To identify metabolomic biomarkers predictive of Intensive Care Unit (ICU) mortality in adults.
Comprehensive metabolomic profiling of plasma at ICU admission to identify biomarkers associated with mortality has recently become feasible.
We performed metabolomic profiling of plasma from 90 ICU subjects enrolled in the BWH Registry of Critical Illness (RoCI). We tested individual metabolites and a Bayesian Network of metabolites for association with 28-day mortality, using logistic regression in R, and the CGBayesNets Package in MATLAB. Both individual metabolites and the network were tested for replication in an independent cohort of 149 adults enrolled in the Community Acquired Pneumonia and Sepsis Outcome Diagnostics (CAPSOD) study.
We tested variable metabolites for association with 28-day mortality. In RoCI, nearly one third of metabolites differed among ICU survivors versus those who died by day 28 (N = 57 metabolites, p<.05). Associations with 28-day mortality replicated for 31 of these metabolites (with p<.05) in the CAPSOD population. Replicating metabolites included lipids (N = 14), amino acids or amino acid breakdown products (N = 12), carbohydrates (N = 1), nucleotides (N = 3), and 1 peptide. Among 31 replicated metabolites, 25 were higher in subjects who progressed to die; all 6 metabolites that are lower in those who die are lipids. We used Bayesian modeling to form a metabolomic network of 7 metabolites associated with death (gamma-glutamylphenylalanine, gamma-glutamyltyrosine, 1-arachidonoylGPC(20:4), taurochenodeoxycholate, 3-(4-hydroxyphenyl) lactate, sucrose, kynurenine). This network achieved a 91% AUC predicting 28-day mortality in RoCI, and 74% of the AUC in CAPSOD (p<.001 in both populations).
Citation: Rogers AJ, McGeachie M, Baron RM, Gazourian L, Haspel JA, Nakahira K, et al. (2014) Metabolomic Derangements Are Associated with Mortality in Critically Ill Adult Patients. PLoS ONE 9(1): e87538. https://doi.org/10.1371/journal.pone.0087538
Editor: Samir M. Parikh, Beth Israel Deaconess Medical Center, United States of America
Received: September 16, 2013; Accepted: December 21, 2013; Published: January 30, 2014
Copyright: © 2014 Rogers et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: Dr. Rogers is supported by a grant from the Parker B. Francis foundation (http://www.francisfellowships.org/). Dr. McGeachie is supported by K12 HL089990; Dr. Gazourian is supported by T32 HL7118 and T32 HL007680; Dr. Fredenburgh is supported by K08 GM083207; Dr. Baron is supported by R01 HL091957, R01 HL112747, and P01 108801; Dr. Hunninghake is supported by NIH grant: K08 HL092222; and Dr. Choi is supported by P01 HL108801, R01 HL079904, and R01 HL112747. Dr. Matthay is supported by NHLBI R37 HL51856 and HL51854. The CAPSOD study was supported by grants from NIH (U01AI066569, P20RR016480, HHSN266200400064C), Pfizer Inc. (www.pfizer.com) and Roche Diagnostics Inc. (www.roche-diagnostics.us). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The CAPSOD trial was supported in part by Pfizer and Roche diagnostics. This does not alter the authors’ adherence to all the PLOS ONE policies on sharing data and materials.
More than 4 million patients are admitted to intensive care units (ICUs) each year in the United States.  Despite marked improvements in care for these critically ill patients, ,  approximately 500,000 die in the ICU each year.  Identifying biomarkers that distinguish the patients at highest risk for poor outcomes who could be targeted for novel therapies and clinical trials is critical. .
Investigating metabolomics to identify novel biomarkers in the ICU is promising for several reasons. First, the most widely used biomarker of ICU mortality identified to date is lactate, a byproduct of anaerobic metabolism. Second, in the spectrum from genotype to phenotype, metabolites, like proteomics, are correlated most closely to phenotype, and are thus likely much more highly correlated with disease.  While metabolomics was previously highly time-intensive and technologically limited, and thus could not plausibly be done on a large scale, technological advances in gas and lipid chromatography and mass spectroscopy now make such study possible. ,  Finally, metabolomic profiling has been performed in multiple complex trait diseases (notably recent work in cancer , diabetes , tuberculosis  and, most recently, septic shock in children ), in each case identifying important novel biologic pathways that contribute to pathogenesis and prognosis.
