Fecal sphingolipids predict parenteral nutrition–associated cholestasis in the neonatal intensive care unit (2025)

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

Sample collection and processing

Participants were chosen from a larger cohort of ongoing neonatal microbiome studies from the level IV NICU at Inova Fairfax Hospital in Virginia (Western Institutional Review Board approval number 20210065). Between 2016 and 2019, after informed consent, preterm and term neonates aged ≤5 days with an anticipated medical intermediate care unit length of stay of >5 days were recruited. We collected detailed maternal, pregnancy, and delivery data. While in the NICU, infants had stool collected twice weekly when possible. Stool was frozen at −80°C within 12 h. To address the aim of the current study, 60 infants were selected from the larger cohort who all had received PN for >5 days. Of these, all were receiving partial enteral nutrition at the time of the first stool sample collection. For stool metabolomics analyses, we enriched sample selection for PNAC cases and excluded samples in which PNAC was presumed to be secondary to infection. Infants were monitored for the development of cholestasis (defined as a conjugated bilirubin ≥ 1 mg/dl16), and the cause of cholestasis was noted as assessed by the treating physicians. Clinical data were collected at baseline and with each stool collection; this included gestational age, birth weight, antibiotic use, infections, and length of time receiving PN.

Statistical analysis

We analyzed the clinical metadata for each infant enrolled in the study as well as the metabolomics data for each of the fecal samples. For clinical data, Mann‐Whitney U tests were performed to correlate clinical metrics with conjugated bilirubin levels. For biomarker identification, if metabolites were correlated with conjugated bilirubin levels, Mann‐Whitney U tests were performed followed by a multiple testing correction using the Bonferroni method to identify metabolites differentially abundant in healthy vs control groups. Distributions were confirmed to be nonnormal using the Shapiro‐Wilk test.

Data processing and machine learning

We used both QIIME and DADA2 software to process the 16S data.25, 26, 27 The forward reads from the 16S data were used to align with the Greengenes 16S rRNA gene database to determine taxonomic calls. We processed the metabolomics data using Python, Jupyter notebooks, pandas, scipy, numpy, and seaborn. We generated the machine learning models using Scikit Learn in Python. All Python scripts and data are available on github (https://github.com/Tjmoutinho/TPNAC_data_processing).

Clinical metrics can identify an at‐risk population of infants receiving PN

We collected relevant clinical data for all infants enrolled in this study, including but not limited to their gestational age at birth, birth weight, number of days of PN before developing PNAC, antibiotic use, and calculated metrics such as birth weight percentile adjusted for gestational age (TableS1). Among the clinical metrics measured, only birth weight and the number of days of PN before occurrence of PNAC were statistically different between the disease and control groups (P < 0.05) (Figure1A). We found that infants with a birth weight percentile of >40% and a birth weight of >1.1 kg were less likely to develop PNAC. These clinical criteria are consistent with known clinical risk factors (Figure1B). Since length of PN administration is associated with the development of PNAC, we investigated the relationship further in our cohort to better understand how PNAC is impacted by duration of PN. We observed a positive correlation between PNAC diagnosis and the amount of time an infant receives PN before diagnosis (P < 0.05). Two‐thirds of the infants diagnosed with PNAC received PN for longer than 20 days. However, the length of time receiving PN is limited as an early clinical predictor of the risk of development of PNAC, as it is not concretely known until later in the clinical course, although certain clinical scenarios predict a longer length of PN.

Bacterial taxa correlate with PNAC but lack predictive potential in this cohort

Within this cohort of NICU infants, we identified several microbial taxa statistically different between the disease and control groups (P < 0.05). Among these taxa, the Enterobacteriaceae were present at greater abundance in stool samples collected from infants with serum conjugated bilirubin levels ≥ 1 mg/dl (FigureS1A). However, the three known genera within the Enterobacteriaceae family all demonstrate an opposite trend, with high abundance in the control group (FigureS1B). The unknown genera within the Enterobacteriaceae family showed a higher abundance in samples from infants with PNAC (FigureS1C). Although this result is consistent with our previous work,14 more research is needed to identify the primary Enterobacteriaceae genera present at elevated abundance in infants with PNAC. Finally, at the species level, Veillonella dispar was at an elevated relative abundance in infants with PNAC (FigureS1D). It is important to note that none of these microbes showed strong associations for determining infants at risk of developing PNAC; all are simply correlated with elevated levels of conjugated bilirubin.

