Acute otitis media (AOM) is among the commonest childhood infections; its incidence is highest in children 1 to 4 years of age with 61 new AOM episodes per 100 children per year.1
Classically, the 3 major bacteria involved in AOM are Streptococcus pneumoniae, Haemophilus influenzae and Moraxella catarrhalis,2 although there is ongoing debate about the role of the latter.3,4 It is assumed that these bacteria enter the middle ear cavity from the nasopharyngeal (NP) niche by ascending through the Eustachian tube upon a virus-induced inflammatory cascade.5 Whereas microbial analysis of middle ear fluid (MEF) is regarded as the gold standard to determine AOM etiology,6 this requires an invasive procedure such as tympanocentesis or myringotomy to obtain a sample; therefore, NP samples are often used as proxy. A recent systematic review of this approach, however, showed only a moderate concordance between conventional cultures of NP and MEF samples.7 This may reflect limitations of conventional culture techniques, which are less sensitive than molecular methods and do not consider relative abundance of pathogens in the context of the complete microbial ecosystem nor the role of commensals in the pathophysiology.8
Episodes of acute ear discharge in children with tympanostomy tubes are thought to be the result of AOM, in which MEF drains through the tube.5 The bacteria involved in AOM in children with tympanostomy tubes (AOMT) include the major bacteria found in AOM, as well as Staphylococcus aureus and Pseudomonas aeruginosa.9 Because tympanostomy tube otorrhea (TTO) can be easily obtained in children with AOMT, this population is of particular interest when studying the pathogenesis of AOM(T). In addition, characterization of the complete microbial community composition through next-generation sequencing techniques holds a great promise to better understand the relation between NP and TTO microbiota in children with AOM(T). The 3 studies thus far comparing microbiota compositions of the NP and MEF/TTO in children with otitis media focused either on differences between the NP and MEF/TTO samples or on otitis media with effusion rather than AOM(T) and/or were too small to extensively study the relation between these 2 niches on the individual patient level.8,10,11
Our group recently performed a randomized controlled trial on the treatment of AOMT.12 As part of this trial, we collected NP and TTO samples of all participants. In the present study, we aim to assess the relevance of the respiratory ecosystem in childhood AOMT by analyzing the relationship between the NP and TTO microbiota in baseline samples of 94 participants. Moreover, we explore whether microbial community profiling predict natural disease course of AOMT.
MATERIALS AND METHODS
We obtained baseline NP and TTO samples from children under 5 years of age who participated in our recent randomized controlled trial of treatment of AOMT. Children were randomly allocated to either antibiotic-corticosteroid (hydrocortisone–bacitracin–colistin) eardrops, oral antibiotics (amoxicillin–clavulanate suspension) or initial observation (no treatment).9,12 Deep transnasal NP swabs were obtained according to World Health Organization standard procedures,13 whereas TTO samples were retrieved by swabbing discharge in the ear canal, avoiding skin contact. During follow-up, otoscopy was performed at 2 weeks to assess presence or absence of otorrhea and the parents of participating children kept a daily diary of ear-related symptoms for 6 months. Further details of the trial entry criteria and methodology are described elsewhere.12
Bacterial High-Throughput Sequencing and Bioinformatic Processing
Bacterial DNA of the matching TTO and NP sample pairs was isolated, polymerase chain reaction (PCR) amplicon libraries were generated, 16S ribosomal RNA gene-sequencing was executed and amplicon pools were processed in our bioinformatics pipeline as previously described and detailed in the supplements.14 All samples fulfilled our quality control standards for reliable analyses, having DNA levels of >0.3 pg/µL over negative controls. The 4 highest PCR and DNA isolation blanks were also sequenced and yielded only a median number of 113.5 reads (range, 8–667 reads/blank), whereas all samples yielded more than 10,000 sequences. Finally, none of the reagent contaminants published by Salter et al15 were present in more than half of our negative controls, all indicating that our strict sequencing protocol and bioinformatics pipeline resulted in no apparent contamination. Turicella was not present in any of the negative controls. In addition, culture results of Streptococcus pneumoniae, H. influenzae, M. catarrhalis, S. aureus and P. aeruginosa were used for the posthoc species-level annotations of the corresponding operational taxonomic unit (OTU) (eFigure 1, Supplemental Digital Content, http://links.lww.com/INF/D311). We generated an abundance-filtered dataset by including only those OTUs that were present at or above a confidence level of detection (0.1% relative abundance) in at least two samples, retaining 138 OTUs in total.16 To avoid OTUs with identical annotations, we refer to OTUs using their taxonomical annotations combined with a rank number based on the abundance of each given OTU. The raw OTU counts table was used for calculations of α-diversity and analyses using the metagenomeSeq package.17 The OTU-proportions table was used for all other downstream analyses, including hierarchical clustering and random forest modelling. β-Diversity was assessed using the Bray-Curtis similarity metric (calculated by 1 – Bray-Curtis dissimilarity).
