Several articles have proposed echocardiographic reference values in normal pediatric subjects, but adequate validation is often lacking and has not been reviewed. The aim of this study was to review published reference values in pediatric two-dimensional and M-mode echocardiography with a specific focus on the adequacy of the statistical and mathematical methods used to normalize echocardiographic measurements. All articles proposing reference values for transthoracic pediatric echocardiography were reviewed. The types of measurements, the methods of normalization, the regression models used, and the methods used to detect potential bias in proposed reference values were abstracted. The detection of residual associations, residual heteroscedasticity, and departures from the normal distribution theory predictions were specifically analyzed. Fifty-two studies met the inclusion criteria. Most authors (87%) used parametric normalization to account for body size, but their approaches were very heterogeneous. Linear regression and indexing were the most common models. Heteroscedasticity was often present but was mentioned in only 27% of studies. The absence of residual heteroscedasticity and residual associations between the normalized measurements and the independent variables were mentioned in only 9% and 22% of the studies, respectively. Only 14% of studies documented that the distribution of the residual values was appropriate for *Z *score calculation or that the proportion of subjects falling outside the reference range was appropriate. Statistical suitability of the proposed reference ranges was often incompletely documented. This review underlines the great need for better standardization in echocardiographic measurement normalization.

Echocardiography is a reliable, noninvasive tool to evaluate heart structure and function in children and adults. Many important clinical decisions are routinely based on the absolute sizes of cardiac structures. Evaluation is highly dependent on the quality of the measurements but also on the quality of the reference values with which these measurements are compared. The American Society of Echocardiography Pediatric and Congenital Heart Disease Council recently published recommendations for quantification methods during the performance of pediatric echocardiography. However, reference values for the proposed methods often lack adequate validation.

Unbiased reference values require appropriate “normal” subjects, standardized reproducible measurements, and appropriate sample sizes. In children, reference values are also highly dependent on accurate adjustment for body size. Although nonparametric approaches have sometimes been used, parametric methods, such as *Z* scores, are now becoming the standard for body size adjustment in pediatric echocardiography. However, parametric methods rely on an appropriate distribution of the data, on the absence of residual associations, and on constant variance of the normalized measurements throughout the entire sample. These important requirements have not always received the attention they deserve.

A recent review by Cantinotti *et al.* underlined several limitations of the available reference values in pediatric echocardiography, including a lack of standardization in data acquisition, a limited number of healthy subjects, and heterogeneous methods of normalizing and reporting reference values. However, their review did not specifically address the statistical methods used or the potential pitfalls of parametric normalization. In this article, we present a systematic review of available reference values in two-dimensional (2D) and M-mode echocardiography in infants, children, and adolescents with a focus on the statistical validity of the methods used to generate the proposed reference ranges. For each reviewed article, we analyzed how the reference values were estimated, what type of normalization was used, and how the authors documented the detection of potential bias.

## Methods

## Literature Search Strategy

A search of the National Library of Medicine’s PubMed database was performed using the Medical Subject Headings controlled vocabulary from the National Library of Medicine. The search strategy was built to retrieve all articles containing the Medical Subject Headings terms “echocardiography” and “reference values” or their equivalents: (reference values OR biometry OR anthropometry OR regression analysis) AND {echocardiography OR [ultrasonography AND (heart OR cardiovascular system)]}. We limited the search results to articles whose subjects were <18 years of age and that were available in English or French. References of selected articles were also reviewed for other, potentially missed relevant articles.

## Article Selection

The search was performed in July 2011. Titles and abstracts were first reviewed for identification of relevant studies. Full-text articles were reviewed if they matched the inclusion criteria or if they could not be confidently excluded from the abstracts alone.

We included studies proposing reference values for cardiac dimensions measured using 2D or M-mode imaging by transthoracic echocardiography in a normal healthy pediatric population. Articles were excluded when one or more of the following was applicable: inclusion of subjects with potential cardiopathies or other conditions that could alter cardiac dimensions, measures from imaging methods other than transthoracic 2D or M-mode echocardiography (fetal echocardiography, transesophageal echocardiography, etc), functional studies (ventricular function or strain, speckle tracking, tissue Doppler, etc), use of reference values from other studies, and publication before 1980. Thirteen studies focusing on reference values in pediatric echocardiography also included young adult subjects, and these were included in our analysis to better represent the upper range of growth.

