Cesarean delivery is one of the most common surgical procedures in the United States. The cesarean rate in the United States has steadily risen over the past three decades but has apparently leveled off at 22% to 25%.1,2 The factors that have contributed to the increase in cesarean delivery rates over the last 25 years are not completely known. However, when cesarean delivery rates are examined using patient characteristics, the highest variation occurs in term, nulliparous patients with singleton fetuses in the vertex presentation.3
The overall mortality rate from cesarean delivery is 6/100,000, which is three to seven times greater than that for vaginal delivery.3 Cesarean deliveries that are performed after a patient has been in labor have higher rates of morbidity and mortality when compared with vaginal delivery or elective cesarean deliveries.4,5 If one were able to predict which patients were at highest risk for cesarean delivery when they entered labor and offer a cesarean before a prolonged course of labor, then unnecessary time in labor, increased costs, and complications might be avoided.
The objective of this study was to identify risk factors that, upon arrival to labor and delivery, place a term nulliparous patient in labor at high risk for cesarean delivery, and to build a model for predicting a patient's risk.
MATERIALS AND METHODS
The Colorado Multiple Institutional Review Board, which reviews human research protocols for several health institutions in Denver, including the University of Colorado Hospital, approved this study. All deliveries occurred at the University of Colorado Hospital. Patients were excluded if there were major fetal anomalies or significant maternal disease processes. This was a case-control design that consisted of 325 women.6 The 174 patients were all of the nulliparous patients other than those excluded at term, who presented in labor, aged 14 to 45 years, with singleton gestations in the vertex presentation and who had a cesarean delivery at University of Colorado Hospital between 1998 and 2001. The 151 controls were defined as the next nulliparous patient who met all inclusion criteria and delivered vaginally. A chart review was conducted extracting 51 variables on a standardized data form. The variables were common descriptors of pregnancy or previously cited risk factors for cesarean delivery.7–10 A sample size of 150 cesarean patients and 150 controls was chosen, on the basis of the following rationale. This sample size provides 80% power to detect a difference in incidence of a risk factor or outcome between patients and controls with an odds ratio of 2.0 (25% versus 40%, for example) with a two-sided α of 0.05. In addition, for risk factors or outcomes measured on a continuous scale, it provides 80% power, with a two-sided α of 0.05 to detect a moderate effect size of one-third or greater.
During the study period, obstetric patients were treated by house staff under supervision of attending physicians. In approximately 97% of cases, the attending physician was a full-time faculty member. Across attending and resident physicians, practices are highly consistent, regardless of insurance status of the patient. For example, labor augmentation with oxytocin is performed if there is inadequate progress in labor, defined as being in the active phase of labor for nulliparous women, less the 1.2 cm dilatation per hour for 2 hours. Oxytocin is administered by intravenous pump beginning at 1 to 2 mU/min and increasing linearly every 20 to 30 minutes until 150 or more Montevideo units of uterine activity are achieved. The criteria for cesarean delivery for arrest of labor in the active phase is at least 2 hours of adequate uterine activity (ie, 150 or more Montevideo units) and no cervical dilatation. The cesarean delivery rate for nulliparous women at term with a singleton vertex presentation is approximately 15.5%, as determined by a review of 167 such women from the delivery log during the first half of March and of September of each year of the study. Dichotomous variables were analyzed by the χ2 test or Fisher exact test, as appropriate, and continuous variables were analyzed by the Wilcoxon two-sample test.
A multiple logistic regression analysis was performed to identify independent risk factors for cesarean delivery. Factors were chosen for inclusion in the multiple regression analysis if they were significant by univariate analysis and would be easy to measure during the first 2 hours after admission. A multiple regression model for predicting risk was built and evaluated on our study population. The model was constructed by stepwise selection from a pool of univariately significant variables with P = .05 used as the criterion for both entry into and retention in the model. The subset of variables presented for selection were those judged likely to be the most clinically important. All analyses were performed by SAS 8.2 (SAS Institute, Cary, NC).
The logistic regression model was validated by the training/testing sets approach described by Osborne.11 For our cross-validation assessment, we randomly split our data into equally sized training and test sets, each containing equal numbers of patients and controls. A logistic model that used the five selected variables (see below) was then built on the training set, and these coefficients were then used to predict cesarean delivery for the patients in the test set. This process was repeated 10,000 times, and the average sensitivity and specificity across all scores from 0 to 1 in 0.05 increments were calculated and plotted.
