Over the past three decades, there has been substantial global progress in reducing maternal mortality.1 This progress has been observed in many developing and industrialized countries; however, the United States remains the only industrialized nation with a rising rate of maternal mortality.2 From 2000 to 2014, the maternal mortality rate in the United States increased 26% to 18 per 100,000 live births.2 In response to this observed trend, there have been renewed efforts by organizations such as the Alliance for Innovation on Maternal Health to reduce maternal mortality and morbidity.3 In 2016, core Alliance for Innovation on Maternal Health partners, the American College of Obstetricians and Gynecologists and the Society for Maternal-Fetal Medicine, in collaboration with the Centers for Disease Control and Prevention, released an Obstetric Care Consensus on the screening for and review of severe maternal morbidity.4 Although there is no consensus on the conditions and complications that constitute severe maternal morbidity, the underlying concept is that these outcomes may be preventable, provide an opportunity to improve obstetric care in at-risk women, and serve as a surrogate or “near miss” for maternal mortality.4 The Consensus proposed a two-criteria screening protocol for severe maternal morbidity: 1) transfusion of 4 or more units of blood and 2) admission of pregnant or postpartum woman to an intensive care unit (ICU), which achieves a high positive predictive value of 0.85 for the identification of severe maternal morbidity.5–7 Subsequent population-based research has demonstrated that maternal ICU admission was the most frequent (45%) severe maternal morbidity indicator among women who died in the peripartum period.8
Although there are known high morbidity-risk obstetric conditions, such as morbidly adherent placenta, that confer drastically increased risk for maternal morbidity, there is limited literature regarding common maternal and obstetric comorbidities and the composite effect of these factors on the likelihood of maternal ICU admission. We sought to evaluate prenatal risk factors associated with maternal ICU admission to develop a predictive model that would help identify women at higher relative risk of maternal ICU admission.
We performed a retrospective population-based cohort study of all live births between 20 and 44 weeks of gestation in the United States during 2012–2016. Frequency of maternal ICU admission was determined for the time period, using U.S. live birth records. The primary objective of this study was to identify prenatal factors associated with increased risk of maternal ICU admission to build a multivariable predictive model to estimate the association of these factors on ICU admission risk using variables that could be obtained before delivery to aid in medical decision making such as delivery location. Data for this study were obtained from the National Center for Health Statistics U.S. live birth records, using the 2003, newest revision of the U.S. Standard Certificate of Live Birth.9,10 Methods for live birth record collection, preparation, quality control, and validation have been previously described.9–11 More than 99% of all live births occurring in the United States are registered in the deidentified National Center for Health Statistics publicly available linked birth–death database.9 Observed ICU admission rates reported in our study were consistent with those previously published.12–14 These data did not meet criteria for human subject research by federal standards and therefore was exempt from review by the institutional review board at our institution. The separate database used for external validation was obtained from the Ohio Department of Health after approval by the Human Subjects Institutional Board Review (IRB 2018-70). The outcome variable, maternal ICU admission, was abstracted from medical and delivery records.9,10
Variables considered for use in the predictive model included factors that could be readily obtained from the prenatal record before delivery. These factors included: 1) maternal demographic variables such as maternal age, race, cigarette smoking status, socioeconomic status (type of health insurance, educational level, use of Women Infants and Children supplementation), and whether prenatal care was received; 2) variables representing preexisting medical comorbidities (pregestational diabetes, chronic hypertension, and obesity); 3) variables indicative of pregnancy-specific conditions associated with adverse pregnancy outcomes (assisted reproductive technology, interpregnancy interval, parity, gestational diabetes, gestational hypertension, preeclampsia, weight gain during pregnancy, and sexually transmitted infections such as gonorrhea, hepatitis B or C, syphilis, and chlamydia); and 4) variables representative of other obstetric and delivery characteristics (prior cesarean delivery, fetal plurality, gestational age at delivery, fetal presentation, and route of delivery [cesarean not preceded by induction or labor]). The variable, scheduled cesarean delivery, was constructed from women who underwent a cesarean birth not including women who had a trial of labor before cesarean. Live births occurring before 20 0/7 weeks of gestation or after 44 6/7 weeks of gestation were excluded from analysis.