For all of these reasons, we performed metabolic profiling to identify prognosticators of 28-day mortality in a cohort of 90 critically ill adult ICU patients. We identified both widespread differences in individual metabolites and a network of interacting metabolites associated with 28-day mortality. The association of these metabolites with 28-day mortality was then replicated in an independent cohort of 150 adults with sepsis. We then compare our results to recent work that examined metabolomics in septic shock using these same metabolomics profiles but different analytical methods. .
Brigham and Women’s Hospital (BWH) Registry of Critical Illness (RoCI).
The BWH RoCI is approved by the Partners IRB committee (2008-P-000495). The protocol for recruitment for RoCI has been published in detail elsewhere.  Briefly, adult patients admitted to the BWH Medical Intensive Care Unit are eligible for inclusion in the RoCI within 72 hours of presentation, unless certain exclusion criteria are met (see Methods S1). Plasma is obtained on days 1, 3, and 7 of enrollment. Extensive phenotypic data (including age, gender, key comorbidities, and APACHE II score), laboratory, radiologic, and mortality data are recorded for all subjects. Between September 2008 and May 2010, 225 subjects were enrolled in the RoCI. Among these 225 subjects, 90 subjects were selected for metabolic profiling: 29 with SIRS, 30 with Sepsis, and 31 with sepsis-induced ARDS. Cases were selected for profiling based in part on IL-18 levels as part of a separate analysis  (sepsis and SIRS patients with low IL-18 levels, ARDS with high IL-18 levels; 31 of the 34 ARDS cases at that time were used). Cases were not selected with regards to risk of death or any known metabolic feature.
The protocol for enrollment in the Community Acquired Pneumonia and Sepsis Outcome Diagnostics (CAPSOD) study has been previously published. , ,  Briefly, 1152 patients with sepsis (≥2 Systemic Inflammatory Response (SIRS) criteria and presumed infection) were enrolled in emergency rooms associated with 3 US Hospitals. Blood samples and extensive phenotypic and laboratory data were recorded at enrollment. Survival/death was the primary outcome. The validation set of 149 patients (13% of the total CAPSOD cohort) was selected in five groups that reflected conventional concepts of sepsis progression as a pyramid from SIRS to sepsis and septic shock (see Methods S1).
Metabolomic profiles were generated independently for the BWH RoCI samples (N = 90 plasma samples from Day 1 of enrollment (within 72 hours of ICU admission), targeting 411 metabolites) and the CAPSOD samples (N = 150 targeting 439 metabolites) by Metabolon, Inc. Gas and liquid chromatography and mass spectroscopy was performed as described previously ,  and in the Methods S1. We removed metabolites with the lowest IQR of variability in the RoCI data, leaving 308 metabolites. We then limited statistical analyses to the metabolites that were also profiled in the CAPSOD database (N = 167 identical metabolites).
All raw metabolite concentrations were log2 transformed and normalized. We performed logistic regression in Rv.2.14, after adjustment for age, gender, race, malignancy status, and renal function (as estimated by the GFR-MDRD). Metabolites significantly associated with mortality in the RoCI cohort (p<.05) were tested for replication in CAPSOD; the significance level of p<.05 (one-sided) was used to define replication in the CAPSOD cohort.
Comprehensive details of Bayesian network selection are available in the Methods S1. Briefly, metabolite levels were log2 transformed and normalized separately in the RoCI and CAPSOD datasets. We then used 5-fold cross-validation on RoCI to arrive at hyper parameters for the Bayesian likelihood calculations. We performed 2500 bootstrap realizations of the RoCI dataset, and learned a pheno-centric conditional Gaussian Bayesian network  for each bootstrap realization. From the sample of 2500 networks, we built a consensus network by starting with the phenotype node and then adding, in sequence, the most frequent edge occurring in the bootstrap networks, and measuring the performance of that network on the dataset in cross-validation to estimate value of adding additional nodes to the network. Demographic variables were eligible for inclusion in the model, but none were ultimately selected. The network was then tested for association in CAPSOD without any parameter refitting. Predictive performance of the network was assessed by the convex hull of the Area Under the Receiver Operator Characteristic Curve (AUC).  Statistical significance was determined using the method of DeLong et al.  All Bayesian and related statistical computations were performed using the CGBayesNets package ,  in MATLAB. To compare our metabolomic network to that identified in a recent manuscript derived in CAPSOD and replicated in RoCI using the same metabolomics profiling data , we constructed a Bayesian network comprised of their variables and compared the predictive performance using the CGBayesNets package described above. Please see the Methods S1 for more details.