Patient characteristics and significant variables are listed in Table 1. There were no significant differences in maternal height, maternal temperature at admission, number of cigarettes smoked per day, gestational age at first prenatal visit, or number of prelabor visits to triage on the labor and delivery unit. Of the 51 variables examined, 22 were statistically significantly different between patients and controls in the univariate analysis. Of these 22, cervical dilatation, effacement, and station at admission and changes in these values by 2 hours, maternal weight, gestational age, diabetes, preeclampsia, and abnormal amniotic fluid volume (abnormal amniotic fluid index or subjectively low or absent at the time of rupture of membranes) were selected as being likely to be known within 2 hours of admission and were used in the initial model building steps. Abnormal fetal heart rate tracing (recurrent late decelerations or severe variables), although determined in the first 2 hours, was not included in the model building because it could be viewed as a direct reason to choose cesarean delivery and could confound the analysis.
Of the variables examined (Table 1), five remained significant in a multiple logistic regression model developed to predict which patients were at highest risk for cesarean delivery. These variables were as follows: change in cervical dilatation within the first 2 hours, maternal weight, gestational age, fetal station at 2 hours, and preeclampsia (Table 2). Patients with missing data for any of these five variables did not contribute to the final model; there were 141 cesarean and 149 vaginal deliveries in the final model set, from 174 cesarean and 151 vaginal deliveries for which data were collected. In all 35 patients, the missing data were the 2-hour postad-mission measurements, indicating that these patients were delivered shortly after admission.
In the 33 cesarean deliveries with missing data, the primary indications for cesarean delivery were arrest of labor in 22, nonreassuring fetal heart tones in nine, and other in two patients. Because our primary goal was determining which patients were at highest risk for cesarean delivery after prolonged labor, the absence of very early deliveries from the model data is not a concern.
Table 2 shows, for each variable, the relative risk for cesarean delivery associated with a one-unit change (within the range of values covered by our data) in the value of that variable, controlling for the other four variables in the model. For example, if two patients had identical values for four of the variables but one patient was preeclamptic, the preeclamptic patient would be 5.75 times as likely to have a cesarean delivery. As another example, suppose that two patients had identical characteristics, except that one patient was 25 pounds heavier than the other. Then the heavier patient would be 1.49 times as likely to have a cesarean delivery as the lighter patient. Likewise, if two otherwise identical patients were laboring and the first experienced a 5-cm increase in dilatation, whereas the second experienced only a 3-cm increase (a 2-cm difference in dilatation change), the patient with the larger dilatation change would have 0.23 times the likelihood of delivering by cesarean as the second patient.
Table 3 lists the overall results for this model in our population. Patients were sorted by scores calculated from the model, divided into quintiles, and compared with the actual cesarean delivery data. Patients in our case-control population with scores in the lowest quintile had a cesarean rate of 5%, compared with an almost 90% rate of cesarean delivery among those with scores in the highest quintile.
The receiver operating characteristic curve analysis of the risk score in predicting the probability of cesarean delivery is shown in Figure 1. The numbers on the curve show the position on the curve for selected risk scores from the model. For example, the sensitivity for a score of 0.7 was 44.0% (62 of 141), and the 1-specificity (false positive) was 8.1% (12 of 149). Thus, choosing a score of 0.7 or greater as a cutoff identifies nearly half of the patients needing a cesarean delivery while erroneously predicting a need for cesarean delivery for only 8% of the patients who would deliver vaginally if left to labor.
Because logistic regression models can show positive bias when validated on the model building data set, we performed cross-validation by using the training/testing sets approach.11 Figure 1 shows the results of this cross-validation; the lower, smooth curve is the average receiver operating characteristic for 10,000 iterations of cross-validation. The fact that the cross-validated receiver operating characteristic curve lies so close to that for the model built from the full data set indicates that the model built by using these five variables has strong predictive utility.
We were able to identify significant risk factors for cesarean delivery for nulliparous patients in labor. By means of independent risk factors, as determined by logistic regression analysis, we developed a model for identifying, within 2 hours of presentation in labor, patients who were at highest and lowest risk for cesarean delivery. The performance of the model was impressive in determining women at as lowas 5%risk for cesarean delivery and as high as 88% risk in our case-control population.