Gestational age was obtained using the best obstetric estimate variable in the live birth record.10 Prepregnancy weight and height were used to calculate body mass index (BMI, calculated as weight in kilograms divided by height in meters squared), as recorded in the medical record. Body mass index class was categorized by World Health Organization classification.15 Prepregnancy weight, prepregnancy BMI, maternal weight at delivery, and weight gain during pregnancy were all analyzed as both continuous and categorical variables in the predictive model using standard scores (z-score) to identify the most predictive variable or discriminatory cut point.
Differences in baseline characteristics between women admitted to the ICU in the peripartum period were compared with those not admitted to the ICU using Student's t test for continuous data and Chi square for categorical data. For continuous variables, optimal dichotomized cut-points were determined to maximize sensitivity and specificity and then normalized for comparison with the original continuous variable.16 The variable format most predictive of outcome was chosen for the final regression model. All factors noted to be significant in univariate analyses were selected for inclusion in the multivariable logistic regression model to estimate the adjusted odds ratios (aORs) of each variable on the outcome of maternal ICU admission. The final regression model was developed using a stepwise backward selection process of the most significant variables as determined by normalized z-scores and data exploratory analysis until a parsimonious model of the 14 most significant variables was obtained. Receiver operating characteristic curve analysis was performed for this predictive multivariable model. The maximum likelihood estimations from the regression model were used to create a logistic regression equation for prediction of ICU admission (Table 1). A calculator was created to represent the logistic regression model, which can be used to generate a patient-specific risk for ICU admission, https://ob.tools/mat-icu-calc). We conducted a k-fold cross-validation in which the dataset was randomly divided into 10 equal groups for internal validation of the predictive model (Appendix 1, available online at http://links.lww.com/AOG/B412).17,18 The predicted minimal and maximal risk for maternal ICU admission from the original dataset ranged 0.00–24.7%. Predicted risk of ICU admission exceeding 5% was rare in the delivery cohort. As such, the model was considered to be overfitted beyond this cutoff, as the number of predictor variables in the regression model (14) would exceed a limit of 10 events per variable according to Peduzzi et al.19,20 Sensitivity, specificity, positive likelihood ratios (LR), positive predictive value (PPV), and negative predictive value (NPV) for the various risk cutoff points were calculated for the predictive model (Appendix 2, available online at http://links.lww.com/AOG/B413).
Using a separate, external database of live births collected from Ohio (2006–2011, n=856,255), external validation of the predictive model was performed using a Hosmer-Lemeshow goodness-of-fit test. We evaluated whether observed binary responses, ICU admission, conditional on a vector of selected risk factors, were consistent with predictions up to 5% predicted risk. The predicted probabilities of ICU admission were partitioned into 10 groups (0–0.25%, 0.25–0.5%, 0.5–0.75%, 0.75–1.0%, 1.0–1.25%, 1.25–1.5%, 1.5–2.0%, 2.0–3.0%, 3.0–4.0%, 4.0–5.0%). For each partitioned group, the observed proportion of women admitted to the ICU was plotted on a calibration curve. Next, the mean expected probability of ICU admission for each predicted risk group with 95% CI were plotted against the observed rate.
The variable multifetal gestation was evaluated for inclusion in the predictive model but did not reach statistical significance in the adjusted model. The final model thus excluded multifetal gestations to limit one outcome (ICU admission) per each live birth. Statistical analyses were performed using STATA Release 15.1.
There were 19,844,580 live births in the United States during 2012–2016, of which 18,745,615 delivered between 20 0/7 and 44 6/7 weeks of gestation and had imputed data on maternal morbidity. Among the mothers of these live newborns, 27,602 (0.15%) were admitted to the ICU in the peripartum period (Fig. 1). Women were excluded if gestational age at delivery was missing, delivery occurred at less than 20 weeks of gestation, greater than 44 weeks, or if data were not recorded regarding maternal morbidity (n=1,098,965, 5.5%, Fig. 1). Women admitted to the ICU were more likely to be of advanced maternal age, non-Hispanic black race, have Medicaid insurance, smoke cigarettes, and have a sexually transmitted infection during the pregnancy (Table 2). As for comorbid medical conditions, women admitted to the ICU were more likely to have pregestational diabetes, gestational hypertension or preeclampsia, chronic hypertension, and gestational diabetes (Table 2). Women admitted to the ICU had higher rates of obesity, excessive weight gain (more than 22.67 kg [50 lb]) during pregnancy, delivery weight greater than 136 kg (300 lb), and BMI of 50.0 or more. They were also more likely to have had a prior cesarean birth, prior preterm birth, and had longer interpregnancy intervals (Table 3). The mean gestational age at delivery was 35.9 vs 38.5 weeks between those admitted to the ICU and those not admitted. When corrected for gestational age, there was a higher prevalence of small for gestational age birth weights in the ICU admission group. Cesarean delivery without labor or induction preceding delivery was more common in the ICU cohort (73.1% vs 32.2%, P<.001) as seen in Table 3.