Baseline characteristics of subjects participating in the RoCI and CAPSOD cohorts are shown in Table 1. RoCI subjects were recruited after admission from the ICU, while CAPSOD subjects were enrolled in the ED with suspected sepsis. Of 149 CAPSOD subjects, 65 were admitted directly to the ICU on the day of admission and metabolomic profiling. As noted, RoCI subjects were younger, more frequently had cancer, had a higher baseline APACHE II score, were more often Caucasian, and had a higher mortality rate (all p<.05 by Wilcoxson or Fisher’s exact tests, Table 1). Of the 90 RoCI subjects, 60 had sepsis, including 25 of the 30 patients who died. As shown in Table 1, patients who die in the ICU differ in several phenotypic ways from those who survive, including higher APACHE score, more malignancy, and older age, among other variables.
RoCI Initial Cohort
We first assessed each of the 187 metabolites measured in both cohorts for an association with 28-day mortality. Nearly one third of the metabolites (N = 57) were associated with 28-day mortality in the RoCI cohort, after adjusting for age, gender, race, cancer status, and renal function (p<.05, Table S1). While p values were attenuated, the majority of metabolites remain significantly associated with death, even after further adjustment for APACHE II score, diabetic status, and use of immunosuppressive medication, including chemotherapy or systemic corticosteroids (Table S1). A majority of the 57 metabolites associated with death in the entire cohort were significant even when limited to the 55 RoCI patients without cancer (N = 35 of the 57 metabolites, p<.05, Table S1).
We then tested the 57 metabolites for replication in the CAPSOD cohort. Among the 57 metabolites associated with death in RoCI, 31 replicated in this independent population (again, after adjustment for gender, age, race, malignancy status, and renal function, p1-sided<.05). Replication of 31 metabolites is greater than the ∼3 replications expected by chance (pFisher’s exact = 6×10−9). Top metabolites are shown in Table 2, with a comprehensive list in Table S2. As shown, elevated lactate level was significantly associated with death in both populations, though it was not the top predictor. Plasma level of 1-arachidonoylglycero-phosphoethanolamine, the most significantly associated metabolite in RoCI, is shown in Figure 1a and 1b, with lower levels associated with mortality in both cohorts. Conversely, higher levels of 3-(4-hydroxyphenyl) lactate, a product of tyrosine and phenylalanine catabolism are associated with mortality in both cohorts (Figure 1c and 1d). Replicating metabolites cross multiple metabolic pathways, and include lipids (N = 14) amino acids and their breakdown products (N = 12), carbohydrates (N = 1), nucleotides (N = 3), and one peptide.
Log2-normalized level of 1-arachidonoylGPE(20:4), the most significantly associated metabolite in the RoCI cohort, is higher among ICU survivors in both the RoCI (light) and CAPSOD (dark) cohorts (Figure 1A and 1B). 3-(4-hydroxyphenyl) lactate (HPLA) is lower among ICU survivors in both cohorts (Figure 1C and 1D). Boxplots depict median (line) and range (edges of boxplots). P values are based on Wilcoxon’s test and are 1-sided in CAPSOD.
Among the 31 metabolites associated with mortality in both cohorts, the vast majority (25) were higher in patients who died. This result is much higher than the ∼15 one would expect by chance (pFisher’s exact = .02). All 6 metabolites that were lower in non-survivors were part of the lipid metabolism pathway (Table S2), though lipid metabolites represented only a minority of metabolites studied (76 of the 187).