All analyses were performed in R version 3.3.2. Good’s estimator of coverage was calculated using the formula: [1 − (singletons/total number of sequences)] × 100.18 α-Diversity was estimated by the Chao 1 estimate of richness and the Shannon’s diversity index, which takes into account both richness and evenness of the samples. Statistical significance of the differences in α-diversity was calculated using linear mixed models with the participant as random factor. Nonmetric multidimensional scaling (NMDS) plots were used to visualize differences of total microbiota communities between groups, and statistical significance was calculated by adonis and Multi-Response Permutation Procedures (both 9,999 permutations) with samples from the same participant grouped in the analysis (as random factor). The overall qualitative concordance between NP and TTO microbiota was evaluated according to previously described methods.7 In short, we calculated the prevalence in both niches, the positive predictive value, negative predictive value, sensitivity and specificity using the TTO sample as the reference. The quantitative correlations were calculated with Spearman’s rank correlation coefficient. Average linkage hierarchical clustering including the determination of biomarker species was performed as described previously.19 We used metagenomeSeq to identify the microbial taxa associated between groups (ie, NP vs. TTO).17
To confirm with an unsupervised quantitative method whether the abundances of NP biomarker species were related to their respective abundances of the paired TTO samples, we used a random forest approach. This also allowed us to determine the relation of biomarker species in the NP with other species in the paired TTO samples. We performed 100-times repeated, 10-times cross-validated sparse random forest models generating 10,000 trees (train function, randomForest package) for each of the biomarker species. Variables for this sparse model were selected using the bacterial species determined by the interpretation step of a 20-times cross-validated VSURF procedure, generating 10,000 trees each iteration, with 100 iterations for the thresholding step and 50 iterations for the interpretation step.20 The direction of the associations was estimated post-hoc using the partial Spearman’s correlations. The importance of each bacterial species is determined by evaluating the increase in the mean square error (MSE; ie, the decrease in prediction accuracy) between observations and model when the data for that bacterial species are randomly permuted. The increase in MSE averaged over all trees produces the final measure of importance.21
To assess whether respiratory microbiota composition predicts AOMT natural disease course, we studied the association between NP and TTO microbiota of the 27 children who were not treated (initial observation group). We used the trial’s prespecified clinical outcome measures, that is, otoscopically confirmed otorrhea 2 weeks after randomization (binary outcome), the duration of the initial otorrhea episode, total number of days with otorrhea and number of recurrent otorrhea episodes during 6 months of follow-up (numerical outcomes). To this purpose, we built separate cross-validated sparse random forest classification and prediction models as described above for the clinical outcomes, respectively. The performance of the classification models was evaluated by calculating the area under the receiver operating characteristic curve (AUC) using the out-of-bag predictions for classification (pROC package22). The performance of the prediction models was assessed by calculating the Spearman’s rank correlations between the model predicted and the observed outcome values.
A P value of less than 0.05 for single parameter outcome or Benjamini-Hochberg adjusted q value less than 0.05 when multiple variables were tested was considered statistically significant.
In 98 of 107 (92%) children under 5 years of age from whom paired NP and TTO samples were available, a sufficient amount of DNA was isolated for reliable 16S ribosomal RNA–based sequencing analyses.14 MiSeq PCR followed by MiSeq sequencing was successful in 94 of 98 children (96%). Fifteen of these children had bilateral AOMT, resulting in 94 NP samples and 109 paired TTO samples (eFigure 2, Supplemental Digital Content, http://links.lww.com/INF/D311). Characteristics of the study population are shown in eTable 1 (Supplemental Digital Content, http://links.lww.com/INF/D311).