The stool metabolome contains valuable biomarkers for infants at risk of PNAC

Among the 19 of 60 infants in our cohort who developed PNAC (Figure2A), there were nine for whom we were able to collect stool samples from before a diagnosis of PNAC. The interval between fecal metabolomic measurements and the initial detection of PNAC varied based on timing of sample collection. The timing of stool sample measurements relative to the detection of PNAC in 19 patients is shown in Figure2A. In the nine children (indicated by the dotted blue circles in Figure2A) from whom we collected ≥1 fecal samples before PNAC was diagnosed, the interval between measured elevations in fecal sphingomyelins and the detection of PNAC ranged from 5 to 46 days, with a mean and median of 20.6 and 15 days, respectively. These early samples of stools from children who went on to develop PNAC were used as a filter to provide a glimpse into metabolites with the greatest association for the risk of developing PNAC prior to disease progression. We first identified >100 metabolites differentially abundant between the disease group and the control group. For each candidate metabolite identified, we then compared the scaled quantity of that metabolite in each of the nine case study samples to the median value in all of the control samples. For metabolites found to be more abundant in PNAC samples, we selected only those with scaled intensities in >90% of the nine case study samples above the median value of the PNAC distribution (Figure2B). For metabolites negatively correlated with disease, we selected PNAC biomarkers by identifying those with scaled intensities in >90% of the nine case study samples falling below the median value of the disease distribution. We performed Mann‐Whitney U tests to assess the statistical significance of each metabolite between the samples with conjugated bilirubin below the clinical threshold and above. We used a Bonferroni multiple comparison correction to control for the number of metabolites tested. This analysis resulted in the identification of a subset of metabolites with the greatest potential to be early predictors for the risk of developing PNAC (Table1).

After multiple testing corrections, we identified a total of 45 biomarkers identified in early stool samples associated with the risk of developing PNAC. Specifically, we found 25 biomarkers positively correlated with a subsequent rise in conjugated bilirubin levels (Table1) and 20 metabolites negatively correlated with a subsequent increase in conjugated bilirubin levels (FigureS2). The biomarkers elevated in PNAC samples provide a glimpse into the pathophysiology that may be governing the disease. Although these biomarkers have early predictive value based on the case study infants, they require additional validation in future studies via targeted metabolomics.

Individual biomarkers are quite useful; however, a comparison of metabolites based on general classifications provides a measure of uniqueness for each biomarker. For example, we identified 1000 known metabolites across all samples and a total of 19 sphingomyelin metabolites. We determined 18 of 19 known sphingomyelin metabolites were biomarkers for elevated serum conjugated bilirubin levels. Separately, we classified PNAC biomarkers into seven additional groups (Table2). Several other classifications of membrane lipids present at greater abundance in PNAC samples indicate a general dysregulation of lipid metabolism in the liver or gastrointestinal (GI) tract. Several primary and secondary bile acids were reduced in PNAC samples, validating well‐known pathophysiology. Long‐chain carnitines were also reduced in PNAC samples.

The sphingomyelin biomarkers presented in Tables1 and2 and FigureS2 are directly applicable to the clinical setting. They have the potential to predict the development of PNAC before clinical markers of cholestasis rise, that is, diagnosing PNAC after detecting elevations in serum conjugated bilirubin levels.

Birth weight and birth weight percentile (adjusted for gestational age) were the most important predictors of infants at risk of developing PNAC (Figure3A). When including metabolic biomarkers, we were able to achieve >70% overall predictive accuracy, with five fold cross‐validation when classifying samples as control or PNAC (Figure3B). Five of the top six biomarkers determined to be the most predictive were sphingomyelins, as were 16 of the top 20 predictive biomarkers. Birth weight percentile and birth weight, again, demonstrate their predictive utility while accounting for the metabolic biomarkers. Other important biomarkers in our random forest models include several long‐chain carnitines and bile acids.

A small set of biomarkers and clinical variables provide early indication of infants at risk of PNAC

With predictive potential demonstrated using fivefold cross‐validation with a set of random forest machine learning models, we next calculated the classification accuracy within our cohort using simple criteria that would be possible to implement in the clinic. The ideal implementation of these biomarkers in the clinic would be a simple point‐of‐care diagnostic that provides neonatologists with additional information regarding an infant's gut health. This ideal scenario requires basic thresholds to be applied to a small set of metabolites present in the stool in the first weeks of life recorded at birth, which provide valuable PNAC risk stratification. In our cohort, infants with birth weights >40^th percentile for gestational age and >1.1 kg were at a significantly lower risk for developing PNAC. We assigned the other 41 infants who did not meet thesec cutoffs in a high‐risk group to determine which fecal biomarkers provided the most robust discriminatory accuracy for the development of PNAC (Figure4). We found the 12 best biomarkers, selected from a complete set of 78, were all sphingomyelin metabolites. Each of these metabolites provided a classification accuracy of the infants in our study between 82% and 88%, superior to the 66% classification accuracy provided by anthropometrics alone.