A Metabolomic Network Associated with 28-day Mortality
Given the substantial proportion of metabolites that were altered in ICU survivors vs. non-survivors, we hypothesized that developing a predictive network of interacting metabolites would facilitate identification of individuals at high risk of death. We used a Bayesian network approach to identify metabolic networks, developing Bayesian networks in 2500 bootstrap realizations of the RoCI cohort and identifying those metabolites that appeared in the largest number of networks. No clinical variables were included in a sufficient number of networks to be included in the final model. Results of this iterative Bayesian Network approach are shown in Figure 2. This network contains seven metabolites, including two lipids, 2 amino acid/amino-acid break-down derivatives, 2 peptides, and 1 carbohydrate (kynurenine, sucrose, gamma-glutamylphenylalanine, gamma-glutamyltyrosine, 1-arachidonoyl-GPC (20:4), taurochenodeoxycholate, 3-(4-hydroxyphenyl) lactate) to predict 28-day mortality. Interestingly, the Bayesian network overlaps substantially with the top metabolites selected with logistic regression (all 7 are among the top 12 metabolites in Roci, Table S1).
This network maximizes the posterior likelihood of data over all networks predictive of mortality. This network achieves 91% AUC in RoCI and 74% AUC in CAPSOD. Directed edges indicate statistical dependence and do not represent causative links.
To evaluate the performance of this network, we tested the AUC for death in both the BWH and CAPSOD cohorts. The network achieved an AUC for 28-day mortality of 91% in BWH and 74% in CAPSOD. Both of these numbers represent statistically significant prediction (p<0.001 in both cohorts).
Comparison of Our Model with Prior Metabolomics Analyses in these Cohorts
These same cohorts and metabolomics profiles were previously analyzed, using the CAPSOD cohort as discovery and RoCI as replication, to identify a metabolomics profile associated with sepsis-related mortality.  Using different analytical methods, their final model included 5 metabolites (cis-4-decenoylcarnitite, 2-methylbytyroylcarnitine, butyroylcarnitine, hexanoylcarnitine, lactate, age, and hematocrit) and two phenotypic variables, hematocrit and age. Notably, none of the 5 metabolites were selected in our final model, although 4 of the metabolites were significant by logistic regression in RoCI (all but cis-4-decenoylcarnitite, Table 2). When the 5 metabolites and 2 phenotypes were combined into a Bayesian Network model (with Bayes Factors optimized in CAPSOD, and then tested without refitting in the RoCI cohort), the model performed similarly to our 7-metabolite model, explaining 84% of the AUC for death in CAPSOD and 84% in RoCI (p = .06 and.23 for comparisons with our models in CAPSOD and RoCI, respectively).
In this analysis, we tested 187 metabolites in two distinct cohorts of acutely ill patients to assess the potential for metabolomic profiling to identify biomarkers associated with death. The most important findings can be summarized as follows. First, in both populations, a substantial proportion (approximately one-third of all metabolites evaluated) were different in subjects who lived or died with a p<.05; this is a much greater number than would be expected by chance (∼8), and suggests major derangements in metabolism that are present within 24–72 hours of ICU admission in these patients. There were 31 individual metabolites associated with 28-day mortality in both cohorts; these metabolites were not restricted to a given metabolic pathway, but instead included diverse lipid, carbohydrate, amino acid, and nucleotide products. In addition to the analysis of individual metabolites, we developed a network model using a novel Bayesian Network technique that similarly replicated in both populations, explaining 91% of the AUC of death in the RoCI cohort, and 74% of the AUC in CAPSOD.
It is important to evaluate our analyses in the context of two recent metabolomics studies in critical illness, those of Mickiewicz et al.  in septic shock, and Langley et al.,  who analyzed the same two adult cohorts described in this work, focusing on septic shock mortality. Mickiewicz et al. tested 57 metabolites and identified a network of 18 metabolites that were associated with mortality in a single pediatric population with septic shock using principal components analysis. Several of our top metabolites were not studied. Metabolomics may enable identification of important differences in the pathophysiology of sepsis in adults versus children, but further work, including inclusion of identical metabolites, is needed. Langley et al. approached our same data sets, but tested first in the CAPSOD population, replicating in RoCI, and using different analytical methods to identify a predictive network. It is interesting to note that, even using the same metabolomics datasets, differences in analytic strategies led to completely different networks identified, without a single common metabolite in the 2 networks, yet both highly predictive of 28 day mortality in both cohorts. This difference highlights the fact that it is probably premature to focus on a single metabolite or two, or even a single network, as several metabolites have the potential to uncover novel biology. The fact that the derangements are so widespread (with nearly 1/3 of metabolites differing in ICU survivors vs. those who die) is itself noteworthy.