Characterization of Sequencing Results and Diversity
A total of 8,758,772 reads were used for analysis (mean 43,147 ± 16,199 reads per sample). These were binned into 138 97%-identity OTUs, representing 66 taxonomic genera from 8 phyla. Good’s coverage of >99.9% was reached for all samples, and rarefaction curves on raw count data approached plateau in all samples (eFigure 3, Supplemental Digital Content, http://links.lww.com/INF/D311), suggesting that the sequence results of each sample represented the majority of bacteria present in the NP and TTO samples under study.
The estimated number of species and Shannon diversity was higher in NP samples than in TTO samples (Chao mean 37.8 and 25.6 species for NP and TTO, respectively; Shannon mean 0.97 and 0.73 for NP and TTO, respectively; both P < 0.001) (eFigure 4, Supplemental Digital Content, http://links.lww.com/INF/D311).
The total microbiota composition differed significantly between NP and TTO (adonis, R2 = 0.054, P < 0.001; Multi-Response Permutation Procedures, A = 0.031, P < 0.001) (Fig. 1A). However, paired NP and TTO samples were considerably more similar than unpaired samples underlining the same biological source (median Bray-Curtis similarity 0.26 and 0.04, respectively, P < 0.001) (Fig. 1B). The similarity of paired NP and TTO samples did vary slightly with age (median Bray-Curtis similarity; <2 years, 0.27; >2 years, 0.11; P = 0.093), but not with number of previous tympanostomy tubes (1 tube, 0.25; >1 tube, 0.15; P = 0.446), duration of tube presence (0–5 days, 0.20; >5 days, 0.16; P = 0.849), history of prior adenoidectomy (yes, 0.14; no 0.26; P = 0.595), nor with season of sampling (P = 0.899) (eFigure 5, Supplemental Digital Content, http://links.lww.com/INF/D311). TTO samples from both ears of the same child (n = 15 with bilateral AOMT) were substantially more similar than TTO samples of different children (Bray-Curtis similarity 0.50 and 0.02, respectively, P < 0.001).
Microbiota Profiles and Biomarker Species
Hierarchical clustering showed the presence of 10 distinct microbiota profiles, which were mainly driven by the abundance of 12 biomarker species (Fig. 2A). Most biomarker species were differentially abundant in NP and TTO samples, except for Streptococcus (7), Klebsiella and Haemophilus (91), which showed high concordance for presence as well as abundance between niches. In contrast, Moraxella spp., Streptococcus pneumoniae (6), H. influenzae (1), Corynebacterium and Dolosigranulum were stronger associated with the NP, whereas Turicella, P. aeruginosa (5) and S. aureus (2) abundances were more associated with TTO (metagenomeSeq absolute log2 fold change, all >2; q < 0.01) (Fig. 2B). A posteriori plotting of all biomarker species in the NMDS ordination supported the niche-preferential abundance as described above (Fig. 1A).
On the individual level, 30% of the paired NP and TTO samples, however, shared the exact same microbiota profile (Fig. 2C–D). This one-to-one association was most obvious for the Haemophilus-, S. aureus- (2), Streptococcus- (7) and Klebsiella-dominated profiles. The Streptococcus (7) NP-profile was additionally associated with the same profile in TTO, also associated with a Streptococcus pneumoniae-dominated TTO profile. The M. catarrhalis NP profile was rarely found in TTO. However, a strong association was observed between Moraxella-dominated NP and P. aeruginosa-dominated TTO (Fig. 2C–D).
Agreement in Microbiota Composition
In contrast to the relatively low correlation between paired NP and TTO samples on total microbiota profile level (Fig. 2), the concordance on the single bacterial species level (OTU level) was considerably higher with a substantial agreement of 79% for the presence/absence of individual species (95% confidence interval: 78%–80%) (eTable 2, Supplemental Digital Content, http://links.lww.com/INF/D311). The high negative predictive value underlines that the NP might be the common biological source of TTO bacteria (91%; 95% confidence interval: 91%–92%).
The quantitative correlation between the bacterial abundances of individual species in the paired NP and TTO samples was in line with the qualitative results, with 12 of the 15 most abundant bacterial species showing a significant correlation between NP and TTO (P < 0.05; Spearman’s rho range, 0.193–0.548) (Fig. 3); H. influenzae (1) (Spearman’s rho 0.548, P < 0.001), P. aeruginosa (5) (Spearman’s rho 0.489, P < 0.001) and S. aureus (2) abundance (Spearman’s rho 0.439, P < 0.001) showed the strongest correlations, whereas Moraxella spp. (including M. catarrhalis ) and Streptococcus spp. including Streptococcus pneumoniae , Spearman’s rho 0.180, P = 0.061) abundances were clearly not correlated between NP and TTO. When analyzing also the lower abundant bacterial species, only 46 of the 138 species showed a significant correlation (P < 0.05; median Spearman’s rho 0.337; interquartile range, 0.247–0.436; combined relative abundance of 81.5%), suggesting low abundant species are less likely seeded from NP to middle ear.
