not identical across the entire model. This means that variables present on one branch of the OCT can be different from the variables included in other branches. This allows the model to account for new information being fed into it in real time, causing it to prompt users with questions based on previous answers it is provided with. Finally, the OCT model, in contrast with other approaches to machine learning that rely on “black box” neural networks that map inputs to outputs without explicitly revealing the underlying decision-making process, is transparent. This means they offer full explanations behind the predictions they make, allowing for a more robust analysis of their performance.

To develop the model, a test set was withheld from the population of patients who met inclusion and exclusion criteria. After the model was trained, the OCT's performance was evaluated on the patients in the previously withheld test set. A receiver–operator characteristics curve (ROC) was developed to analyze the model's performance. A calculator tool was developed for the prediction of successful VBAC and associated maternal and neonatal adverse outcomes. Prior to the development of the algorithm or the analysis of data, this study was deemed exempt by the Women and Infants Institutional Review Board (WIH 22-0022). All statistical analyses were done using Julia with statistical significance being determined by p < 0.05.[14]

 

MissingLinkText

Results

The study population included 73,262 deliveries of which 12,942 patients met inclusion and exclusion criteria. After removing cases for the test set, the included patients totaled 8,078, of which 5,970 people had a successful VBAC (73.9%). Parity, number of years since prior CD, prepregnancy BMI, delivery BMI, maternal age, and previous VBAC were associated with successful VBAC. Most women in this cohort were non-Hispanic and either White or Black ([Table 1]).

Table 1 Demographics of included women

Variable name	Included patients
Maternal age (y)	28.6 ± 5.8
Self-reported race
White, non-Hispanic	4,465 (38.7%)
Black, non-Hispanic	4,184 (36.3%)
Hispanic	2,302 (19.9%)
Asian	227 (2.0%)
Native American	22 (0.2%)
Other race	339 (2.9%)
BMI (prepregnancy)	26.35 ± 6.28
Parity	1.87 ± 1.23
Prior SVD	5,503 (47.9%)
Years between delivery and the last CD	5.21 ± 3.63
Prior VBAC	3,588 (32.5%)
Arrest disorder indication for prior cesarean	3,829 (33.2%)
Gestational age	277.46 ± 8.23
Insurance
Government assisted	5,008 (45.0%)
Private	5,059 (45.5%)
Self-pay	1,055 (9.5%)
Diabetes	583 (5.1%)
Hypertension	139 (1.2%)
Cigarette use	1,666 (14.4%)
Asthma	832 (7.2%)
Seizure disorder	90 (0.8%)
Renal disease	66 (0.6%)
Heart disease	126 (1.1%)
Thyroid disease	283 (2.5%)
Connective tissue disorder	50 (0.4%)
Labor
Induction	3,119 (27.0%)
Oxytocin	7,013 (60.8%)
Dilation on admission	3.26 ± 2.04
Station on admission	3.62 ± 1.74
Epidural	7,901 (83.6%)

Abbreviations: BMI, body mass index; SVD, spontaneous vaginal delivery; VBAC, vaginal birth after cesarean.

A risk predictor calculator was created using the OCT that takes into account patient variables found to be independently associated with successful VBAC ([Fig. 2]). To illustrate the calculator tool, four example patients were input into the model and their predicted likelihood of successful VBAC along with both maternal and neonatal outcomes were calculated ([Table 2]). Using patient 1 as an example, we can see that a 22-year-old with a parity of 1, a prepregnancy BMI of 20, delivery BMI of 25, who delivered via CD 1 year ago and has never had a prior VBAC is likely to have a successful VBAC (76%), unlikely to have uterine rupture (0%), and has low risk of bother maternal (2%) and neonatal (6%) adverse outcomes. Whereas, a 40-year-old with a parity of 1, a prepregnancy BMI of 50, a delivery BMI of 56, who delivered via cesarean section 3 years ago, and has never had a prior VBAC is less likely to have a successful VBAC (24%), still unlikely to have uterine rupture (3%) or adverse maternal outcomes (3%), but is more likely to have adverse neonatal outcome (10%). An ROC was developed to assess the effectiveness of the calculator as a tool to predict VBAC with an area under the curve being 0.72. The AUCs describing the calculator's ability to predict uterine rupture, maternal morbidity, and neonatal morbidity were 0.56, 0.66, and 0.59, respectively.

Table 2 Example patients with probabilities of successful VBAC

Input	Case 1	Case 2	Case 3	Case 4
Parity	1	1	3	4
Prepregnancy BMI	20	50	29	22
Delivery BMI	25	56	36	28
Maternal age	22	40	30	35
Interval since prior cesarean (y)	1	3	11	3
Prior VBAC	No	No	Yes	Yes
Output
VBAC probability	76%	24%	78%	95%
Uterine rupture probability	0%	3%	1%	0%
Adverse maternal outcome probability	2%	3%	12%	2%
Adverse neonatal outcome probability	7%	10%	9%	5%

Abbreviations: BMI, body mass index; VBAC, vaginal birth after cesarean.

