Factors affecting the growth of infants diagnosed with cystic fibrosis by newborn screening

A cohort of 144 children diagnosed with CF by NBS was identified. This included all babies with a positive CF NBS from the West Midlands Regional Screening Laboratory between 1st November 2007 and 31st October 2014 who had a confirmed diagnosis of CF (identification of two CF-causing mutations and/or a sweat chloride ≥60 mmol/L). Five children categorized as Cystic Fibrosis Screen Positive Inconclusive Diagnosis (CF SPID) [11] were excluded. Two children had died and two moved out of area. A further six patients were excluded due to incomplete datasets. The remaining 129 children attended one of the two tertiary CF centres in the region: Birmingham Children’s Hospital and Royal Stoke University Hospital. Although these two centres work independently, both follow the UK CF Trust and European CF Society Standards of care resulting in similar clinical practices [12, 13].

Demographic and newborn screening data were collected. These included date of diagnosis (taken as date of first clinic visit or date of diagnosis of meconium ileus), gender, ethnicity, birth weight, the presence of meconium ileus, mean immunoreactive trypsinogen, CFTR mutations, sweat chloride, faecal elastase (measured at first clinic appointment), weight at first clinic visit and mode of feeding (exclusive breast / exclusive formula / mixed). All weight and length measurements recorded in the first 2 years of life were then collated. Insufficient data points were available for head circumference to allow meaningful analysis. We also recorded age at first isolation of Pseudomonas aeruginosa and Staphylococcus aureus. These data were obtained from the West Midlands NBS Laboratory database, patient’s clinical notes and the electronic results systems at the two CF Centres. All height and weight measurements were converted into z scores using the WHO-UK anthropometric calculator. This corrects for sex and gestational age. To summarise growth, the following four outcomes were calculated for each child: mean weight z score in first year of life, mean weight z score in second year of life, mean height z score in the first year of life and mean height z score in second year of life. The median time for children to achieve height and weight z scores of − 2, − 1 and 0 was calculated and children who did not regain their birth weight z score in the first 2 years of life were identified.

Before classification modelling could be undertaken, the cohort had to be validated using a clustering algorithm (k-means) [14]. This ensured it was representative of the UK CF population. The anthropometric measurements as well as the demographic, newborn screening and microbiology data were included as inputs and the mathematically generated clusters were scrutinised by the CF clinicians (MD, WC and FG) to ensure they corresponded to clinical phenotypes. Decision-tree classification models were then generated to determine the variables that predicted future growth. Inputs were the demographic, newborn screening and infection variables. The model analyses all the inputs but only uses those with predictive effect in the decision trees. The four outcomes were the mean z scores for height and weight in the first and second year of life. These outputs were chosen as the z scores frequently varied significantly over time meaning a z score from a single time point would not have been representative. Each classification model was validated using a stratified five-fold cross-validation method. This approach is used when the sample size does not allow for a representative split (such as 70% / 30%) for model training and validation sets. The data is split into five sets ensuring each set has proportional representation of all classes in each set. Modelling is repeated five times each using a different fifth of the data for validation. The final model reported is the one obtained using all the data, however, the error reported is the mean error out of the five repeats.

The mean (SD) birth weight of the entire cohort was 3.17 (0.51) kg giving a mean (SD) z score of − 0.32 (1.13). The mean (SD) birth weight was 3.21 (0.50) kg for boys and 3.12 (0.53) kg for girls. The mean (SD) birth weight z scores were − 0.35 (1.07) for boys and − 0.30 (1.20) for girls. Using a faecal elastase cut-off of 200 μg/g, 113 (88%) children were PI and 16 (12%) PS. When reviewed at their first clinic visit, 29 (22%) infants were exclusively breast-fed, 60 (47%) were exclusively formula milk and 40 (31%) were mixed feeding. There was no difference in mode of feeding between PI and PS. All babies were started on standard treatment on the day of diagnosis. The difference in weight between PI and PS infants was not significant at birth but became so by the time of the first clinic visit. Height and weight was significantly lower in PI children in the first year of life. The difference was less in the second year and no longer significant. See Table 1. In the first 2 years of life, 18 (14%) of infants failed to regain their BW z score, all of these were pancreatic insufficient. The median time for children to achieve a weight z score of − 2, − 1 and 0 was 18, 33 and 65 weeks respectively. The median times to reach the same z scores for height were 30, 51 and 90 weeks.

All the children were on oral Staphylococcus aureus prophylaxis. Staphylococcus aureus was isolated in 16 of the 129 children in the first 2 years of life. The median (IQR) age of the first isolate was 0.22 (0.08–0.63) years. Pseudomonas aeruginosa was isolated in 45 children in the first 2 years of life. The median (IQR) age of the first isolate was 1.18 (0.82–1.49) years. All these isolates were non-mucoid. No child met the Leeds Criteria for chronic Pseudomonas aeruginosa infection in the first 2 years of life [15].

The unsupervised cluster analysis identified two distinct groups. See Table 2 in Appendix. The CF clinicians agreed the clusters were clinically relevant as they represented PI and PS phenotypes. This validated the cohort. Decision tree classification modelling was therefore undertaken to identify the variables that could be used to predict the growth outcomes listed above. See Fig. 1. Across the four models, birth weight z score, change in weight z score from birth to first clinic, faecal elastase, isolation of Pseudomonas aeruginosa, isolation of Staphylococcus aureus and sweat chloride were the variables of use in predicting future growth. Importantly, the mode of feeding was not identified as a predictor of infant growth by any of the classification models.