Random Forest Associations
All results together confirmed our hypothesis that the NP microbiota composition does not fully reflect TTO microbiota in a simple one-to-one fashion. Despite this, we found that microbial profiles of NP samples still predicted the microbial community in the paired TTO samples fairly well, with an almost one-to-one association when dominated by H. influenzae (1) and Haemophilus (91), Klebsiella, Corynebacterium and Streptococcus (7) (Fig. 4). Moreover, S. aureus (2) abundance in the NP was predictive for either S. aureus (2) or Neisseria overgrowth in TTO as well as absence of other species. Similarly, P. aeruginosa (5) abundance in the NP swab was predictive for either Pseudomonas or Staphylococcus abundance in TTO. Dolosigranulum abundance in NP demonstrated a less specific association with TTO bacterial abundances. M. catarrhalis (3) was highly predictive of other species but itself, especially Pseudomonas. Streptococcus pneumoniae (6) abundance in the NP was mostly associated with presence of a diverse group of (oral) anaerobes, though not itself.
Relation Between Microbiota and Clinical Outcome
Although the baseline respiratory microbiota community profiles of the children allocated to the initial observation group could not predict the otoscopically confirmed presence or absence of otorrhea 2 weeks after onset of symptoms very accurately (AUC 0.71 and 0.62 for the sparse RF models using NP and TTO microbiota, respectively), the microbiota composition of NP samples could predict the duration of symptoms and recurrence of otorrhea as reported by the parents fairly well (Pearson’s r between predicted and observed outcome 0.40–0.54, all P< 0.05, random forest R2 = 0.69–0.70) (Fig. 5A), whereas the models using TTO microbiota did not demonstrate a significant correlation between predicted and observed outcome values (all P > 0.10). Within this untreated group, especially the NP abundance of Acinetobacter, followed by Klebsiella, Neisseria and H. influenzae (1) (positive partial Spearman’s correlation) were associated with longer duration of otorrhea, whereas abundance of Corynebacterium, followed by Dolosigranulum and Haemophilus (91) were associated with shorter duration of otorrhea (negative partial Spearman’s correlation) (Fig. 5B).
This study, comparing paired NP and TTO samples of 94 children with AOMT, shows a substantive qualitative and moderate quantitative correlation between NP and TTO thereby supporting the hypothesis that the microbiota in the middle ear originates from the NP. Moreover, NP microbiota composition predicts presence and absence of other microbiota in the TTO well, with for example S. aureus abundance in the NP predicting either the presence of S. aureus or Pseudomonas in the middle ear. Second, our study indicates that the TTO microbiota of children with AOMT is a rich community comprising of on average 26 species, suggesting the existence of a complex middle ear microbiome in those children rather than the presence of a single pathogen.
In accordance with previous small studies, NP samples show a higher α-diversity compared with TTO, and the total microbiota composition differed significantly between both niches.10,23 Although high qualitative concordance was found, our analyses also showed that some biomarker species are overrepresented in NP samples, whereas other biomarker species are more abundant in TTO samples, suggesting niche preference. Especially the association of bacteria like M. catarrhalis, other Moraxella spp., Streptococcus pneumoniae and Corynebacterium/Dolosigranulum with NP rather than TTO presence, confirms previous findings that these microbes are key commensals of the NP niche ecosystem.24–27 Moreover, the association of Turicella, P. aeruginosa and S. aureus within TTO samples corroborates reports that describe these species as otopathogens.10,27–29
The difference in niche preference between bacteria is presumably driven by the niche specific growth condition of both sites such as oxygen tension, temperature, humidity, presence of nutrients or immune cells.26 Moreover, seeding of microorganisms from the NP to the middle ear through the Eustachian tube and local outgrowth might not solely depend on the presence and/or abundance of these microorganisms in the ascending community, but also on their relative abundance, as well as on the presence of other microbial community members that may either support or prevent their dissemination. By analyzing the association of the microbial profiles on the individual level as well as using quantitative correlations associating NP and TTO microbiota, a significant one-to-one relationship between the NP and TTO abundances was found for the majority of the microbiota (81.5%). The strength of the correlations was generally modest but was highest for potential (AOMT) pathogens such as P. aeruginosa, S. aureus, Streptococcus, Turicella and Haemophilus spp. This was confirmed using unsupervised random forest analysis. Moreover, random forest analysis also demonstrated that the abundances of P. aeruginosa, S. aureus and Turicella in NP were additionally associated with a range of Gram-negative oral type of bacteria in the TTO, including Neisseria, Bradyrhizobium, Bergeyella and Actinomyces spp. A possible explanation for this symbiotic behavior might be the ability of P. aeruginosa30 to rapidly reduce these species’ toxic oxygen levels, and vice versa the known facilitation of P. aeruginosa growth by the metabolites of these oral bacteria.31 Interestingly, Streptococcus pneumoniae abundance in the NP predicted mostly the presence of a diverse group of anaerobes in TTO, whereas in the few occasions Streptococcus pneumoniae occurred in TTO this was mostly predicted by the abundance of Streptococcus (7). This suggests that collaboration between both species (which is a well-known phenomenon for streptococcal species26) is needed for the currently circulating serotypes to colonize the middle ear niche and render pathogenic behavior.