Note: Example patients whose probabilities of successful VBAC, along with risks of maternal and neonatal adverse outcomes, have been calculated using the calculator tool developed with the optimal classification tree. Green numbers represent a successful VBAC probability over 50%, and adverse outcome risks lower than 10%. Red numbers represent a successful VBAC probability below 50%, and adverse outcome risks higher than 10%.

Fig. 2 A collapsed section of the optimal classification tree (OCT) algorithm used to develop the risk assessment 33 tools. Green nodes indicate >50% chance of successful vaginal birth after cesarean (VBAC). Red nodes indicate <50% chance of successful VBAC. Gray nodes can be expanded to reveal more possible outcomes. Start at node 1. Then depending on patient characteristics, follow the tree. For example, following the tree to node 21 predicts that for a patient with at least one prior vaginal delivery, and one prior VBAC with a prepregnancy BMI < 46.7 they have a predicted chance of VBAC success of 81.21%. A section of the optimal classification tree (OCT) algorithm is used to develop the risk assessment too.

 

MissingLinkText

Conclusion

In this study of patients with term, singleton gestation, and cephalic presentation pregnancies with a history of one prior low transverse CD a calculator tool for the chance of successful VBAC along with the chance of maternal and neonatal adverse outcomes was developed. This tool can potentially help providers counsel patients who are considering undergoing TOLAC not only about their personalized chance of success but also about their individualized risk to both them and the neonate. Multiple variables (parity, number of years since prior CD, prepregnancy BMI, delivery BMI, maternal age, and previous VBAC) were identified as being independently associated with successful VBAC. Tools to predict the chance of successful VBAC have existed for decades.[15] [16] Multiple calculators have been published, with some having moderate ability to predict successful VBAC in patients using similar variables used in the development of this calculator.[17] However, previous VBAC calculators' reliance on nomograms and simple prediction models represents their major limitation and their inclusion of race as a predictive variable in their original publications has limited their use and validity in recent years. Our novel calculator circumvents these limitations by using more advanced mathematical models and by omitting race from our calculator. The model's notable ROC suggests value for its clinical use. Medical decision-making has always relied on scientific corroboration to maximize patient care. In the current age of big data and advanced analytics, it is wise for medical care providers to recognize the boost that these tools can bring to individual patient analysis. We see tools such as this one serving as an adjunct to current patient counseling.

This study serves to not only expand providers' capabilities for counseling patients on the risks and benefits of TOLAC but also demonstrates a novel approach to using advanced machine learning algorithms in the setting of medical decision-making. The use of an ODT in this calculator is an example of how our growing knowledge of machine learning can be used to improve on previous medical interventions, specifically ones that were designed with less complex mathematical methods. Future directions in this sphere of research would be to use ODTs in other predictive calculators and to develop further iterations of the VBAC calculator as more advanced computational methods become available. Modern medicine is poised to take advantage of database confirmation and augmentation of the solitary clinician's experiences, and the use of predictive tools is a logical extension of centuries of scientific inquiry and knowledge accumulation. Although no two medical situations are identical, a prediction tool that has the ability to analyze and assess the significant parameters related to the specifics of the case in question can quickly elevate the decision-making process. In obstetrics, enhanced data can enrich and strengthen the singular provider's abilities to ensure the safest possible birthing experience for mother and child. Future studies could focus on a practical trial using this prediction model. Confident use of the calculator in the clinical setting would rely on a practical trial using this predictive model.

 

MissingLinkText

Strengths and Limitations

A major strength was selected using a trusted dataset from the original MFMU VAC calculator. This database is a validated, reliable database. Other strengths include that the data analysis was done using current cutting-edge methodology and that the algorithm and calculator tool are straightforward to interpret and easy to use. Additionally, our ROC curve showed that our model was able to distinguish delivery methods with good precision. An additional strength of this study is the use of an OCT. Of note, branches of the OCT are not identical, and this method allows for a prediction that is adaptive and reboots itself with each variable, accounting for nonlinear interactions among variables. The variables used by the tree can change at each level; therefore, questions in the calculator tool change based on the responses to the prior question. Decision trees can therefore capture nonlinear interactions between variables rather than mandate that the variables interact in a linear and additive manner as in classical logistic regression. Additionally, none of the aforementioned studies have looked at the concurrent prediction of maternal and neonatal risks associated with the time of delivery as a balance to the risk of delivery mode. Our study eliminated many of these previous issues with its large sample size, ease of use of the predictor tool, and practical use because our input data are available to any provider. Because of the retrospective nature of this study, there is a chance of misclassification bias and selection bias.

