Abstract Background Clinical risk factors in neonatal cardiac surgery do not fully capture discrepancies in outcomes. Targeted metabolomic analysis of plasma from neonates undergoing heart surgery with cardiopulmonary bypass was performed to determine associations with clinical outcomes. Methods and Result Samples and clinical variables from 149 neonates enrolled in the Corticosteroid Therapy in Neonates Undergoing Cardiopulmonary Bypass trial with surgical treatment for congenital heart disease between 2012 and 2016 were included. Blood samples were collected before skin incision, immediately after cardiopulmonary bypass, and 12 hours after surgery. Outcomes include composite morbidity/mortality (death, extracorporeal membrane oxygenation, cardiac arrest, acute kidney injury, and/or hepatic injury) and a cardiac composite (extracorporeal membrane oxygenation, cardiac arrest, or increase in lactate level), hepatic injury, and acute kidney injury. Targeted metabolite levels were determined by high‐resolution tandem liquid chromatography and mass spectrometry. Principal component and regression analyses were used to assess associations between metabolic profiles and outcomes, with 2 models created: a base clinical model and a base model+metabolites. Of the 193 metabolites examined, 40 were detected and quantified. The first principal component, principal component 1, was composed mostly of preoperative metabolites and was significantly associated with the composite morbidity/mortality, cardiac composite, and hepatic injury outcomes. In regression models, individual metabolites also improved model performance for the composite morbidity/mortality, cardiac composite, and hepatic injury outcomes. Significant disease pathways included myocardial injury (false discovery rate, 0.00091) and heart failure (false discovery rate, 0.041). Conclusions In neonatal cardiac surgery, perioperative metabolites were associated with postoperative outcomes and improved clinical model outcome associations. Preoperative metabolite levels alone may improve risk models and provide a basis for optimizing perioperative care. Keywords: acute kidney injury, cardiac surgical procedures, cardiopulmonary bypass, extracorporeal membrane oxygenation, heart defects, congenital, infant, newborn, risk factors Subject Categories: Metabolism, Cardiovascular Surgery, Congenital Heart Disease __________________________________________________________________ Nonstandard Abbreviations and Acronyms AIC Akaike Information Criterion AKI acute kidney injury CPB cardiopulmonary bypass FDR false discovery rate; HPLC, high performance liquid chromotography; KEGG, Kyoto Encyopledia of Genes and Genomics PC principal component PGE prostaglandin E‐1 Clinical Perspective. What Is New? * This study expands on perioperative metabolic changes, with a specific focus on neonates undergoing cardiopulmonary bypass. This cohort is larger than prior studies, is multicenter, and has a larger panel of metabolites that are included in the analysis. What Are the Clinical Implications? * A panel of metabolites that were primarily preoperative values was associated with postoperative outcomes, including death, renal and hepatic dysfunction, and cardiac arrest, which suggests that preoperative management could be more important than exposure to cardiopulmonary bypass. Despite significant improvement in morbidity and mortality over the past 30 years, congenital heart surgery remains high risk. Neonates are among the most vulnerable of these patients, comprising mortality rates of 10%, compared with 3.0% and 1.7% in infants and adults, respectively.[36] ^1 , [37]^2 Major and overall complication rates in neonates undergoing congenital heart surgery are high, at 19% and 67%, respectively,[38] ^3 leaving room for optimization of care. Current risk models incorporate clinical factors that are associated with increased morbidity and mortality, such as age, weight, preoperative status, including hospitalization before surgery, comorbidities, and surgical complexity. Perhaps the most widely recognized model, the Society of Thoracic Surgeons–European Association for Cardiothoracic Surgery category, combines age, surgical complexity, and preoperative length of stay to model mortality risk of surgery.[39] ^4 However clinical factors are static and mostly not modifiable and therefore not useful for operational improvements in care. Metabolomics is a dynamic measure of the physiological state that can be leveraged in pediatric cardiology. Because metabolites are the result of the complex interplay between the genetic substrate and environmental influences, a patient's metabolic profile offers insight on the phenotype of the patient in the moment of sampling. Metabolic profiling has shown superior risk prediction for death or hospitalization in adult patients with heart failure compared with current standards, such as brain natriuretic peptide.