Over the last years, with the advent of microbial community profiling, evidence is accumulating that M. catarrhalis is associated with a stable bacterial community composition and a state of respiratory health.26 In our study, only three TTO samples had a M. catarrhalis dominated profile, suggesting a limited role of this species in AOM(T) pathogenesis. While our study population consisted of children with previous otitis media episodes requiring tympanostomy tubes, NP abundance of M. catarrhalis showed no association with its presence in TTO samples across all our analyses. These data are corroborated by a recent study from Australia,27 therefore suggest that this bacterium is rather a NP commensal than a pathogen. Other studies however have also reported that M. catarrhalis could play a role in AOM in children,32,33 although in those cases, it generally reflects a mild infection.3 In addition, the historical common detection of M. catarrhalis in conventional culture-based studies might mirror the easy identification of this species by culture, rather then a high abundance in MEF. In all, this warrants a careful consideration of future vaccination strategies against this microorganism.34
Some limitations deserve further attention. First, we did not include children with a body temperature higher than 38.5°C, who might have different microbiota profiles. Second, TTO was sampled from the ear canal after a median otorrhea duration of 2 days (interquartile range, 1–4); the contamination by external ear canal microbiota might have led to an overestimation of P. aeruginosa, Turicella and/or S. aureus detection as these species are common constituent of the microbiota in the ear canal.11,35 However, we have previously compared bacterial presence in otorrhea samples swabbed from the ear canal with those taken from the lumen of the tympanostomy tube in a subset of 20 children participating in the trial and did observe a high concordance, suggesting limited outer ear canal contamination.9 Also, the high correlation between the abundances of P. aeruginosa, Turicella and S. aureus in NP and TTO samples might further indicate that their abundances in TTO are not merely the result of contamination from the outer ear canal, but that these species rather originate from the NP niche. Our results are in line with other recent studies that detect Pseudomonas and Turicella in low abundance in the NP of the majority of children without tympanostomy tubes.27,36,37 We cannot exclude, however, that the presence of P. aeruginosa and T. otitidis in the NP may be the result of its reversed transition from the TTO to the NP.
Although NP microbiota did aptly predict the natural disease course of AOMT as defined by 3 different clinical outcome measures as reported by parents, we did not observe a significant relation between NP microbiota and otorrhea 2 weeks after randomization as confirmed by a physician through otoscopic examination. This ambiguity may well be caused by sample size constraints, as only 27 of the 94 children included in the current study were allocated to the initial observation group. Although our prediction algorithms may not be accurate enough for their direct implementation in the clinical setting, they might open-up new avenues to refrain from treatment in children with AOMT whom NP microbiota profiles indicate a favorable natural disease course and to initiate treatment in those with a less favorable predicted outcome. Further testing and validation in prospective cohorts is, however, warranted. With AOM being the single most important cause of childhood antibiotic prescribing, it does seem worthwhile to further study the potential usefulness of microbiota analysis to predict clinical outcome and its impact on antimicrobial use and subsequent development of antimicrobial resistance. Of particular note is that we could not predict the natural disease course using TTO microbiota, although TTO reflects the site of infection. This may suggest that the NP microbiota not only seed the middle ear with potential pathogens that initiate disease but also determines recovery to health. This is strengthened by the finding of a strong association between Dolosigranulum and Corynebacterium abundance and a better clinical outcome, supporting evidence that these bacteria are associated with respiratory health.26,27 We hypothesize that children colonized with these beneficial microbes have diminished mucosal inflammation, leading to more rapid restoration of Eustachian tube function and subsequently clinical recovery.