## Data Abstraction and Analysis

Each selected article was reviewed separately by two authors (W.M. and F.D.) using a standard data collection form. Discrepancies were resolved by consensus. Information on the study subjects (age, reason for echocardiography, inclusion and exclusion criteria), the cardiac structures measured, and the echocardiographic techniques used (views, mode, cardiac cycle, etc) was extracted. We also noted if interobserver or intraobserver variability was considered by the authors.

We then thoroughly examined the method of adjustment for body size and reviewed the type and number of independent variables used. For parametric normalization, we extracted information on the type of regression, the mathematical transformation of the independent or the dependent variables, and how the authors justified their choice of regression strategy. Because most cardiac structures display inconstant variance across most growth variables (heteroscedasticity), we noted if heteroscedasticity was assessed, by what method (visual assessment of plots or statistical tests), and, if present, how its effect was taken into account in the regression models.

Finally, we extracted information on how the authors assessed their proposed reference values for the absence of bias. For any parametric normalization, we noted if the authors confirmed the absence of residual associations with the selected independent variable and the absence of residual heteroscedasticity. We also extracted information on the distribution of the residual values and on the adequacy of the proportion of subjects falling outside the normal ranges across the entire range of the independent variable.

## Results

## Search Results and Selected Articles

The search strategy returned 1,016 articles. Initial screening by title and abstract identified 117 potentially relevant articles. Sixty-five were further excluded (27 did not propose reference values, five were related to other echocardiographic modalities, 20 did not include pediatric data, five did not include normal subjects, and eight were conducted before 1980), leaving 52 articles for analysis.

The main characteristics of selected recent articles are summarized in Table 1 (2D studies) and Table 2 (M-mode studies). Reference values for almost all cardiac structures have been published, but the dimensions of the left ventricular outflow tract, ascending aorta, and coronary arteries; M-mode measurements of left ventricular size; and estimations of left ventricular volume and mass were the most commonly measured structures. Three studies included reference ranges for ≥10 structures.

Study | n |
Age | Structures | Model and type of normalization | Was the detection of potential biases documented? | |||
---|---|---|---|---|---|---|---|---|

Testing for heteroscedasticity | Testing for residual association | Testing for residual heteroscedasticity | Testing for the distribution of residual or normalized values | |||||

Lytrivi et al. (2011) |
100 | 0 to 3 y | LVEDV (bullet shape formulas) | Indexed to BSA ^{1.38 }(allometric model) |
No | Yes | No | No |

Dallaire and Dahdah (2011) | 1,033 | 2 mo to 18 y | Coronary arteries | Linear with square root of BSA | Yes | Yes | Yes | Yes |

Gautier et al. (2010) |
353 | Mean, 12 ± 4.5 y | AoV, sinus, STJ, AscAo | Log-log with BSA | No | No | No | Percentage of subjects with Z > 2 |

Olivieri et al. (2009) |
432 | 0 to 20 y | Coronary arteries | Log-log model with BSA | Yes | No | No | No |

Neilan et al. (2009) |
4,109 | 1 to 17 y | LA | Log-log model with BSA or weight | Yes | Yes | Yes | Yes |

Pettersen et al. (2008) |
782 | 0 to 18 y | 14 structures by 2D imaging, 7 structures by M-mode imaging | Third-order polynomial with log-transformed dependent variable | Unclear | No | No | No |

Kaldararova et al. (2007) |
702 | 0 to 20 y | AoV, aortic sinus, STJ | Nonparametric | NA | NA | NA | NA |

Suleymanoglu et al. (2007) |
213 | 0 to 15 y | RV volume | Nonparametric normalization; authors report 5th and 95th percentiles for several independent variables | NA | NA | NA | NA |

Poutanen et al. (2006) |
168 | 2 to 27 y | Mitral and aortic area | Linear with BSA (range presented as 5th and 95th percentiles) | NA | No | NA | No |

Zilberman et al. (2005) |
748 | 0 to 18 y | AoV, PV, TV, MV annuli | Log-log model with BSA | No | No | No | No |