This study is limited by the age of the data set used to develop the algorithm and calculator tool. The MFMU network Cesarean Registry database is over 10 years old, which means it does not perfectly reflect the current patient population this calculator is intended to be used on. While ODTs are superior to older machine learning models, they are not without limitations themselves. ODTs have more potential to be overfitted to the data than older, more simplistic models. This is minimized by “pruning” the model while it is being developed (simplifying the model by removing the least relevant branches), but is not completely avoidable.

 

MissingLinkText

Conclusion

In conclusion, we found that VBAC for patients with term, cephalic, and singleton gestations was associated with several variables. This advanced calculator tool will facilitate shared decision-making about the value of a TOLAC regarding the personalized risks of maternal and neonatal morbidity.

 

MissingLinkText

 

MissingLinkText

Conflict of Interest

None declared.

Acknowledgment

The work here was previously presented in poster format at the Society for Maternal–Fetal Medicine Pregnancy Meeting in National Harbor, MD, on February 13, 2024.

• References

• 1 Angolile CM, Max BL, Mushemba J, Mashauri HL. Global increased cesarean section rates and public health implications: a call to action. Health Sci Rep 2023; 6 (05) e1274
• 2 Lazarou A, Oestergaard M, Netzl J, Siedentopf JP, Henrich W. Vaginal birth after cesarean (VBAC): fear it or dare it? An evaluation of potential risk factors. J Perinat Med 2021; 49 (07) 773-782
• 3 Folsom S, Esplin MS, Edmunds S. et al. Patient counseling and preferences for elective repeat cesarean delivery. AJP Rep 2016; 6 (02) e226-e231
• 4 Kurtz Landy C, Sword W, Kathnelson JC. et al. Factors obstetricians, family physicians and midwives consider when counselling women about a trial of labour after caesarean and planned repeat caesarean: a qualitative descriptive study. BMC Pregnancy Childbirth 2020; 20 (01) 367
• 5 Gasim T, Al Jama FE, Rahman MS, Rahman J. Multiple repeat cesarean sections: operative difficulties, maternal complications and outcome. J Reprod Med 2013; 58 (7–8): 312-318
• 6 Maykin MM, Mularz AJ, Lee LK, Valderramos SG. Validation of a prediction model for vaginal birth after cesarean delivery reveals unexpected success in a diverse American population. AJP Rep 2017; 7 (01) e31-e38
• 7 Vyas DA, Jones DS, Meadows AR, Diouf K, Nour NM, Schantz-Dunn J. Challenging the use of race in the vaginal birth after cesarean section calculator. Womens Health Issues 2019; 29 (03) 201-204
• 8 Grobman WA, Sandoval G, Rice MM. et al; Eunice Kennedy Shriver National Institute of Child Health and Human Development Maternal-Fetal Medicine Units Network. Prediction of vaginal birth after cesarean delivery in term gestations: a calculator without race and ethnicity. Am J Obstet Gynecol 2021; 225 (06) 664.e1-664.e7
• 9 Nguyen MT, Hayes-Bautista TM, Hsu P, Bragg C, Chopin I, Shaw KJ. Applying a prediction model for vaginal birth after cesarean to a Latina inner-city population. AJP Rep 2020; 10 (02) e148-e154
• 10 Bertsimas D, Dunn J, Pawlowski C. et al. Applied informatics decision support tool for mortality predictions in patients with cancer. JCO Clin Cancer Inform 2018; 2: 1-11
• 11 Gimovsky AC, Levine JT, Pham A, Dunn J, Zhou D, Peaceman AM. Pushing the bounds of second stage in term nulliparas with a predictive model. Am J Obstet Gynecol MFM 2019; 1 (03) 100028
• 12 Bertsimas D, Dunn J. Optimal classification trees. Mach Learn 2017; 106: 1039-1082
• 13 Landon MB, Hauth JC, Leveno KJ. et al; National Institute of Child Health and Human Development Maternal-Fetal Medicine Units Network. Maternal and perinatal outcomes associated with a trial of labor after prior cesarean delivery. N Engl J Med 2004; 351 (25) 2581-2589
• 14 Bezanson J, Edelman A, Karpinski S, Shah VB. Julia: A fresh approach to numerical computing. SIAM Rev 2017; 59 (01) 65-98
• 15 Troyer LR, Parisi VM. Obstetric parameters affecting success in a trial of labor: designation of a scoring system. Am J Obstet Gynecol 1992; 167 (4 Pt 1): 1099-1104
• 16 Metz TD, Stoddard GJ, Henry E, Jackson M, Holmgren C, Esplin S. Simple, validated vaginal birth after cesarean delivery prediction model for use at the time of admission. Obstet Gynecol 2013; 122 (03) 571-578
• 17 Grobman WA, Lai Y, Landon MB. et al. Does information available at admission for delivery improve prediction of vaginal birth after cesarean?. Am J Perinatol 2009; 26 (10) 693-701

Publication History

Received: 20 February 2025

Accepted: 11 May 2025

Accepted Manuscript online:
12 May 2025

Article published online:
29 May 2025