Beruflich Dokumente
Kultur Dokumente
Regression models for predicting men- S ONOCRAPHIC measurement of the fetal biparietal diameter (BPD)
strual age based on real-time sonographic can provide a good estimate of fetal age in the first half of preg-
measurements of four fetal parameters nancy (2 SD = ±7-10 days) (1, 2), but there is a progressive increase
(biparietal diameter, head circumference, in the variability as pregnancy progresses, with a maximum van-
abdominal circumference, and femur ability of approximately ±3.6 weeks in the last six weeks of pregnancy
length), used alone and in combination, (1, 2). For this reason, efforts have been made to predict fetal age from
were developed in a cross-sectional study other fetal growth parameters such as the head circumference (HC)
of 361 fetuses between 14 and 42 men- (3), abdominal circumference (AC) (4), and femur length (FL) (5, 6).
strual weeks. The head circumference and However, the variability patterns in predicting age from these pa-
femur length were the strongest individ- rameters used individually are similar to those obtained when BPD
ual predictors of age. A number of combi- is used alone (3-7).
nations of fetal parameters, including the In another report we demonstrated statistically significant im-
combination of head circumference and provement in estimation of fetal age in the third trimester of preg-
femur length, provided age estimates that nancy when two or more of these measurements are used in combi-
were significantly better (p 0.05) than nation (8). This method resulted in a reduction in the variability of
those using any single parameter alone. It approximately 25 to 30%, as well as a reduction in the maximum errors
was also demonstrated that simply aver- that are observed when an individual parameter such as the BPD is
aging individual age estimates in a given used alone (8). In this study we expanded the original study popuba-
case could provide results that were not tion from 177 to 361 fetuses to evaluate this method of fetal dating
significantly different from those ob- throughout a broader range of gestational ages (14-42 weeks).
tamed by using the same parameters in a
complex regression equation. The advan-
MATERIALS AND METHODS
tages and potential pitfalls of this system
of fetal dating are discussed. The study group consisted of 361 middle-class Caucasian women from the
Houston area. All of these patients had a history of regular menses, and knew
Index terms: Fetus, growth and development unequivocally the beginning day of the last menstrual period. In all cases the
first-trimester clinical findings were in agreement with the last menstrual
Radiology 1984; 152: 497-501 period. Patients with maternal disease that might adversely effect fetal growth
(e.g., diabetes melbitus) were not included; similarly, patients with multiple
gestations in this pregnancy were not included. In keeping with proper design
of a cross-sectional study, each fetus was measured only once in gestation.
All examinations were performed by physicians using a commercially
available linear array real-time ultrasound system with a 3.5-MHz focused
transducer (ADR-Tempe, AZ). The techniques for measuring the BPD, femur
length, head circumference, and abdominal circumference (Fig. 1) are de-
scnibed in detail elsewhere (2-5). Measurements of the BPD and femur length
were made using electronic calipers. Circumference measurements of the head
and abdomen were either made directly from Polaroid images using an
electronic digitizer (Numonics Corp.) or by calculation from two diameters
using the formula for the circumference of a circle (D1 + D2 X 11/2); these
methods have bee#{252} shown to give equivalent results (9).
Regression models for predicting menstrual age from these parameters,
both individually and in all possible combinations, were obtained by stepwise
regression analysis (10). The regression analysis for the individual parameters
1 From the Department of Radiology, Jefferson included the linear, quadratic, and cubic terms of each parameter, and the
Davis Hospital (F.P.H., S.K.P.), the Departments of Ra-
stepwise regression for the combinations of parameters include the linear,
diology (F.P.H., S.K.P.) and Obstetrics and Gynecolo-
quadratic, and cubic terms of the individual parameters as well as all the
gy (R.L.D.), Baylor College of Medicine, and the De-
partment of Biometry, The University of Texas School cross-products of these terms. For example, stepwise regression analysis of
of Public Health (R.B.H.), Houston, TX. Received Sept. menstrual age as a function of all four parameters included the following
7, 1983; accepted and revision requested Oct. 31; revi- terms: BPD, BPD2, B1’D3, HC, HC2, HC3, AC, AC2, AC3, FL, IL2, FL3, BPD X
sion received March 2, 1984. AC, BPD X HC, BPD X FL, HC X AC, HC X FL, AC X FL. The optimal re-
©l984,RSNA cg gression function chosen by this analysis is the one with the highest r2 and
497
a. b. c.