Several aspects of this analysis are worth highlighting. First, metabolomic profiling in critically ill adults to identify risk metabolites is only now feasible because of advances in mass spectroscopy and lipid chromatography. Second, while the RoCI cohort was recruited within the ICU, the CAPSOD replication cohort was comprised of ED patients with presumed sepsis or community acquired pneumonia. The ED is a critical site for identifying patients at highest risk for mortality , and biomarkers are particularly useful for patient triage or potential enrollment in clinical trials early in the course of illness. Thus, the 31 metabolites associated with mortality in both RoCI and CAPSOD may be especially promising biomarkers. Third, we used Bayesian networks to identify interdependent metabolites of interest that together are associated with increased risk of death. Bayesian networks are an attractive modeling methodology since they can model complex interactions between many variables of interest.  Finally, to our knowledge, this study is the first to use Bayesian networks not only to identify novel metabolomic biomarkers, but to test the predictive ability of a network of metabolites measured in an ICU setting to predict 28-day mortality in an independent replication cohort. A replicated Bayesian network is an important strength of this study, because it allows the network to be used to construct a straightforward “ICU mortality prediction score.”
Several features of the metabolomic results in our study are particularly interesting. Among the 31 metabolites that showed replicated association with mortality, the vast majority (25) were higher among those who died in the ICU. Interestingly, all 6 metabolites that were lower among subjects who died were part of the lipid metabolism pathway (Table S2). Mechanisms for this finding are unclear, but merit future functional validation. Several variables known to be associated with ICU mortality were also identified in our data, adding to the validity of our findings, First, lactate levels were higher in those who died in both RoCI and CAPSOD (p = .003 and.009, respectively, Table 2). Similarly, cortisol has long been recognized to be elevated in critically ill patients because of hypothalamus-pituitary-adrenal axis stimulation, ,  though exact mechanisms for cortisol elevation are still being elucidated . As expected, free cortisol level was lower among ICU survivors in both the RoCI and CAPSOD populations (p = .02 in RoCI,.01 in CAPSOD, Table S1).
The 187 metabolites examined in our analysis represent a small fraction of the more than 5000 human metabolites in plasma. In addition to the 187 metabolites in this analysis, an additional 121 metabolites were measured and variable in the RoCI cohort. We focused only on metabolites that could be replicated in CAPSOD; however, as shown in Table S3, an additional 36 metabolites are nominally associated in RoCI (p<.05), highlighting the existence of numerous additional metabolomic differences among survivors of critical illness.
Our work has several limitations. First, each of the 7 metabolites incorporated in our model was chosen because it was selected in the largest number of models created during iterative bootstrapping. While the resulting model replicated in the independent CAPSOD cohort, reinforcing its potential importance, it is highly probable that other metabolites contribute to these results and were missed. Determining the functional role of these metabolites will require laboratory-based experimental follow-up. We studied plasma of patients early in the course of critical illness; metabolomic changes on a cellular or organ-specific nature (like those identified with glutamine uptake and cancer , for example) could well be missed in a plasma sample. The RoCI subjects were enrolled in the ICU, a median of 2 days after ICU admission; the CAPSOD cohort we studied was at time 0 in the emergency room. Thus, our analytic strategy favored metabolites that remain altered over days in the ICU, and may have missed metabolites with biomarker potential that change more rapidly. Also RoCI subjects derived from a medical ICU, and mortality in the cohort is heavily driven by sepsis; similarly a majority of CAPSOD subjects had infection at enrollment. This may limit the generalizability of the results of this study in a non-medical ICU context. Nutritional status can substantially alter the plasma metabolomic profile; future studies will be strengthened by including dietary information as a covariate. ,  Because we were focused on discovering novel biomarkers and the overall differences in metabolomics profiling in the ICU, we chose a liberal α threshold of.05 in our logistic regression analyses in the RoCI discovery cohort. While we mitigated our risk of false positives by focusing only on replicated metabolites, it is certainly possible that false positives remain. Finally, the two cohorts differ in many baseline covariates, as shown in Table 1. While such heterogeneity likely means that some important metabolites did not replicate (if they are, for example, more important in cancer patients or vary with race), it also suggests that the identified metabolites are highly generalizable, and could be important in a wide variety of settings.