In conclusion, this study offers valuable insights into the association between NP and TTO microbiota compositions in children with AOMT. Our findings of substantial niche-niche relationships endorse the hypothesis that the middle ear microbiota is seeded by the NP microbiota through ascending the Eustachian tube. Moreover, our results suggest that M. catarrhalis could be a NP commensal rather than a pathogen, which is a relevant finding with regard to future vaccine strategies and that warrants further investigation. Finally, our data indicate that NP microbiota profiles may be useful for clinical decision-making in the future, but for this more research is needed.
The authors thank the children and their parents who participated in the study; Pauline Winkler, Nelly van Eden, Lidian Izeboud, Dicky Mooiweer and our team of medical students for administrative and practical support; the participating family physicians and the ear, nose and throat surgeons at the participating hospitals.
AVAILABILITY OF DATA AND MATERIALS
The 16S ribosomal RNA sequence reads were submitted to the National Center for Biotechnology and Information Sequence Read Archive (accession number SRP128433).
1. Monasta L, Ronfani L, Marchetti F, et al. Burden of disease caused by otitis media: systematic review and global estimates. PLoS One. 2012;7:e36226.
2. Coker TR, Chan LS, Newberry SJ, et al. Diagnosis, microbial epidemiology, and antibiotic treatment of acute otitis media
in children: a systematic review. JAMA. 2010;304:2161–2169.
3. Aebi C. Curtis N, Finn A, Pollard AJ. In: Moraxella Catarrhalis – Pathogen or Commensal? Vol 2011.697. New York, NY: Springer New York.
4. Broides A, Dagan R, Greenberg D, et al. Acute otitis media
caused by Moraxella catarrhalis: epidemiologic and clinical characteristics. Clin Infect Dis. 2009;49:1641–1647.
5. Schilder AG, Chonmaitree T, Cripps AW, et al. Otitis media. Nat Rev Dis Primers. 2016;2:16063.
6. Preciado D. Preciado D. In: Otitis Media: State of the Art Concepts and Treatment. 2015.Cham, Switzerland: Springer International Publishing.
7. van Dongen TM, van der Heijden GJ, van Zon A, et al. Evaluation of concordance between the microorganisms detected in the nasopharynx
and middle ear of children with otitis media. Pediatr Infect Dis J. 2013;32:549–552.
8. Liu CM, Cosetti MK, Aziz M, et al. The otologic microbiome: a study of the bacterial microbiota in a pediatric patient with chronic serous otitis media using 16SrRNA gene-based pyrosequencing. Arch Otolaryngol Head Neck Surg. 2011;137:664–668.
9. van Dongen TM, Venekamp RP, Wensing AM, et al. Acute otorrhea in children with tympanostomy tubes: prevalence of bacteria and viruses in the post-pneumococcal conjugate vaccine era. Pediatr Infect Dis J. 2015;34:355–360.
10. Jervis-Bardy J, Rogers GB, Morris PS, et al. The microbiome of otitis media with effusion in Indigenous Australian children. Int J Pediatr Otorhinolaryngol. 2015;79:1548–1555.
11. Chan CL, Wabnitz D, Bassiouni A, et al. Identification of the bacterial reservoirs for the middle ear using phylogenic analysis. JAMA Otolaryngol Neck Surg. 2017;143:155–161.
12. van Dongen TM, van der Heijden GJ, Venekamp RP, et al. A trial of treatment for acute otorrhea in children with tympanostomy tubes. N Engl J Med. 2014;370:723–733.
13. O’Brien KL, Nohynek H; World Health Organization Pneumococcal Vaccine Trials Carriage Working Group. Report from a WHO Working Group: standard method for detecting upper respiratory carriage of Streptococcus pneumoniae
. Pediatr Infect Dis J. 2003;22:e1–e11.
14. Bosch AATM, de Steenhuijsen Piters WAA, van Houten MA, et al. Maturation of the infant respiratory microbiota
, environmental drivers, and health consequences. A prospective cohort study. Am J Respir Crit Care Med. 2017;196:1582–1590.
