Medical Student Université de Sherbrooke Université de Sherbrooke
Disclosure(s):
Jonathan Lauzon-Schnittka, n/a: No financial relationships to disclose
Background: Measures of cardiac dimensions in children must be normalized for body size to account for growth. Z-scores are the most widely used method to do this. The statistical methods used for the creation of Z-scores have been debated for over two decades. The most popular approach has been to normalize dimensions for only one anthropometric variable, most often body surface area (BSA). However, these bivariate models have been shown to introduce a bias to body mass index adjusted for age (BMIz). While multivariable models have been proposed to correct this bias, their optimal form for a wide range of cardiac dimensions remains to be defined. OBJECTIVES. This study aimed to compare a range of multivariate regression methods to identify the best approach to create unbiased Z-scores for all body sizes and compositions.
METHODS AND RESULTS: The reference echocardiography databank of the Canadian Congenital and Pediatric Cardiology Research Network was used. It holds data on > 20,000 normal echocardiograms from 8 Canadian institutions. Four multivariable models were tested with height and weight as normalizing variables, with the aim of minimizing residual associations to height, weight, age, BSA and BMIz. Non-parametric and parametric models were tested with the lambda-mu-sigma (LMS) transformation, as well as conventional parametric multinomial models. These were compared to a BSA-based LMS model.
Significant residual associations to BMIz were found in the BSA-based model, corroborating the evidence that bivariate models are biased in lean and obese populations. Parametric multivariable models reduced this bias. Particularly, a parametric multinomial model where weight was empirically corrected to account for obesity yielded the lowest mean Z-score deviations across cardiac dimensions. Non-parametric multivariable models, despite their complexity, were unexpectedly prone to residual errors. The LMS transformation significantly improved the normality of the Z-score distributions for left ventricle dimensions, but this improvement was marginal for all other measures. The Z-score distributions from three models for the pulmonary valve diameter are presented in the Figure. They are representative of the general results: both multivariable models improved the error seen in the BSA-based model. The model with corrected weight produced results closest to the expected distribution.
Conclusion: Multivariable models with height and weight as normalizing variables perform generally well and reduce the systematic bias related to adiposity that is seen in bivariate models, such as BSA-based models.
