Triplet and higher-order multiple gestations have become more common with the advent of assisted reproductive technologies and ovulation induction. In the last three decades there has been a 500% increase in triplet and high-order births.1 Although multifetal births account for only 3% of live births, they are responsible for a large proportion of perinatal morbidity and mortality.2–4 Perinatal mortality rates are 20 times higher for triplet gestations compared with singleton gestations.
Several studies have shown that the pattern of growth between a singleton gestation and higher-order multiple gestations is different.5–7 Inappropriate fetal growth is a significant complication in multifetal pregnancies. Although there are singleton growth curves that help obstetricians diagnose intrauterine growth restriction, there is little information about the growth patterns of triplets.
Previous published work on triplet ultrasound growth curves was limited by analyses not properly adjusting for the dependence of repeated measurements over time on the same fetus and between fetuses within the same pregnancy.6,8–10 Although several authors have published triplet growth curves,5–11 these authors used a linear regression method, whereas we developed curves using multilevel statistical models to properly account for similarities of growth of fetuses within a mother as well as multiple measurements over time for each fetus. This methodology is important clinically because it may provide more accurate growth curves to evaluate triplet growth.
MATERIALS AND METHODS
In this longitudinal study, we reviewed all triplet pregnancies from 1992–2004 managed at Tufts-New England Medical Center, a single tertiary center. The study sample was limited to triplet pregnancies meeting the following inclusion criteria: 1) pregnancy progressed past 20 weeks of gestation without spontaneous miscarriage, 2) absence of congenital anomalies as documented by the newborn record, 3) no history of selective reduction, and 4) absence of fetal death. Table 1 gives a breakdown of maternal demographics including age, race, parity, smoking status, and body mass index. Twenty-seven pregnancies were excluded for missing records (23), fetal demise (3), and fetal anomaly (1). This study was approved by the institutional review board at Tufts-New England Medical Center. Data were extracted from routine ultrasound examinations done for clinical purposes on fetal measurements of estimated fetal weight, biparietal diameter (BPD), femur length, head circumference, and abdominal circumference. Measurements were done by certified sonographers with real-time imaging using 3.5-MHz transducers, and all images were reviewed by board-certified or board-eligible perinatologists. The ultrasound equipment used included two Phillips HDI 5000 machines (Advanced Technology Laboratories, Bothell, WA), one ATL-Phillips HDI 3000 (Advanced Technology Laboratories), two Acuson 128/XP/10 machines (Acuson, Mountain View, CA), and one Ultramark-9 HDI machine (Advanced Technology Laboratories). Biparietal diameter, head circumference, abdominal circumference, and femur length were measured whenever possible. Estimated fetal weight was derived from Hadlock's formula, with femur length, head circumference, and abdominal circumference,12 or abdominal circumference and femur length by Hadlock et al13 when BPD measurements were not possible. The number of fetuses without a BPD were neglible. These formulas were applied to all fetal measurements in our study. Fetal measurements were performed multiple times on each mother between 6 and 37 weeks of gestation, although the actual schedule of measurements differed between mothers. Sonograms to assess fetal growth were obtained every 2 to 4 weeks, from 16 to 18 weeks until delivery. In total, there were 123 mothers, 368 fetuses, and 1,398 time points included in the derivations of the charts for fetal weight. Totals were slightly different for other growth parameters due to minor differences in missing data patterns. Data from the very early and late weeks of gestation where there were fewer than 30 fetal measurements available for analysis were excluded.
Proper computation of the growth curves necessitates accounting for variation in growth within a single fetus over time, variations in growth between multiple fetuses within a single mother, and variations in fetal growth between mothers.14 To account for these different levels of associations, we used multilevel models,15 which are also referred to as random-effects, hierarchical, or mixed models. This approach involves specifying two or more levels of relationships among study variables and parameters. These levels are arranged in a hierarchy; hence the approach is also referred to as hierarchical modeling. Ordinary regression techniques are different in that they only represent one level of the hierarchy. For the multilevel models, we considered level 1 to account for variation between weights at different gestational ages (over time) within a fetus, level 2 to account for variation between fetuses within a mother, and level 3 to account for variation between mothers. The final form of the multilevel model used included terms for the fixed effects for linear and quadratic gestational age. The model also included random effects for level 1, the variation over time within a fetus using an autoregressive covariance structure, and level 3, the variation between mothers using an unstructured covariance structure. Inclusion of a random effect term for level 2, the variation between fetuses within a mother, did not improve the model, took a very long time to generate the intervals, and added complexity to the model and therefore was not included in the models used to generate the prediction intervals. In a larger database, however, the inclusion of the level 2 term may be more important.