15. Salter SJ, Cox MJ, Turek EM, et al. Reagent and laboratory contamination can critically impact sequence-based microbiome analyses. BMC Biol. 2014;12:87.
16. Subramanian S, Huq S, Yatsunenko T, et al. Persistent gut microbiota immaturity in malnourished Bangladeshi children. Nature. 2014;510:417–421.
17. Paulson JN, Stine OC, Bravo HC, et al. Differential abundance analysis for microbial marker-gene surveys. Nat Methods. 2013;10:1200–1202.
18. Good IJ. The population frequencies of species and the estimation of population parameters. Biometrika. 1953;40:237–264.
19. de Steenhuijsen Piters WA, Heinonen S, Hasrat R, et al. Nasopharyngeal microbiota, host transcriptome, and disease severity in children with respiratory syncytial virus infection. Am J Respir Crit Care Med. 2016;194:1104–1115.
20. Genuer R, Poggi J-M, Tuleau-Malot C. Variable selection using random forests. Pattern Recognit Lett. 2010;31:2225–2236.
21. Breiman L. Random forests. Mach Learn. 2001;45:5–32.
22. Robin X, Turck N, Hainard A, et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinformatics. 2011;12:77.
23. Chan CL, Wabnitz D, Bardy JJ, et al. The microbiome of otitis media with effusion. Laryngoscope. 2016;126:2844–2851.
24. Biesbroek G, Tsivtsivadze E, Sanders EA, et al. Early respiratory microbiota
composition determines bacterial succession patterns and respiratory health in children. Am J Respir Crit Care Med. 2014;190:1283–1292.
25. Pettigrew MM, Laufer AS, Gent JF, et al. Upper respiratory tract microbial communities, acute otitis media
pathogens, and antibiotic use in healthy and sick children. Appl Environ Microbiol. 2012;78:6262–6270.
26. Man WH, de Steenhuijsen Piters WA, Bogaert D. The microbiota of the respiratory tract: gatekeeper to respiratory health. Nat Rev Microbiol. 2017;15:259–270.
27. Lappan R, Imbrogno K, Sikazwe C, et al. A microbiome case-control study of recurrent acute otitis media
identified potentially protective bacterial genera. BMC Microbiol. 2018;18:13.
28. Krueger A, Val S, Pérez-Losada M, et al. Relationship of the middle ear effusion microbiome to secretory mucin production in pediatric patients with chronic otitis media. Pediatr Infect Dis J. 2017;36:635–640.
29. Daniel M, Imtiaz-Umer S, Fergie N, et al. Bacterial involvement in otitis media with effusion. Int J Pediatr Otorhinolaryngol. 2012;76:1416–1422.
30. Kim EJ, Wang W, Deckwer WD, et al. Expression of the quorum-sensing regulatory protein LasR is strongly affected by iron and oxygen concentrations in cultures of Pseudomonas aeruginosa irrespective of cell density. Microbiology. 2005;151(pt 4):1127–1138.
31. Tunney MM, Field TR, Moriarty TF, et al. Detection of anaerobic bacteria in high numbers in sputum from patients with cystic fibrosis. Am J Respir Crit Care Med. 2008;177:995–1001.
32. Sillanpää S, Oikarinen S, Sipilä M, et al. Moraxella catarrhalis might be more common than expected in acute otitis media
in young finnish children. J Clin Microbiol. 2016;54:2373–2379.
33. Hilty M, Qi W, Brugger SD, et al. Nasopharyngeal microbiota in infants with acute otitis media
. J Infect Dis. 2012;205:1048–1055.
34. Murphy TF, Parameswaran GI. Moraxella catarrhalis, a human respiratory tract pathogen. Clin Infect Dis. 2009;49:124–131.
35. Brook I, Yocum P, Shah K. Aerobic and anaerobic bacteriology of otorrhea associated with tympanostomy tubes in children. Acta Otolaryngol. 1998;118:206–210.
36. Chonmaitree T, Jennings K, Golovko G, et al. Miyaji EN. Nasopharyngeal microbiota in infants and changes during viral upper respiratory tract infection and acute otitis media
. In: PLoS One. 2017;12:e0180630.
37. Teo SM, Mok D, Pham K, et al. The infant nasopharyngeal microbiome impacts severity of lower respiratory infection and risk of asthma development. Cell Host Microbe. 2015;17:704–715.