Using stepwise backward selection, we built a multivariable model to predict the outcome, maternal ICU admission, using the 14 most significant factors (https://ob.tools/mat-icu-calc). We ranked the 14 most significant variables selected for the predictive model of maternal ICU admission in descending order of statistical significance (Table 4). The receiver operating characteristic curve for this model achieved an area under the curve (AUC) of 0.81 (95% CI 0.80–0.81, Fig. 2). K-fold cross-validation was performed for internal validation with 10 randomly and equally partitioned cohorts and revealed an AUC 0.80 (95% CI 0.79–0.81) for the predictive model (Appendix 2, http://links.lww.com/AOG/B412). External validation of the logistic regression equation using a separate cohort of live births was performed using a dataset of live births between 2006 and 2011 obtained from Ohio (n=856,255) that included 859 (0.10%) women admitted to the ICU. This model demonstrated consistent findings, achieving an AUC of 0.83 (95% CI 0.82–0.84) (Fig. 3). The Hosmer-Lemeshow goodness-of-fit test revealed a well-calibrated observed-to-expected risk for ICU admission up to 5% (Fig. 4). Sensitivity, specificity, positive LR, PPV, and NPV for predictive model can be found in Appendix 3 (http://links.lww.com/AOG/B413). Using a low cut point of 0.1% predicted risk for ICU admission would detect 87.9% of women admitted to the ICU but would have a PPV of 0.3%. Alternatively, a relatively high cut point of 5.0% or more predicted risk for ICU admission, achieved a PPV of only 4.0% with a sensitivity of 2.3% (Appendix 3, http://links.lww.com/AOG/B413).
Maternal morbidity is an important obstetric quality indicator as serious adverse maternal events may lead to either death or have disabling consequences. Additionally, it has previously been estimated that there is opportunity for improvement in outcome in nearly 45% of severe maternal morbidity cases.21 Monitoring severe maternal morbidity as a surrogate for maternal mortality is valuable given the rarity of maternal mortality events and the potential for which intervention may prevent a maternal death or permanent disability. Although it is important for obstetricians to identify, triage, and counsel at-risk women, there are few guidelines and practical tools available for risk prediction. Risk can be estimated based on either individual risk factors or a subjective assessment of multiple coexisting risk factors. Many well-known, albeit rare, high-risk conditions such as critical maternal cardiac disease (mitral stenosis), active thromboembolic disease, and morbidly adherent placenta are commonly known to be associated with significant maternal morbidity and are typically triaged to a regional center for delivery based on the individual risk factor alone.22 However, increasing rates of obesity, hypertensive disorders, diabetes, and abdominal deliveries also contribute significantly to rising maternal morbidity rates.23,24 As such, we intended to build a model that could estimate an individualized risk based on common maternal and obstetric conditions that may cumulatively increase a woman's risk and therefore assist in delivery planning and allocation of appropriate resources.