Although the 12 metabolites individually discriminated between the disease and control groups in our cohort, it was important to not overfit our model and apply a more robust strategy focused on future implementation in the NICU. To improve the robustness of our calculations, we assessed the accuracy of classification when using all 12 metabolites simultaneously by using an ensemble approach. For this, we analyzed the first available stool samples from infants meeting the following criteria: (1) high risk by clinical metrics and (2) stool samples collected within the first 3 weeks of life. For the 34 samples meeting these criteria, a majority vote ensemble classifier for our cohort provides an overall accuracy of 85%, which demonstrates consistency in the alignment across the 12 metabolites identified as optimal candidates for a rigorous follow‐up validation study (Figure5).

This article includes online‐only Supplemental Data.

Supplementary Figure S1.

Supplementary Figure S2.

Supplementary Figure S3.

Supplementary Table S1.

• 1.Khalaf RT, Sokol RJ. New insights into intestinal failure‐associated liver disease in children. Hepatology. 2020;71(4):1486‐1498. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 2.Venigalla S, Gourley GR. Neonatal cholestasis. Semin Perinatol. 2004;28(5):348‐355. [DOI] [PubMed] [Google Scholar]
• 3.Hsieh M‐H, Pai W, Tseng H‐I, Yang S‐N, Lu C‐C, Chen H‐L. Parenteral nutrition‐associated cholestasis in premature babies: risk factors and predictors. Pediatr Neonatol. 2009;50(5):202‐207. [DOI] [PubMed] [Google Scholar]
• 4.Alkharfy TM, Ba‐Abbad R, Hadi A, Sobaih BH, AlFaleh KM. Total parenteral nutrition‐associated cholestasis and risk factors in preterm infants. Saudi J Gastroenterol. 2014;20(5):293‐296. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 5.Christensen RD, Henry E, Wiedmeier SE, Burnett J, Lambert DK. Identifying patients, on the first day of life, at high‐risk of developing parenteral nutrition‐associated liver disease. J Perinatol. 2007;27(5):284‐290. [DOI] [PubMed] [Google Scholar]
• 6.Mangalat N, Bell C, Graves A, Imseis EM. Natural history of conjugated bilirubin trajectory in neonates following parenteral nutrition cessation. BMC Pediatr. 2014;14(1):298. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 7.Park HW, Lee NM, Kim JH, Kim KS, Kim S‐N. Parenteral fish oil‐containing lipid emulsions may reverse parenteral nutrition‐associated cholestasis in neonates: a systematic review and meta‐analysis. J Nutr. 2015;145(2):277‐283. [DOI] [PubMed] [Google Scholar]
• 8.Kasirer Y, Bin‐Nun A, Raveh A, Schorrs I, Mimouni FB, Hammerman C. SMOFlipid protects preterm neonates against perinatal nutrition‐associated cholestasis. Am J Perinatol. 2019;36(13):1382‐1386. [DOI] [PubMed] [Google Scholar]
• 9.Casson C, Nguyen V, Nayak P, et al. A comparison of Smoflipid® and Intralipid® in the early management of infants with intestinal failure. J Pediatr Surg. 2020;55(1):153‐157. [DOI] [PubMed] [Google Scholar]
• 10.Zhgun ES, Ilina EN. Fecal metabolites as non‐invasive biomarkers of gut diseases. Acta Naturae. 2020;12(2):4‐14. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 11.Bouatra S, Aziat F, Mandal R, et al. The human urine metabolome. PLoS One. 2013;8(9):e73076. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 12.Cahova M, Bratova M, Wohl P. Parenteral nutrition‐associated liver disease: the role of the gut microbiota. Nutrients. 2017;9(9):987. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 13.Lee WS, Sokol RJ. Intestinal microbiota, lipids, and the pathogenesis of intestinal failure‐associated liver disease. J Pediatr. 2015;167(3):519‐526. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 14.Hourigan SK, Moutinho TJ Jr, Berenz A, et al. Gram‐negative microbiota blooms in premature twins discordant for parenteral nutrition associated cholestasis. J Pediatr Gastroenterol Nutr. 2020;70(5):640‐644. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 15.Li Y, Tang R, Leung PSC, Gershwin ME, Ma X. Bile acids and intestinal microbiota in autoimmune cholestatic liver diseases. Autoimmun Rev. 2017;16(9):885‐896. [DOI] [PubMed] [Google Scholar]