a. Demonstration of the appropriate axial section of the fetal head for measurement of the BPD (broken squares in vertical axis) and head cm-
cumfenence (broken lines). The small arrowhead is a landmark thought to represent the cavum septi pellucidi.
b. Demonstration of the appropriate axial section of the fetal abdomen for measurement of the abdominal circumference. In this case circum-
ference is calculated using the formula (Dl + D2) X 1.57. The small arrowhead indicates the umbilical portion of the left portal vein and
the large arrowhead indicates the stomach.
c. Demonstration of the appropriate section for measurement of the femur length (+ indicates electronic caliper markers).
the lowest standard deviation, and the ability with increasing menstrual age. The struab age in the six-week subintervabs
method requires that all terms in the re- results using the simple averaging tech- using the individual parameters and
gression equation be statistically signifi- nique in these intervals were also evaluated,
the combinations of parameters is in-
cantly different from zero (p < 0.05). and were compared with the results using
dicated in TABLE II. There was no sig-
The ability of various combinations of the the regression models to determine if there
nificant difference between the van-
individual parameters to predict menstrual were significant differences within these
ability among the optimal models
age accurately was also evaluated by a sim- age intervals.
ple averaging technique (8), which gives within each subintervab, but the results
equal weight to each age estimate based on RESULTS using these models were significantly
our previously published nomograms (2-5) better than those using the best mdi-
for each parameter. For example, if all four
The results from the stepwise me- viduab models. It appears from this
parameters were to be evaluated as a group, gression analysis of menstrual age on analysis that the HC is the most accu-
individual estimates of age would be made the individual parameters and the nate individual parameter in predicting
based on BPD, HC, AC, and FL, and these combinations of parameters are sum- age, a point that has also been observed
four estimates would simply be added to- manized in TABLE I. The results using recently by Law and MacRae (11).
gether and divided by four to provide a the simple averaging technique were There was a progressive increase in
composite age estimate (8).
not significantly different (p < 0.05) variability with time between 12 and
The efficacy of each individual parameter
from those using the various regression 30 weeks for all the individuab param-
and various groups of parameters as esti-
models. The best results (as defined by etems and groups of parameters, but no
mators of fetal age was assessed by com-
paring the magnitude of the variability for
the lowest standard deviation, the significant differences in variability
each regression model, which is repre- highest m2 value, and the smallest were demonstrated when the van-
sented in this context by the standard de- maximum error) were obtained using ability for a given regression model
viation of the regression. The F test was all four parameters in combination between 30 and 36 weeks was com-
used to determine if there were significant (Fig. 2). The results using this model pared with the variability for the same
differences (p = 0.05) among the various were significantly better
(p = 0.05) regression model during the period of
models (10); each regression model was
than any of the individual parameters 36 to 42 weeks (p = 0.05).
compared with every other regression
used alone, and were also
significantly
model in this manner. This test was also
better than the combinations of BPD
used to determine if there were significant DISCUSSION
differences in the variability between the
and AC, BPD and HC, HC and AC, and
regression model and the simple averaging BPD, HC, AC. Theme were no signifi- The vamiabibity in predicting men-
technique when both used the same mdi- cant differences, however, in the van- strual age from any individual fetal
vidual parameters to produce a composite abilities of the remaining combinations measurement such as BPD, head cm-
age estimate. of parameters. It would appear from cumfenence, abdominal circumference,
Finally, the 361 known menstrual age the magnitude of the standard devia- or femur length is known to increase
data points were divided into six-week in- tions, the m2 values, and the maximum progressively throughout gestation,
tervals by age as follows: 12-17.99 weeks,
errors that five of these regression reaching a maximum variability of
18-23.99 weeks, 24-29.99 weeks, 30-35.99
models could be considered optimal ±3.5 weeks in the third trimester of
weeks, and 36-41.99 weeks. The regression