One of the strengths of these results is in the ability to determine an individual's risk early in labor and to avoid the increased morbidity and mortality associated with a later cesarean delivery. We limited enrollment to those patients in whom the cesarean rates are the most variable: nulliparous patients with singleton vertex fetuses. We considered a large number of variables, some of which had been reported in the literature to affect cesarean delivery rates in other populations.7,8,12–17 Our patients delivered over a brief time period at a single hospital where practices have been fairly well standardized. We suspect that individual practices did not vary as much as in previous studies.18,19
A weakness of our study is that it is retrospective, and thus, data collection was incomplete in some of our subjects. This accounts for the different numbers of patients and controls, although as mentioned previously, patients with missing data were those who delivered early and are not of interest in developing a model for predicting patients at risk for cesarean delivery after a longer trial of labor.
Our model worked well when applied to this case-control population where the cesarean delivery rate was approximately 50%, but it is unknown whether it would perform equally well in a population-based study. Thus, we attempted to estimate how this model would perform on a population with a cesarean delivery rate closer to the reported national average of 22%. To accomplish this, we used our actual data on cesarean rates associated with a given model score (Table 3) to predict what the rate would be in a hypothetical population with a cesarean delivery rate of 22%, under the assumption that the distribution of characteristics would be the same for each patient class, as in our population. These results are provided in Table 4. For any given score, the percentage of patients who would deliver by cesarean delivery is lower than that of our case-control population, but there were still groups who would be very low risk (2%) and high risk (65%).
We also evaluated the model in hypothetical populations with cesarean delivery rates from 20% to 8% (in 2% decrements). It is clear that the ability to predict risk of cesarean delivery decreases as the rate of cesarean delivery decreases in a given population. For example, in a hypothetical population with a cesarean delivery rate of 16% (closer to our estimated rate for nulliparous women at term with a single vertex presentation presenting in spontaneous labor), the predicted rate of cesarean delivery in the highest risk group was only 56%, but this would still be more than 3.5-fold greater risk than in the overall hypothetical population, and the predicted rate in the lowest risk group would be only 1.1% (Table 4). Thus, this score may be helpful in decision making for a woman and her physician. They may be reassured by a predicted rate of vaginal delivery of 99% in the lowest risk group, but a risk of cesarean delivery as high as 56% might influence early decisions.
There have been attempts in the past to develop cesarean delivery risk scores, but these were not limited to nulliparous patients in labor. These seemed to more accurately identify women at low risk for cesarean delivery and to lack the ability to identify women who were at highest risk.9 Several studies have limited enrollment to nulliparous patients, or they separately analyzed the data for nulliparous patients but failed to determine an increased risk early in the course of labor.10,20,21 The study that most closely resembled ours22 also limited enrollment to nulliparous women in spontaneous labor at term with singleton vertex fetuses. That study was prospective and used operative delivery (midforceps and cesarean delivery) as a primary outcome. In addition, they attempted to identify variables that can be documented at admission and those that were available later in labor, as we did. Among those that were available at admission, maternal age and height, pregnancy weight gain, smoking status, gestational age, and admission cervical dilatation were independent predictors of operative delivery. When the variables that were available later in labor were added to the first model, maternal age and height, smoking status, presence of dystocia, epidural use, and fetal heart rate abnormalities were retained as important predictors.
Our study adds to the understanding of the factors that might place this very specific patient population at highest and lowest risk for cesarean delivery. Application of this model might allow us to avoid the pitfall of proceeding with a long labor or a failed operative vaginal delivery and to offer early cesarean delivery to those who are at highest risk after the first 2 hours of labor. Accordingly, this model more accurately applies to nulliparous women admitted in labor at term and undelivered 2 hours later. This model could be used to predict an individual patient's risk for cesarean delivery by calculating the score by using the previously defined formula, which could be entered into a personal digital assistant for easy application in the clinical setting. Then, the patient's estimated approximate risk of cesarean delivery could be quickly found by reading the rate corresponding to a given score range.
Before applying the model in a patient care setting, it must be validated prospectively in a series of consecutive nulliparous women admitted in labor. Future research should involve recruitment of a group of patients who meet entry criteria, applying the model, and then following them prospectively to observe the outcome of their delivery to determine whether the relationship we have observed in this study persists. It should also be acknowledged that the regression equation may vary from hospital to hospital because of local demographics and practices.
