Fecal sphingolipids predict parenteral nutrition-associated cholestasis in the neonatal intensive care unit

Abstract

Background: Parenteral nutrition-associated cholestasis (PNAC) in the neonatal intensive care unit (NICU) causes significant morbidity and associated healthcare costs. Laboratory detection of PNAC currently relies on elevated serum conjugated bilirubin levels in the aftermath of impaired bile flow. Here, we sought to identify fecal biomarkers, which when integrated with clinical data, would better predict risk for developing PNAC.

Methods: Using untargeted metabolomics in 200 serial stool samples from 60 infants, we applied statistical and machine learning approaches to identify clinical features and metabolic biomarkers with the greatest associative potential for risk of developing PNAC. Stools were collected prospectively from infants receiving PN with soybean oil-based lipid emulsion at a level IV NICU.

Results: Low birth weight, extreme prematurity, longer duration of PN, and greater number of antibiotic courses were all risk factors for PNAC (P < 0.05). We identified 78 stool biomarkers with early predictive potential (P < 0.05). From these 78 biomarkers, we further identified 12 sphingomyelin lipids with high association for the development of PNAC in precholestasis stool samples when combined with birth anthropometry.

Conclusion: We demonstrate the potential for stool metabolomics to enhance early identification of PNAC risk. Earlier detection of high-risk infants would empower proactive mitigation with alterations to PN for at-risk infants and optimization of energy nutrition with PN for infants at lower risk.

Keywords: early detection; infant; metabolomics; neonatal intensive care unit; parenteral nutrition-associated cholestasis; sphingomyelin; stool.

Figure 1

Clinical characteristics of infants with and without parenteral nutrition–associated cholestasis (PNAC). (A) Continuous clinical variables were tested for a statistically significant difference between the control and PNAC groups using a Wilcoxon rank sum test. We determined that five of the six variables are statistically different (P < 0.05). (B) The comparison of two clinical metrics, birth weight and birth weight percentile, reveals that there are simple thresholds that classify infants in our cohort as high or low risk. Infants born above the 40^th weight percentile (adjusted for gestational age) and also weigh >1.1 kg at birth are at a low risk of developing PNAC compared with the rest of the infants. (C) Although PNAC diagnosis is correlated with the amount of time an infant receives PN (panel A), two‐thirds of infants diagnosed with PNAC received PN for >20 days.

Figure 2

Metabolomics data and predictive biomarker selection. (A) There are 60 infants in this study and 19 were diagnosed with parenteral nutrition–associated cholestasis (PNAC). We collected 200 fecal samples with accompanied metabolomics data. The samples from each infant are plotted on an individual plot. The clinical conjugated bilirubin threshold, used to diagnose PNAC, is displayed with a dashed gray line on each panel. There are nine cholestatic infants for which we were able to collect fecal samples before conjugated bilirubin levels were above the diagnostic threshold of 1 mg/dl, as indicated by the dashed blue circle. (B) The y‐axis displays the scaled intensity for each metabolite. The case study samples were plotted as dashed blue lines across the boxplots. We selected biomarkers based on statistical significance (P < 0.05) and consistency among the case study samples.

Figure 3

Random forest machine learning with fivefold cross‐validation. (A) We performed feature reduction random forest machine learning to determine the minimal set of clinical metrics that provide the greatest predictive potential in our cohort. The optimal random forest consists of two clinical metrics and has an average fivefold cross‐validation overall accuracy of 57%. Birth weight percentile adjusted by gestational age and birth weight each contribute significantly to this model. (B) When we include the 78 biomarkers that we identified to have predictive potential, we are able to generate a set of random forest models with >70% cross‐validation accuracy on average. The second set of models also demonstrates that the two previously identified clinical variables maintain predictive potential when in the context of the stool biomarkers. These models use all of the metabolomics samples in this study and classify samples as high or low bilirubin levels; therefore, they do not predict if an infant will develop parenteral nutrition–associated cholestasis from the early stool samples collected.

Figure 4

Biomarkers with the strongest discriminatory accuracy. This analysis includes only the infants who fall outside of the low‐risk group identified in Figure 1, with a birth weight percentile of >40% and a birth weight of >1.1 kg. Additionally, we reduced the number of stool samples to only include the first sample for each infant. There are 12 metabolites from our complete set of 78 that demonstrate particularly accurate discriminatory potential within our cohort. These metabolites range from being 88% to 82% accurate at classifying the infants in our cohort based on only the first fecal sample that was collected for each infant. Although these accuracies are not properly validated with independent data, they demonstrate that there are several metabolites present in neonatal intensive care unit stool samples that have predictive capabilities. All 12 of these metabolites are various types of sphingomyelin.

Figure 5

The 12 best biomarkers show high agreement across our cohort. There is one infant in particular who contributes the majority of false‐negative classifications across all 12 metabolites. Among the infants, we see only false‐positive classifications when using an ensemble majority vote across the 12 metabolites. There were four false‐positive classifications based on majority vote and one false‐negative classification, resulting in an overall accuracy of 85%.