Data processing and statistical analyses

Automated measurements of skin layer thickness involved the use of two different software programs for digital image processing. The DermaLab^® embedded software provided dermal measurements (Fig 1A and 1B) but not epidermal thickness. For the epidermal axial dimension estimation, we developed a dedicated software in Python language [24]. The ultrasonography skin image (Fig 1A) with a 356- × 276-pixel resolution automatically identified the epidermis (white layer in Fig 1C, which was separated in a new frame in Fig 1D). The calculation of the epidermal thickness for each line in the image array was based on the size of the pixel in the frame obtaining 356 thickness values. Using a bootstrap technique in the selected sample size with 70 lines, the software resampled the thickness values 2,000 times, obtaining a frequency distribution. From a Gaussian distribution adjustment curve to the epidermis measurement, the peak was the thickness value, and the full width at half-maximum (FWHM) area was the sigma error value. During the image quality analysis, the measurements with non-normal distributions were promptly indicated to promote a visual inspection of the ultrasonography image.

(A) Skin ultrasonography image. (B) The red lines and the grid in between indicate the area over which dermal thickness is calculated using the DermaLab^® software. (C) The red color corresponds to the automated epidermal boundary detection by our dedicated software. (D) The white and black limits of the epidermis correspond to the automated mean thickness estimation by our dedicated software.

Descriptive statistics was used to assess the clinical, skin measure, and environmental variables during the assessment. Depending on the data distribution, quantitative variables were presented as means (95% CI), standard deviations (SDs), medians (minimum and maximum), or interquartile ranges (IQRs). The magnitude of the skin layer thickness was presented using histograms of frequency. Qualitative variables were presented as absolute values and percentages. The neonatal characteristics and skin layer thickness of the infants were described in accordance with the standard fetal growth (AGA, SGA, and LGA) [21, 22] and compared using Kruskall-Wallis or Chi-square tests.

Growth standard-dependent differences in the skin layer thickness were determined using a one-way analysis of variance or Kruskal-Wallis according to the data distribution and post hoc test. A regression analysis was performed to determine the correlation between GA and skin thickness parameters for each body site. Nonlinear models were adjusted to fit the correlation between predictors and outcomes better. The best body site on the newborn skin to correlate with GA was inferred from the regression coefficients obtained in the scatter plot of skin layer thickness versus GA. Multiple regression analysis included predictor variables from the univariate models, considering the effect modifiers from incubator stay, and the standard fetal growth, using the enter method of model arrangement. Coefficients of determination (adjusted R²) were determined on the basis of the hypothesis that all coefficients were 0. A normality test for the residual analysis was performed. The statistical program SPSS^® 22.0 was used for the analysis. The significance level adjusted for the hypothesis test was set at 5% with 95% CIs.