In summary, the combination of metabolomic profiling and Bayesian network analysis identifies both individual metabolites and a metabolomic network that are present in plasma early in the course of critical illness in adults and are associated with higher 28-day mortality in two different cohorts. Metabolomic profiling may provide new insights into both pathogenesis and prognosis in critically ill patients. Conceivably, better risk stratification for clinical trials and therapeutics can be achieved, perhaps by combining metabolomics measures with known protein biomarkers and standard clinical predictors.
Comprehensive Results, 187 Metabolites in RoCI. 1 P value and β are for association of a given metabolite with 28-day mortality in the RoCI cohort, using logistic regression after adjustment for age, gender, race, and renal function. Metabolite values are log2-transformed for testing. 2 P value and β are for association of a given metabolite with 28-day mortality in the RoCI cohort, using logistic regression after adjustment for all variables as described above (1), but additionally adjusted for baseline APACHE score. 3 P value and β are for association of a given metabolite with 28-day mortality in the RoCI cohort, using logistic regression after adjustment for all variables as described above (2), but additionally adjusted for DM status and use of immunosuppressive medications, including chemotherapy and/or systemic corticosteroids. 4 P value and β are for association of a given metabolite with 28-day mortality in the RoCI cohort using logistic regression after adjustment for all variables as described above (1), but limited to the 55 subjects without malignancy, of whom 10 died.
31 metabolites associated with mortality in both cohorts. 1Metabolite class is defined by Metabolon, inc. 2P value and β are for association of a given metabolite with 28-day mortality in the RoCI cohort, using logistic regression after adjustment for age, gender, race, malignancy status, and renal function. Metabolite values are log2-transformed for testing. 3P value in CAPSOD is similarly for logistic regression after the same adjustments. P values are 1-sided, as only metabolites with consistent direction are considered replicated.
Association with mortality in all 308 Metabolites tested in RoCI. Y: Metabolite tested only in RoCI, not in CAPSOD. P values and β are for association of a given metabolite with 28-day mortality in the RoCI cohort, using logistic regression after adjustment for age, gender, race, malignancy status, and renal function. Metabolite values are log2-transformed for testing.
Conceived and designed the experiments: AJR MM RMB LG KN LF BAR RMO VGF EPR CWW SK RL AMC. Performed the experiments: LG RMB RL AMC. Analyzed the data: AJR MM GMH BAR AMC. Contributed reagents/materials/analysis tools: RL SK AMC. Wrote the paper: AJR MM MAM RMB LG JAH KN GMH SK AMC.
- 1. Joint Commission Resources (2004) Improving Care in the ICU. Oakbrook TerraceIllinois: Joint Commission Resources, Inc. 128 p.
- 2. Zimmerman JE, Kramer AA, Knaus WA (2013) Changes in hospital mortality for United States intensive care unit admissions from 1988 to 2012. Crit Care 17: R81.
- 3. Li G, Malinchoc M, Cartin-Ceba R, Venkata CV, Kor DJ, et al. (2011) Eight-Year Trend of Acute Respiratory Distress Syndrome. American Journal of Respiratory and Critical Care Medicine 183: 59–66.
- 4. Calfee CS, Pugin J (2012) The search for diagnostic markers in sepsis: many miles yet to go. Am J Respir Crit Care Med 186: 2–4.
- 5. Gerszten RE, Wang TJ (2008) The search for new cardiovascular biomarkers. Nature 451: 949–952.
- 6. Veenstra TD (2012) Metabolomics: the final frontier? Genome Med 4: 40.
- 7. Dettmer K, Aronov PA, Hammock BD (2007) Mass spectrometry-based metabolomics. Mass Spectrom Rev 26: 51–78.
- 8. Jain M, Nilsson R, Sharma S, Madhusudhan N, Kitami T, et al. (2012) Metabolite profiling identifies a key role for glycine in rapid cancer cell proliferation. Science 336: 1040–1044.
- 9. Rhee EP, Cheng S, Larson MG, Walford GA, Lewis GD, et al. (2011) Lipid profiling identifies a triacylglycerol signature of insulin resistance and improves diabetes prediction in humans. J Clin Invest 121: 1402–1411.