Grobman et al, previously developed a scoring system to identify severe maternal morbidity and found placenta accreta as the most significant risk factor associated with severe morbidity. Preterm delivery, antenatal anticoagulant use, cigarette use, hypertension, diabetes mellitus, abruptio placentae, and prior cesarean delivery were also associated with severe maternal morbidity to lesser degrees. Unfortunately, certain obstetric risk factors such as placental abruption occur concurrently with the severe maternal morbidity event and thus are difficult to anticipate. In their model, which had an AUC of 0.8, using a higher score cutoff resulted in a PPV of 39% but missed 93% of women with severe maternal morbidity owing to the overall low prevalence of severe maternal morbidity in the population, whereas using a low cutoff point identified more women with severe maternal morbidity but had a PPV of less than 1%.25
Although maternal ICU admission is only one of many indicators on the pathway to maternal death, it has been validated to be predictive of maternal mortality and is endorsed by the American College of Obstetricians and Gynecologists as one of two indicators for severe maternal morbidity screening.4 Recently, Ray et al8 evaluated severe maternal morbidity indicators among 1.9 million hospital births in Ontario and found ICU admission as the most prevalent indicator (45%) for maternal death among 10 morbidity indicators evaluated. Panchal et al studied risk factors associated with ICU admission and found advanced maternal age, black race, maternal hospital transfer, and cesarean delivery as all associated with ICU admission.26 We previously demonstrated that preterm birth was a significant risk factor for composite adverse maternal outcomes, with periviable birth in particular being associated with a sixfold increased risk for adverse maternal outcome and a 9.6-fold increased risk for ICU admission.27 Consistent with these findings, we found decreasing gestational age (weeks) was highly predictive of ICU admission (aOR 0.86, 95% CI 0.8–0.87, z-score −33.0) and found to be the second most predictive factor, only behind scheduled cesarean delivery (aOR 3.0, CI 2.9–3.1, z-score 57.4). Body mass index as a continuous variable was not associated with ICU admission, however extreme obesity (BMI 50 or more) was found to be moderately associated with the outcome.
Limitations to this study are notable in that the dataset for this study was obtained using national vital statistics data, which include potential inaccuracies in reporting.11,28 In addition, there were missing data for 1,098,965 (5.5%) women regarding maternal morbidities, which, if not randomly selected, could limit the precision of estimates. Women in this cohort may have had more than one delivery during this time period and thus been included more than once, which may lead to overestimation of the influence for each variable. We are unable to exclude multiple deliveries from the same woman as each birth was de-identified and not linked. Although underreporting of medical comorbid outcomes such as ICU admission is possible, the ICU admission rates observed in our study (0.15%) were consistent with those previously published, which range 0.06–0.42%.12–14 Additionally, there are likely regional and hospital variation differences between the degree of morbidity for a maternal patient admitted to the ICU at one institution compared with another, which was not accounted for in our study. We also excluded multifetal gestations from the primary analysis and therefore this study is not generalizable to this group. The retrospective nature of our study precludes the determination of causality between factors evaluated and the outcome of ICU admission. Women with other significant known morbidities such as morbidly adherent placenta or those with significant cardiac, renal, or pulmonary disease would be expected to have a priori risk exceeding the risk range evaluated in our model.
Despite an AUC higher than 0.8, the model cannot precisely identify women who will and will not be admitted to the ICU, as the outcome of ICU admission is a rare event and therefore the PPV of the model remains low. For instance, the PPV in our study remained unacceptably low (4%) even when using a high-risk prediction cutoff (ie, 5.0%) and missed in the identification of nearly 98% of ICU admission cases at this level. Nonetheless, even with a low PPV (less than 5%) for all risk categories, the positive LRs for the risk categories of 0.6%, 1.5%, 3.0%, and 4.0% remain relatively high at 12, 20, 28, and 32, respectively. Thus, we believe this model could be clinically useful in triaging patients found to be at substantially increased risk (ie, 30-fold increased risk) to delivery hospitals with higher levels of maternal care, with 24 hour anesthesiology, obstetrician, and possibly perinatology coverage, and to delivery hospitals with massive transfusion and ICU availability.
We created an online calculator to represent the logistic regression model, which can be used to generate a patient-specific risk for ICU admission. This calculator is available online at https://ob.tools/mat-icu-calc.
1. Alkema L, Chou D, Hogan D, Zhang S, Moller AB, Gemmill A, et al. Global, regional, and national levels and trends in maternal mortality between 1990 and 2015, with scenario-based projections to 2030: a systematic analysis by the UN Maternal Mortality Estimation Inter-Agency Group. Lancet 2016;387:462–74.
2. MacDorman M, Declercq E, Cabral H, Morton C. Recent increases in the U.S. maternal mortality rate: disentangling trends from measurement issues. Obstet Gynecol 2016;128:447–55.
4. Severe maternal morbidity: screening and review. Obstetric Care Consensus No. 5. American College of Obstetricians and Gynecologists. Obstet Gynecol 2016;128:e54–60.
5. You WB, Chandrasekaran S, Sullivan J, Grobman W. Validation of a scoring system to identify women with near-miss maternal morbidity. Am J Perinatol 2013;30:21–4.
6. Main EK, Abreo A, McNulty J, Gilbert W, McNally C, Poeltler D, et al. Measuring severe maternal morbidity: validation of potential measures. Am J Obstet Gynecol 2016;214:643.e1–10.