If the model remains precise in predicting an individual's cesarean delivery risk, it would be easy to use during labor and delivery by means of a palm handheld device. This would generate a score that would permit earlier decisions and avoid risks accompanying longer labors.
1. Taffel SM, Placek PJ, Moien M. 1988 US cesarean-section rate at 24.7 per 100 births—A plateau? N Engl J Med 1990;323:199–200.
2. Clarke SC, Taffel S. Changes in cesarean delivery in the United States, 1988 and 1993. Birth 1995;22:63–7.
3. American College of Obstetricians and Gynecologists. Evaluation of cesarean delivery. Washington, DC: American College of Obstetricians and Gynecologists, 2000.
4. Lilford RJ, van Coeverden de Groot HA, Moore PJ, Bingham P. The relative risks of caesarean section (intrapartum and elective) and vaginal delivery: A detailed analysis to exclude the effects of medical disorders and other acute pre-existing physiological disturbances. Br J Obstet Gynaecol 1990;97:883–92.
5. Gibbs RS. Clinical risk factors for puerperal infection. Obstet Gynecol 1980;55:178S–84S.
6. Gordis L. Epidemiology. Philadelphia: WB Saunders, 1996.
7. Martel M, Wacholder S, Lippman A, Brohan J, Hamilton E. Maternal age and primary cesarean section rates: A multivariate analysis. Am J Obstet Gynecol 1987;156:305–8.
8. Mor-Yosef S, Samueloff A, Modan B, Navot D, Schenker JG. Ranking the risk factors for cesarean: Logistic regression analysis of a nationwide study. Obstet Gynecol 1990;75:944–7.
9. Hueston WJ. Development of a cesarean delivery risk score. Obstet Gynecol 1994;84:965–8.
10. Harlow BL, Frigoletto FD, Cramer DW, Evans JK, Bain RP, Ewigman B, et al. Epidemiologic predictors of cesarean section in nulliparous patients at low risk. RADIUS Study Group. Routine Antenatal Diagnostic Imaging with Ultrasound Study. Am J Obstet Gynecol 1995;172:156–62.
11. Osborne JW. Prediction in multiple regression. Prac Assess Res Eval 2000;7.
12. Gordon D, Milberg J, Daling J, Hickok D. Advanced maternal age as a risk factor for cesarean delivery. Obstet Gynecol 1991;77:493–7.
13. Dougherty CR, Jones AD. Obstetric management and outcome related to maternal characteristics. Am J Obstet Gynecol 1988;158:470–4.
14. Parsons MT, Winegar A, Siefert L, Spellacy WN. Pregnancy outcomes in short women. J Reprod Med 1989;34:357–61.
15. Gregory KD, Ramicone E, Chan L, Kahn KL. Cesarean deliveries for Medicaid patients: A comparison in public and private hospitals in Los Angeles county. Am J Obstet Gynecol 1999;180:1177–84.
16. Braveman P, Egerter S, Edmonston F, Verdon M. Racial/ethnic differences in the likelihood of cesarean delivery, California. Am J Public Health 1995;85:625–30.
17. Gould JB, Davey B, Stafford RS. Socioeconomic differences in rates of cesarean section. N Engl J Med 1989;321:233–9.
18. Carpenter MW, Soule D, Yates WT, Meeker CI. Practice environment is associated with obstetric decision making regarding abnormal labor. Obstet Gynecol 1987;70:657–62.
19. Goyert GL, Bottoms SF, Treadwell MC, Nehra PC. The physician factor in cesarean birth rates. N Engl J Med 1989;320:706–9.
20. Sims CJ, Meyn L, Caruana R, Rao RB, Mitchell T, Krohn M. Predicting cesarean delivery with decision tree models. Am J Obstet Gynecol 2000;183:1198–206.
21. Irwin DE, Savitz DA, Bowes WA Jr, St Andre KA. Race, age, and cesarean delivery in a military population. Obstet Gynecol 1996;88:530–3.
22. Turcot L, Marcoux S, Fraser WD. Multivariate analysis of risk factors for operative delivery in nulliparous women. Canadian Early Amniotomy Study Group. Am J Obstet Gynecol 1997;176:395–402.