The Assessment of Fecal Volatile Organic Compounds in Healthy Infants: Electronic Nose Device Predicts Patient Demographics and Microbial Enterotype

Abstract

Background: The assessment of fecal volatile organic compounds (VOCs) has emerged as a noninvasive biomarker in many different pathologies. Before assessing whether VOCs can be used to diagnose intestinal diseases, including necrotizing enterocolitis (NEC), it is necessary to measure the impact of variable infant demographic factors on VOC signals.

Materials and methods: Stool samples were collected from term infants at four hospitals in a large metropolitan area. Samples were heated, and fecal VOCs assessed by the Cyranose 320 Electronic Nose. Twenty-eight sensors were combined into an overall smellprint and were also assessed individually. 16s rRNA gene sequencing was used to categorize infant microbiomes. Smellprints were correlated to feeding type (formula versus breastmilk), sex, hospital of birth, and microbial enterotype. Overall smellprints were assessed by PERMANOVA with Euclidean distances, and individual sensors from each smellprint were assessed by Mann-Whitney U-tests. P < 0.05 was significant.

Results: Overall smellprints were significantly different according to diet. Individual sensors were significantly different according to sex and hospital of birth, but overall smellprints were not significantly different. Using a decision tree model, two individual sensors could reliably predict microbial enterotype.

Conclusions: Assessment of fecal VOCs with an electronic nose is impacted by several demographic characteristics of infants and can be used to predict microbiome composition. Further studies are needed to design appropriate algorithms that are able to predict NEC based on fecal VOC profiles.

Keywords: Biomarkers; Electronic nose; Enterotype; Fecal volatile organic compounds; Necrotizing enterocolitis.

Figure 1.

Cyranose® 320 Electronic Nose device was used to analyze fecal volatile organic compounds from infant stool.

Figure 2.

Differences in fecal VOCs between diet, calories, sex, and hospital location. (A) Overall smellprints were significantly different between infants consuming exclusive human breastmilk (HBM) and infants consuming exclusive formula. (B) Sensor 5 and 25 were the most different. (C) Overall smellprints were not significantly different according to caloric intake. (D) Sensor 16 had the strongest negative correlation, and sensor 12 had the strongest negative correlation between calories. (E) Overall smellprints were not significantly different between male and female infants. (F) Two sensors were significantly different. (G) Overall smellprints were not significantly different between four hospital locations. (H) Sensors 11 and 23 were the most different based on hospital of birth.

Figure 3.

16s rRNA analysis of stool samples. (A) Relative abundance of bacterial families grouped by enterotype. (B) NMDS Ordinated based on Bray-Curtis dissimilarity of family abundances. The microbiota composition was significantly different between enterotypes (p=0.00002).

Figure 4.

Three individual sensors were significantly associated with enterotype

Figure 5.

Decision tree model created by the Fast Frugal Tree algorithm was used to test whether smellprint sensors could be used to predict the two most common enterotypes. The top panel shows the training set population, which consisted of 12 samples with microbiomes belonging to Enterotype 1 (squares) and 14 from Enterotype 2 (triangles). The middle panel shows the structure of the decision tree, which splits samples based on the readings from two smell print sensors. Green shapes indicate correct classifications, while red shapes indicate erroneous classifications. The bottom panel summarizes the performance of the model using a variety of metrics, including the error matrix, sensitivity, specificity, accuracy, and a receiver operator characteristic (ROC) curve.