We used a combination of a methodology described by Jiang and Zhang16 and bootstrapping to create percentiles of normal fetal weight, or prediction intervals, at each gestational age.14,17–19 A series of 100 bootstrap samples (with replacement) of 123 mothers and all measurements for the fetal weight for their associated fetuses were generated. A multilevel model was created for each bootstrap sample resulting in 100 multilevel models for weight. For each fetus at each time point within each sample we then calculated a predicted weight based on the estimates of the fixed effects from the quadratic model from that bootstrap sample: Predicted weight = Intercept + B1 (gestational age) + B2 (gestational age) (2)
The residuals were then computed for each fetus at each time point for each sample as the difference between the actual weights minus the predicted weights. The distribution of the residuals at each time point, across all bootstrapped samples, were used to create prediction intervals of weights at each week as follows. First, the 10th, 50th, and 90th percentiles of the residuals at each time point were calculated. These intervals are based on the rank order of the actual residuals and are therefore referred to as “distribution-free” prediction intervals by Jiang and Zhang.16 Next, the mean predicted weight, across all bootstrapped samples, was calculated for each gestational age. Finally, the associated residual (10th, 50th, and 90th percentiles) for each gestational age was added to the mean predicted weight for that gestational age. This same approach was repeated to generate the percentiles for each of the other growth parameters. The SAS System for Windows 9.3 (SAS Institute, Inc., Cary, NC) was used to do these analyses.17
One hundred fifty triplet pregnancies were identified. Twenty seven (18.0%) pregnancies were excluded for the following reasons: missing records (23), fetal demise (3), and fetal anomaly (a triplet with an omphalocele; 1). The gestational age range was restricted to 14–34 weeks because few scans were available after 34 weeks or before 14 weeks. Table 2 shows distribution-free percentiles of estimated fetal weight at each gestational age. Tables 3 through 6 show these percentiles for abdominal circumference, head circumference, BPD, and femur length. Figures 1 through 4 show medians, 10th and 90th percentiles for abdominal circumference, head circumference, BPD, and femur length. Figure 5 shows medians, 10th and 90th percentiles for estimated fetal weight. If the calculated percentiles are accurate, then actual percentage of fetuses that fall within the calculated 5th to 95th percentile should be 90%, or have a coverage of 90% of the fetuses. Figure 6 shows the coverage of various fetal weight intervals for the entire cohort, as well as for the fetuses measured each week. This figure shows that the intervals are accurate overall and have close-to-the-correct coverage for each gestational week. Accuracy of the model was tested by comparing the percentages of actual patients who fell within the calculated coverage intervals and replicating the analysis with random subsets of the data. Note the slightly larger deviations at the more extreme times may be a reflection of the smaller sample size, which is also plotted over time in this figure.
To make an informed clinical judgement about a growth-restricted fetus, accurate standards for intrauterine fetal growth must be available. Triplet pregnancies are often felt to be complicated by intrauterine growth restriction because measurement of growth is based on singleton growth curves. Use of singleton growth curves to plot intrauterine growth restriction of multiple pregnancies is controversial. As demonstrated by Alexander et al5 in 1998 using National Center for Health Statistics, fetal growth in triplets does not follow the growth curves of singletons or twins. They compared birth weights of singletons born alive in 1995 with birth weights of triplets and twins born alive from 1991 to 1995 and found that fetal growth in triplets begins to deviate from singletons and twins at around 31 weeks of gestation. After 31 weeks, the birth weights were significantly lower at each gestational age. Similarly, Blickstein et al7 in 2002 did a cohort study of 3,238 liveborn triplets and found that, after 33 weeks of gestation, triplet birth weights were lower at each gestational age. Min et al20 in 2004 did a cohort study of 188 pregnancies of liveborn triplets and showed that well-grown triplets fell below singletons by 30 weeks of gestation and twins after 34 weeks of gestation. Studies using ultrasonographic measurements of triplet gestations also found that triplet growth patterns deviate from singletons. For example, Weissman et al6 in 1990 found that triplet growth patterns deviated from singletons starting at 28 weeks of gestation. By standard singleton growth curves, more than 50% of triplets are small for gestational age by 35 weeks of gestation, and by 38 weeks more than 80% of triplets are small for gestational age.5
There is no consensus in the literature regarding whether singleton curves should be used for triplet gestations. Some authors recommend the use of specific charts generated from twin pregnancies (Sokol, 1984) and others suggest that the use of singleton growth curves for twins but not triplets. 21 The available data suggests that triplet growth falls off the singleton growth curve at later gestational ages. The reason for this is unknown. Hata et al23 (1991) suggested that deposition of soft tissue seen in normal singletons in the third trimester occurs to a lesser extent in normal twins and triplets. Since triplet growth tends to fall off the singleton growth curve at later gestational ages, a limitation of our study is that we excluded data from the very late weeks of gestation where there were fewer than 30 fetal measurements available for analysis. This could be addressed by using a much larger sample size.