7. Geller SE, Rosenberg D, Cox S, Brown M, Simonson L, Kilpatrick S. A scoring system identified near-miss maternal morbidity during pregnancy. J Clin Epidemiol 2004;57:716–20.
8. Ray JG, Park AL, Dzakpasu S, Dayan N, Deb-Rinker P, Luo W, et al. Prevalence of severe maternal morbidity and factors associated with maternal mortality in Ontario, Canada. JAMA Netw Open 2018;1:e184571.
11. Martin JA, Osterman MH, Kimeyer SE, Gregory EC. Measuring gestational age in vital statistics data: transitioning to the obstetric estimate. Natl Vital Stat Rep 2015;64:1–20.
12. Oud L. Epidemiology of pregnancy-associated ICU utilization in Texas: 2001–2010. J Clin Med Res 2016;9:143–53.
13. Wanderer JP, Leffert LR, Mhyre JM, Kuklina EV, Callaghan WM, Bateman BT. Epidemiology of obstetric-related ICU admissions in Maryland: 1999-2008*. Crit Care Med 2013;41:1844–52.
14. Curtin SC, Gregory KD, Korst LM, Uddin SFG. Maternal morbidity for vaginal and cesarean deliveries, according to previous cesarean history: new data from the birth certificate, 2013. Natl Vital Stat Rep 2015;64:1–13.
15. Obesity: preventing and managing the global epidemic: report of a WHO consultation. World Health Organ Tech Rep Ser 2000;894:i–xii, 1–253.
16. Liu X. Classification accuracy and cut point selection. Stat Med 2012;31:2676–86.
17. LeDell E, Petersen M, van der Laan M. Computationally efficient confidence intervals for cross-validated area under the ROC curve estimates. Electron J Statist 2015;9:1583–607.
18. Kohavi R. A study of cross-validation and bootstrap for accuracy estimation and model selection. In: Mellish CS, editor. IJCAI’95 proceedings of the 14th international joint conference on artificial intelligence. Vol 2. San Francisco (CA): Morgan Kaufmann Publishers Inc; 1995.
19. Peduzzi PN, Concato J, Holford TR, Feinstein AR. The importance of events per independent variable in multivariable analysis, II: accuracy and precision of regression estimates. J Clin Epidemiol 1995;48:1503–10.
20. Peduzzi PN, Concato J, Kemper E, Holford TR, Feinstein AR. A simulation study of the number of events per variable in logistic regression analysis. J Clin Epidemiol 1996;49:1373–9.
21. Ozimek JA, Eddins RM, Greene N, Karagyozyan D, Pak S, Wong M, et al. Opportunities for improvement in care among women with severe maternal morbidity. Am J Obstet Gynecol 2016;215:509.e1–6.
22. Levels of maternal care. ACOG Obstetric Care Consensus No. 9. American College of Obstetricians and Gynecologists. Obstet Gynecol 2019;134:e41–55.
23. Burlingame J, Horiuchi B, Ohana P, Onaka A, Sauvage LM. The contribution of heart disease to pregnancy-related mortality according to the pregnancy mortality surveillance system. J Perinatol 2012;32:163–9.
24. May AL, Freedman D, Sherry B, Blanck HM; Centers for Disease Control and Prevention (CDC). Obesity—United States, 1999–2010. MMWR Suppl 2013;62(suppl 3):120–8.
25. Reichman NE, Schwartz-Soicher O. Accuracy of birth certificate data by risk factors and outcomes: analysis of data from New Jersey. Am J Obstet Gynecol 2007;197:32.e1–8.
26. Panchal S, Arria A, Harris P. Intensive care utilization during hospital admission for delivery: prevalence, risk factors, and outcomes in a statewide population. Anesthesiology 2000;92:1537–44.
27. Rossi RM, DeFranco EA. Maternal complications associated with periviable birth. Obstet Gynecol 2018;132:107–14.
28. Grobman WA, Bailit JL, Rice MM, Wapner RJ, Reddy UM, Varner MW, et al. Frequency of and factors associated with severe maternal morbidity. Obstet Gynecol 2014;123:804–10.
Supplemental Digital Content
© 2019 by the American College of Obstetricians and Gynecologists. Published by Wolters Kluwer Health, Inc. All rights reserved.