- 10. Weiner J 3rd, Parida SK, Maertzdorf J, Black GF, Repsilber D, et al. (2012) Biomarkers of inflammation, immunosuppression and stress with active disease are revealed by metabolomic profiling of tuberculosis patients. PLoS One 7: e40221.
- 11. Mickiewicz B, Vogel HJ, Wong HR, Winston BW (2013) Metabolomics as a novel approach for early diagnosis of pediatric septic shock and its mortality. Am J Respir Crit Care Med 187: 967–976.
- 12. Langley RJ, Tsalik EL, Velkinburgh JC, Glickman SW, Rice BJ, et al. (2013) An integrated clinico-metabolomic model improves prediction of death in sepsis. Sci Transl Med 5: 195ra195.
- 13. Dolinay T, Kim YS, Howrylak J, Hunninghake GM, An CH, et al. (2012) Inflammasome-regulated cytokines are critical mediators of acute lung injury. Am J Respir Crit Care Med 185: 1225–1234.
- 14. Tsalik EL, Jaggers LB, Glickman SW, Langley RJ, van Velkinburgh JC, et al. (2012) Discriminative value of inflammatory biomarkers for suspected sepsis. J Emerg Med 43: 97–106.
- 15. Tsalik EL, Jones D, Nicholson B, Waring L, Liesenfeld O, et al. (2010) Multiplex PCR to diagnose bloodstream infections in patients admitted from the emergency department with sepsis. J Clin Microbiol 48: 26–33.
- 16. Lawton KA, Berger A, Mitchell M, Milgram KE, Evans AM, et al. (2008) Analysis of the adult human plasma metabolome. Pharmacogenomics 9: 383–397.
- 17. Sreekumar A, Poisson LM, Rajendiran TM, Khan AP, Cao Q, et al. (2009) Metabolomic profiles delineate potential role for sarcosine in prostate cancer progression. Nature 457: 910–914.
- 18. Heckerman D, Chickering M (1995) Learning Bayesian networks: The combination of knowledge and statistical data. Machine Learning 20: 197–243.
- 19. Lasko TA, Bhagwat JG, Zou KH, Ohno-Machado L (2005) The use of receiver operating characteristic curves in biomedical informatics. J Biomed Inform 38: 404–415.
- 20. DeLong ER, DeLong DM, Clarke-Pearson DL (1988) Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 44: 837–845.
- 21. McGeachie M, Chang HH (2011) MultiModal Conditional Gaussian Bayesian Networks Package.
- 22. Chang HH, McGeachie MJ (2011) Phenotype prediction by integrative network analysis of SNP and gene expression microarrays. 33rd Annual Intl Conf of the IEEE EMBS. Boston, MA, USA. 6849–6852.
- 23. Rivers E, Nguyen B, Havstad S, Ressler J, Muzzin A, et al. (2001) Early goal-directed therapy in the treatment of severe sepsis and septic shock. N Engl J Med 345: 1368–1377.
- 24. Rodin AS, Boerwinkle E (2005) Mining genetic epidemiology data with Bayesian networks I: Bayesian networks and example application (plasma apoE levels). Bioinformatics 21: 3273–3278.
- 25. Annane D, Sebille V, Troche G, Raphael JC, Gajdos P, et al. (2000) A 3-level prognostic classification in septic shock based on cortisol levels and cortisol response to corticotropin. JAMA 283: 1038–1045.
- 26. Hamrahian AH, Oseni TS, Arafah BM (2004) Measurements of serum free cortisol in critically ill patients. N Engl J Med 350: 1629–1638.
- 27. Boonen E, Vervenne H, Meersseman P, Andrew R, Mortier L, et al. (2013) Reduced cortisol metabolism during critical illness. N Engl J Med 368: 1477–1488.
- 28. Gibney MJ, Walsh M, Brennan L, Roche HM, German B, et al. (2005) Metabolomics in human nutrition: opportunities and challenges. Am J Clin Nutr 82: 497–503.
- 29. Focker M, Timmesfeld N, Scherag S, Knoll N, Singmann P, et al. (2012) Comparison of metabolic profiles of acutely ill and short-term weight recovered patients with anorexia nervosa reveals alterations of 33 out of 163 metabolites. J Psychiatr Res 46: 1600–1609.