We derived growth curves based on ultrasonographic data rather than data based on live births. There is inherent error in using ultrasonography, and this error is known to increase with later gestational ages. There is little data regarding the accuracy of ultrasonography in multiple gestations. Lynch et al24 (1995) did a retrospective analysis of ultrasound data for fetuses that underwent an ultrasound examination one week before delivery (singletons 1,832, twins 518, triplets 51). They found that the accuracy of estimated fetal weight in triplets was the same for singletons at weights below 2,500 g. They also found no difference in accuracy of ultrasound measurments between twins and singletons greater than 2,500 g and concluded that ultrasound estimation of fetal weight is as accurate in twins and triplets as it is in singletons. The greatest triplet weight in our study was 2,510 g (95th percentile at 34 weeks of gestation). The study by L ynch et al24 is applicable to this study in that almost all the weights in our triplet gestations are below 2,500 g. Using this study, we can postulate that the estimated fetal weights in our study are at least as accurate as singleton weights. However, more data are needed to address the question of whether accuracy of ultrasonography in multiple gestations differs from that of singleton gestations.
Several other authors have attempted to create triplet growth curves.6,8–10 However, these authors have used the individual regression-lines method, which we believe create less-accurate fetal growth curves than the multilevel model method. They either can underestimate the true variation by not accounting for clustering of measurements from the three fetuses within a single mother, or they may lose information by not using all available data and using only one fetus' measurement per mother. The multilevel modeling methodology is an important and powerful technique because many kinds of data have a clustered structure or longitudinal structure and this methodology allows one to properly adjust for these associations. For example, offspring from the same parents tend to be more alike in their physical characteristics than individuals chosen at random from the population at large. The multilevel approach allows the statistical modeling of data in which a variable (such as weight or length) is repeatedly measured on the same individuals, giving rise to important correlations within a subject. It is able to accommodate measurements made at unequal intervals, and is efficient even when some data are randomly missing, because data are pooled across subjects in the estimation procedure. Thus, using multilevel models to create triplet growth curves is a more appropriate method than the standard linear-regression lines method. The growth curves we have derived are unique because we used multilevel models to take into account variation in growth within a single fetus over time, variations in growth between multiple fetuses within a single mother, and variations in fetal growth between mothers. In addition to the methodology we used, when comparing our data to previously published data based on triplet ultrasound measurements (123 triplet sets compared with 33 [Rodis et al8], 24 [Weissman et al6], 40 [Fountain et al9], 47 [Shushan et al25], 12 [Kuno et al10], and 36 [Mordel et all2411]) our sample size is larger, and thus, clinicians may be more confident in the validity of our results. When comparing our curves with the previously published curves, there is a difference of a few percent (3–5%) between the measurements we obtained and the measurements obtained by the linear regression models. Although only a small difference, we maintain that the curves we derived are more clinically accurate because of the methodology and sample size and may provide a more accurate estimate of fetal weight. The anticipated use of these curves is for the prenatal diagnosis of intrauterine growth restriction. Intrauterine growth restriction is defined with a rigid cutoff. With a more accurate cutoff, fewer triplets will be misclassified as intrauterine growth restricted. These curves allow clinicians the opportunity to determine which measurements are better predictors of triplet growth.
1. Martin JA, Hamilton BE, Sutton PD, Ventura SJ, Menacker F, Munson ML. Births: final data for 2002. Natl Vital Stat Rep 2003;42(10):1–102.