Makan et al. (2005) |
250 controls, 900 elite athletes | 14 to 18 y | LV dimensions and mass, LA, aortic root | No normalization, mean ± 2SDs according to age | No | NA | NA | NA |

Poutanen et al. (2003) |
168 | 2 to 27 y | AoV, sinus of Valsalva, STJ, aortic arch | Linear with BSA (range presented as 5th and 95th percentiles) | NA | No | NA | No |

Tan et al. (2003) |
390 | 2 mo to 8 y | Coronary arteries | Linear (results presented with 5 different independent variables | No | No | No | No |

Joyce et al. (2001) |
44 | 0 to 17 y | RV free wall mass | Indexed to BSA | No | No | No | No |

Tan et al. (2001) |
62 | Preterm infants | AoV, PV, PAs | Linear with weight | No | No | No | No |

Tacy et al. (1995) |
70 | 0 to 10 d | AoV, PV, MV, TV | Linear with weight | No | No | No | No |

Sheil et al. (1995) |
48 | 0 to 23 y | LVOT, AoV, sinus of Valsalva, STJ, AscAo | Linear with height and indexed to AoV | No | No | No | No |

Nidorf et al. (1992) |
196 | 6 d to 18 y | AoV, LA, LVEDD, LV length | Indexed to height | No | Yes | No | No |

Domanski et al. (1991) |
10 | 4 to 12 y | LVEDD, LA | Indexed to LVOT dimension | No | Yes | No | No |

Pearlman et al. (1990) |
196 children, 72 adults | 6 d to 18 y (children) | LA | Power model with BSA | Yes | No | No | No |

Study | n |
Age | Structures | Type of normalization | Was the detection of potential biases documented? | |||
---|---|---|---|---|---|---|---|---|

Testing for heteroscedasticity | Testing for residual association | Testing for residual heteroscedasticity | Testing for the distribution of residual or normalized values | |||||

Nagasawa (2010) | 243 | 0 to 12 mo | LVEDD | Linear with height (with 2 inflexion points) | No | No | No | No |

Foster et al. (2008) |
440 | 0 to 21 y | LVM | Nonparametric. lambda-mu-sigma method | NA | NA | NA | NA |

Bonatto et al. (2006) |
595 | 1 mo to 12 y | 8 structures by M-mode imaging and LVM | Nonparametric (centile curves with BSA) | NA | NA | NA | NA |

Overbeek et al. (2006) |
747 | 0 to 18 y | Left ventricular dimensions | Log-log model with weight | Yes | Yes | Yes | No |

Kervancioglu et al. (2006) |
229 | 0 to 15 y | Aortic sinus | Nonparametric | NA | NA | NA | NA |

Huicho et al (2005) |
321 (subjects living at high altitude) | 2 mo to 19 y | 10 structures | Linear with BSA | No | No | No | No |

Joyce et al. (2004) |
45 | 0 to 4 mo | Left and right ventricular volumes, masses, wall thicknesses | No normalization, mean ± SD according to age | NA | NA | NA | NA |

Skelton et al. (1998) |
79 | Newborns including preterm infants | LV, Ao, LA | Nonparametric, mean and SD for strata of GA and BW | NA | NA | NA | NA |

Nagasawa et al. (1996) |
437 | 1 mo to 17 y | LVEDD | Linear with height | No | No | No | No |

Daniels et al. (1995) |
192 | 6 to 17 y | LV | Indexed to height (allometric model) | No | Yes | No | No |

Huwez et al. (1994) |
127 | 7 mo to 19 y | LV, aortic root, LA, RV | Linear with age and linear with BSA | No | No | No | No |

Malcolm et al. (1993) |
904 | 6 to 16 y | LVM | Multiple regression with height, age, and height × age interaction | No | No | No | No |

de Simone et al. (1992) |
444 | 4 mo to 23 y | LVM | Indexed to height ^{2.7 }(allometric model) |
No | Yes | No | Yes |

Gupta and Jain (1991) | 183 | 3 to 12 y | LV | Indexed to height | No | No | No | No |

## Population Studied

The populations studied ranged in age from infancy to early adulthood. Twelve studies included young adults up to 27 years of age. One study including adults of all ages was also included in our analysis because almost 75% of the subjects were children. Seven studies included only newborns, and three focused only on preterm infants. Most studies included strictly healthy subjects, but some did not exclude current or past history of a minor congenital heart defect such as a small atrial septal defect, a patent foramen ovale, or a small patent ductus arteriosus. One study included patients with histories of Kawasaki disease without documented coronary artery abnormalities. Twenty-one studies (39.6%) recruited subjects for research purposes only, and 18 studies (34.0%) relied on echocardiographic studies performed on clinical grounds but subsequently read as normal. The authors of the remaining 15 studies (28.3%) failed to specify why echocardiography was performed.