models. These include the combina- pregnancy (1-7). Ourapproach in this
models obtained from the stepwise regres-
sion analysis described in the preceding tions of HC and FL, BPD, AC, and FL, investigation, in contrast with a pre-
paragraph were then evaluated as predic- BPD, HC, and FL, HC, AC, and FL, and vious study limited to the third tn-
tors of menstrual age in the subgroups, to BPD, HC, AC, and FL. mesten, was to develop regression
determine if there was any change in van- The variability in predicting men- equations for predicting menstrual age
498 Radiology
#{149} August 1984
TABLE I: Regression Equations for Predicting Menstrual Age (MA) from Fetal Measurements (12-42 wks)
throughout the entire range of gesta- TABLE II: Subgroup Variability in Predicting Menstrual Age Using the Regression
tion using various combinations of Equations in TABLE I
parameters. The rationale for using
Subgroup Variability (±2 SD) in Weeks
more than one parameter is as follows:
Fetal 12-18 Weeks 18-24 Weeks 24-30 Weeks 30-36 Weeks 36-42 Weeks
(a) the demonstration by other inves- Parameters (N = 43) (N 69) (N 76) (N 95) (N 78)
tigators that the use of a combination of
BPD ±1.19 ±1.73 ±2.18 ±3.08 ±3.20
parameters provides better results then HC ±1.19 ±1.48 ±2.06 ±2.98 ±2.70
a single parameter in estimating neo- AC ±1.66 ±2.06 ±2.18 ±2.96 ±3.04
natal age by the Dubowitz examination FL ±1.38 ±1.80 ±2.08 ±2.96 ±3.12
BPD, AC ±1.26 ±1.68 ±1.92 ±2.60 ±2.88
(12), estimating fetal weight using ul- BPD, HC ±1.08 ±1.49 ±1.99 ±2.86 ±2.64
trasound (13), determining fetal well- BPD, FL ±1.12 ±1.46 ±1.84 ±2.60 ±2.62
HC, AC ±1.20 ±1.52 ±1.98 ±2.68 ±2.52
being with a biophysical profile (14),
HC, FL ±1.08 ±1.34 ±1.86 ±2.52 ±2.28
and predicting fetal age in the first AC, FL ±1.32 ±1.64 ±1.88 ±2.66 ±2.60
trimester using ultrasound (15); (b) the BPD, AC, FL ±1.20 ±1.52 ±1.82 ±2.50 ±2.52
BPD, HC, FL ±1.04 ±1.35 ±1.81 ±2.52 ±2.34
fact that any of these parameters used HC, AC, FL ±1.14 ±1.46 ±1.86 ±2.52 ±2.34
alone may be limited not only by the HC, AC, BPD ±1.21 ±1.58 ±1.94 ±2.60 ±2.52
BPD, HC, AC, FL ±1.08 ±1.40 ±1.80 ±2.44 ±2.30
biological variability but also by subtle
technical problems that may be ap-
parent only to the very experienced
sonogmapher; (c) the observation that
in a normal fetus any of these parame- Figure 2
tens may be generally banger or smaller NUMBER ERROR(wks)
-3.0
than the mean value expected for the -2.5
menstrual age, and that these dis- -2.0
A histogram of the residuals
16 -1.5
crepancies are not always in the same 46 -1.0 (e.g. , differences between ac-
64 -0.5
direction, e.g., the fetus with a 75th 77 0.0 tual age and predicted age)
67 0.5 using the regression model
percentile head size and a 25th per- 42 1.0
based on all four fetal mea-
centile body size (16); and (d) routine 18 1.5
2.0 surements (residuals rounded
measurements of BPD, head cincum- 2.5
3.0 to the nearest half week).
ference, abdominal circumference, and
femur length are now considered part *.2 cases
of a standard ultrasound examination
(17, 18).
The results in this study demonstrate
that there is a significant reduction in
the overall variability and the maxi- The choice of parameters used to this case coubd be weaker because head
mum observed errors when an optimal predict age in a given case will depend circumference has a stronger relation
combination of parameters is used to both on the philosophy of the sono- to age than the BPD and is more shape
estimate age instead of any single pa- graphem and the number of technically independent. In our department we
rameter. For example, if one chose to satisfactory measurements obtained. use all four measurements routinely,
use the BPD alone to estimate age in For exampbe, the data in TABLE I dem- since BPD and abdominal circumfer-
place of using the optimal combination onstrate that head circumference and ence are already available because of
of four parameters, the overall van- femur length in combination provide their robe in estimating weight (13).
ability as estimated by the standard age estimates that are not significantly Another important feature of our me-
deviation of the regression would be different (p = 0.05) from those obtained sults is that they provide information
increased by 33% (1.02 vs. 1.36 wks), using all four measurements in com- about the magnitude of overall van-
and the maximum observed error bination. One could argue justifiably ability and maximum errors that one
would be increased by 55% (3.2 vs. 5.1 that only these two measurements are can expect when certain measurements
wks). We feel the magnitude of these needed to predict age; a similar angu- are technically impossible (TABLES I,
differences warrants routine use of this ment could be made for the use of BPD II).
dating method. and femur length in combination, but These regression models are easily