2. Multiple gestation: complicated twin, triplet, and high-order multifetal pregnancy. ACOG Practice Bulletin No. 56. American College of Obstetricians Gynecologists. Obstet Gynecol 2004;104:869–83.
3. Kiely JL, Kleinman JC, Kiely M. Triplets and higher-order multiple births: time trends and infant mortality. Am J Dis Child 1992;146:862–8.
4. Luke B, Keith LG. The contribution of singletons, twins, and triplets to low birth weight, infant mortality, and handicap in the United States. J Reprod Med 1992;37:661–6.
5. Alexander G, Kogan M, Martin J, Papiernik E. What are the fetal growth patterns of singletons, twins, and triplets in the United States? Clin Obstet Gynecol 1998;41:1:115–25.
6. Weissman A, Jakobi P, Yoffe N, Zimmer EZ, Paldi E, Brandes JM. Sonographic growth measurements in triplet pregnancies. Obstet Gynecol 1990;75:324–8.
7. Blickstein I, Jacques D, Keith L. Total and individual triplet birth weights as function of gestational age. Am J Obstet Gynecol 186:1372–5.
8. Rodis JF, Arky L, Egan JF, Borgida AF, Leo MV, Campbell WA. Comprehensive fetal ultrasonographic growth measurements in triplet gestations. Am J Obstet Gynecol 1999;181:1128–32.
9. Fountain SA, Morrison JJ, Smith SK, Winston RM. Ultrasonographic growth measurements in triplet pregnancies. J Perinat. Med 1995;23:257–63.
10. Kuno A, Akiyama M, Yanagihara T, Hata T. Comparison of fetal growth in singleton, twin, and triplet pregnancies. Hum Reprod 1999;14:1352–60.
11. Mordel N, Laufer N, Zajecek G, Shalev Z, Lewin A, Schenker J, et al. Sonographic growth curves of triplet conception. Am J Perinatol 1993;10:239–242.
12. Hadlock FP, Harrist RB, Sharman RS, Deter RL, Park SK. Estimation of fetal weight with the use of head, body, and femur measurements: a prospective study. Am J Obstet Gynecol 1985;151:333–7.
13. Hadlock FP, Harrist RB, Carpenter RJ, Deter RL, Park SK. Sonographic estimation of fetal weight. Radiology 1984;150:535–40.
14. Royston P, Altman DG. Design and analysis of longitudinal studies of fetal size. Ultrasound Obstet Gynecol 1995;6:307–12.
15. Goldstein H. Multilevel models in educational and social research. London: Griffin & Co; 1987.
16. Jiang J, Zhang W. Distribution-free prediction intervals in mixed linear models. Statistica Sinica 2002;12:537–53.
17. Littell RC, Henry PR, Ammerman CB. Statistical analysis of repeated measures data using SAS procedures. J Anim Sci 1998;76:1216–31.
18. Altman DG, Chitty LS. Charts of fetal size: 1. Methodology. Br J Obstet Gynaecol 1994;101:29–34.
19. Royston P, Wright EM. How to contruct “normal ranges” for fetal variables. Ultrasound Obstest Gynecol 1998;11:30–8.
20. Min SJ, Luke B, Min L, Misiunas R, Nugent C, Van de Ven C, et al. Birthweight references for triplets. Am J Obstet Gynecol 2004;191:809–14.
21. Socol ML, Tamura RK, Sabbagha RE, Chen T, Vaisrub N. Diminished biparietal diameter and abdominal circumference growth in twins. Obstet Gynecol 1984;64:235–8.
22. Sheh YG, Graham D, Stinton SK, Render TD. Biparietal diameter growth in uncomplicated twin gestation. Am J Perinatol 1987;4:229–32.
23. Hata T, Deter RL, Hill RM. Reduction of soft tissue deposition in normal triplets. J Clin Ultrasound 1991;19:541–5.
24. Lynch L, Lapinski R, Alvarez M, Lockwood CJ. Accuracy of ultrasound estimation of fetal weight in multiple pregnancies. Ultrasound Obstet Gynecol 1995;6:349–52.
© 2006 The American College of Obstetricians and Gynecologists
25. Shushan A, Mordel N, Zajicek G, Lewin A, Schenker JG, Sadovsky E. A comparison of sonographic growth curves of triplet and twin fetuses. Am J Perinatol 1993;10;388–91.