## Adjustment for Growth Parameters

Seven studies (13.5%) did not attempt parametric normalization and presented reference ranges as percentile limits or means and standard deviations stratified by age or weight. The majority of authors (45 studies [86.8%]) used various methods of parametric normalization to account for growth. Sixteen studies (31.4%) considered only one independent variable (body surface area [BSA] in nine studies). The authors of the remaining studies tested two or more independent variables, the most common being BSA, weight, and height. BSA was used as the independent variable in 22 studies (42.3%). Weight was used in seven studies (mostly for infants), and height was used in eight studies (mostly for older children). Normalization using multivariate regression was performed in nine studies.

The approach to parametric normalization was heterogeneous among the reviewed articles. Twenty of the studies (44.2%) using parametric normalization tested one regression model only. Of them, eight articles reported linear regression only, six reported indexing (including two using allometric indexing), two reported power or logarithmic models only, and three used multivariate regression but tested only one model. In the remaining reports, two or more models were assessed, usually linear, polynomial, power, and log-log models. Pearson’s correlation coefficient ( *R *^{2 }) was often used to assess the goodness of fit of the regressions. Various mathematical transformations of both the dependent and the independent variables were commonly used to deal with the nonlinearity of most echocardiographic measurements with growth parameters. Logarithmic transformation of either the dependent or the independent variable, or both, was the most common strategy to adjust for nonlinearity, followed by the square and cube roots of the independent variable.

## Unequal Variance and Heteroscedasticity

Heteroscedasticity is defined as the inconstant variance of the dependent variable across the entire range of the independent variable. In parametric normalization, one uses the magnitude of dispersion around the mean to establish reference ranges. Such ranges are thus dependent on any variation of that dispersion. Heteroscedasticity was mentioned in the text of 12 of the 45 studies in which some form of parametric normalization was done. However, there were four studies in which heterogeneous variance seemed to have been taken into account, although we could not find any mention of it in the text. The most common methods to account for increasing variance were logarithmic transformation of the dependent variables (seven studies) and weighted regression models (seven studies).

## Types of Reference Values

The reference values proposed by the selected studies were expressed in a variety of ways. *Z *score equations were suggested in eight studies. The remaining studies proposed graphs (14 studies), tables (two studies), equations to derive the upper and lower ranges of normal (12 studies), indexes (10 studies), means and standard deviations (four studies), or other methods (two studies).

## Evaluation of Bias

Only 14 studies using parametric normalization mentioned any type of further evaluation of the proposed reference values. The authors of 10 studies assessed if residual associations between the normalized measurements and the chosen independent variables were still present. The authors of six studies stated that the distribution of the dependent variable was appropriate for *Z *score calculation or that the proportion of subjects falling outside the reference range was appropriate. Testing for absence of residual heteroscedasticity was mentioned in only four studies. Finally, intraobserver and interobserver variability was evaluated in 13 and 19 studies, respectively.

## Results

## Search Results and Selected Articles

The search strategy returned 1,016 articles. Initial screening by title and abstract identified 117 potentially relevant articles. Sixty-five were further excluded (27 did not propose reference values, five were related to other echocardiographic modalities, 20 did not include pediatric data, five did not include normal subjects, and eight were conducted before 1980), leaving 52 articles for analysis.

The main characteristics of selected recent articles are summarized in Table 1 (2D studies) and Table 2 (M-mode studies). Reference values for almost all cardiac structures have been published, but the dimensions of the left ventricular outflow tract, ascending aorta, and coronary arteries; M-mode measurements of left ventricular size; and estimations of left ventricular volume and mass were the most commonly measured structures. Three studies included reference ranges for ≥10 structures.