• 16.Fawaz R, Baumann U, Ekong U, et al. Guideline for the evaluation of cholestatic jaundice in infants: joint recommendations of the North American society for pediatric gastroenterology, hepatology, and nutrition and the European society for pediatric gastroenterology, hepatology, and nutrition. J Pediatr Gastroenterol Nutr. 2017;64(1):154‐168. [DOI] [PubMed] [Google Scholar]
• 17.Almonacid DE, Kraal L, Ossandon FJ, et al. 16S rRNA gene sequencing and healthy reference ranges for 28 clinically relevant microbial taxa from the human gut microbiome. PLoS One. 2017;12(5):e0176555. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 18.Ramamoorthy S, Levy S, Mohamed M, et al. An ambient‐temperature storage and stabilization device performs comparably to flash‐frozen collection for stool metabolomics in infants. BMC Microbiol. 2021;21(1):59. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 19.Loftfield E, Vogtmann E, Sampson JN, et al. Comparison of collection methods for fecal samples for discovery metabolomics in epidemiologic studies. Cancer Epidemiol Biomarkers Prev. 2016;25(11):1483‐1490. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 20.Zierer J, Jackson MA, Kastenmüller G, et al. The fecal metabolome as a functional readout of the gut microbiome. Nat Genet. 2018;50(6):790‐795. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 21.Evans AM, Bridgewater B, Liu Q, et al. High resolution mass spectrometry improves data quantity and quality as compared to unit mass resolution mass spectrometry in high‐throughput profiling metabolomics. Metabolomics. 2014;4(2):1. [Google Scholar]
• 22.Ford L, Kennedy AD, Goodman KD, et al. Precision of a clinical metabolomics profiling platform for use in the identification of inborn errors of metabolism. J Appl Lab Med. 2020;5(2):342‐356. [DOI] [PubMed] [Google Scholar]
• 23.Evans AM, DeHaven CD, Barrett T, Mitchell M, Milgram E. Integrated, nontargeted ultrahigh performance liquid chromatography/electrospray ionization tandem mass spectrometry platform for the identification and relative quantification of the small‐molecule complement of biological systems. Anal Chem. 2009;81(16):6656‐6667. [DOI] [PubMed] [Google Scholar]
• 24.DeHaven CD, Evans AM, Dai H, Lawton KA. Organization of GC/MS and LC/MS metabolomics data into chemical libraries. J Cheminform. 2010;2(1):1‐12. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 25.Callahan BJ, McMurdie PJ, Rosen MJ, Han AW, Johnson AJA, Holmes SP. DADA2: high‐resolution sample inference from Illumina amplicon data. Nat Meth. 2016;13(7):581‐583. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 26.Callahan BJ, McMurdie PJ, Holmes SP. Exact sequence variants should replace operational taxonomic units in marker‐gene data analysis. ISME J. 2017;11(12):2639‐2643. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 27.Bolyen E, Rideout JR, Dillon MR, et al. Reproducible, interactive, scalable and extensible microbiome data science using QIIME 2. Nat Biotechnol. 2019;37(8):852‐857. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 28.Bock J, Liebisch G, Schweimer J, Schmitz G, Rogler G. Exogenous sphingomyelinase causes impaired intestinal epithelial barrier function. World J Gastroenterol. 2007;13(39):5217‐5225. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 29.Qi Y, Jiang C, Cheng J, et al. Bile acid signaling in lipid metabolism: metabolomic and lipidomic analysis of lipid and bile acid markers linked to anti‐obesity and anti‐diabetes in mice. Biochim Biophys Acta. 2015;1851(1):19‐29. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 30.Kurek K, Łukaszuk B, Piotrowska DM, Wiesiołek P, Chabowska AM, Żendzian‐Piotrowska M. Metabolism, physiological role, and clinical implications of sphingolipids in gastrointestinal tract. BioMed Res Int. 2013;2013(20):e908907. [DOI] [PMC free article] [PubMed] [Google Scholar]
• 31.Soo I, Madsen KL, Tejpar Q, et al. VSL#3 probiotic upregulates intestinal mucosal alkaline sphingomyelinase and reduces inflammation. Can J Gastroenterol. 2008;22(3):237‐242. [DOI] [PMC free article] [PubMed] [Google Scholar]

Supplementary Materials

Data Availability Statement

Data described in the article, code book, and analytic code will be made available on request pending application and approval.