[40] ^5 Published data are less prevalent in the pediatric population, although studies have examined metabolomic fingerprinting in children undergoing cardiac surgery, with associations to outcome measures, such as mortality and intensive care unit length of stay.[41] ^6 , [42]^7 The physiological disruptions experienced in neonates undergoing congenital heart surgery are why metabolomics could be enlightening for this vulnerable population. Disruptive factors include smaller body size, cyanosis, nutrition challenges, high likelihood of prostaglandin exposure preoperatively, and use of cardiopulmonary bypass (CPB). Although integral to cardiac surgery and thus life saving, CPB contributes to physiological derangements through exposure to nonphysiological pump circuit surfaces, organ ischemia/reperfusion, cooling, and rewarming.[43] ^8 , [44]^9 The aims of this study are to: (1) describe perioperative changes in metabolic profiles in neonates requiring CPB, (2) associate metabolites with outcomes, and (3) identify potential metabolic pathways involved. METHODS Study Population This study is a secondary analysis of the National Heart, Lung, and Blood Institute–funded, Corticosteroid Therapy in Neonates Undergoing Cardiopulmonary Bypass trial, a randomized controlled trial of intraoperative methylprednisolone or placebo ([45]ClinicalTrials.gov Identifier: [46]NCT01579513). Participants were enrolled between 2012 and 2016 from 2 enrolling sites. The trial primary end point was the incidence of a clinically derived composite morbidity‐mortality outcome with additional secondary end points. The trial design and clinical outcomes have been previously reported.[47] ^10 , [48]^11 Data supporting the findings of this study are available on reasonable request from the corresponding author. Inclusion criteria for the parent study consisted of infants aged <1 month undergoing cardiac surgery with CPB. Exclusion criteria included prematurity, defined as <37 weeks postgestational age at the time of surgery, steroids within the 2 days before surgery, suspected infection, hypersensitivity that would be a contraindication to methylprednisolone, or use of mechanical circulatory support or active resuscitation at the time of proposed randomization. The protocol was approved by the institutional review board at the participating centers, and written informed consent was obtained from a parent/guardian before randomization. This secondary study was approved by the Johns Hopkins Institutional Review Board. Outcomes The parent study used a composite morbidity‐mortality outcome, which has been previously validated in neonates undergoing cardiac surgery with CPB.[49] ^12 The composite outcome was met if any of the following occurred after surgery but before hospital discharge: death, cardiac arrest, extracorporeal membrane oxygenation, renal injury (defined as creatinine >2 times normal), hepatic injury (defined as aspartate aminotransferase or alanine aminotransferase >2 times normal >36 hours postoperatively), or increasing lactate level (defined as >5 mmol/L). To focus on a more cardiac functional measure, we developed a cardiac composite outcome that only included death, extracorporeal membrane oxygenation, cardiac arrest, and increasing lactate level. Hepatic injury, using the same definition as in the composite outcome, and acute kidney injury (AKI), using the AKI Network's criteria, were further assessed individually from the composite morbidity/mortality outcome.[50] ^13 AKI was defined using the AKI Network's criteria for acute kidney injury as an increase in serum creatinine above the preoperative level by an absolute value of >0.3 mg/dL, a ≥50% increase, or urine output of <0.5 mL/kg per hour over a 6‐hour period.[51] ^14 Metabolite Identification and Quantification Blood samples were collected from neonates at 3 time points: (1) before skin incision and steroid randomization, (2) immediately after CPB, and (3) 12 hours postoperatively. Samples were held on ice until the plasma was isolated, aliquoted, and then stored at −80 °C until assayed for metabolite concentrations using a targeted panel of 193 metabolites representing >366 metabolic pathways (Molecular Determinants Core at Johns Hopkins All Children's Hospital, St. Petersburg, FL). The full assay list is in Table [52]S1. Samples were initially filtered, with extraction blanks (water) versus extraction reference (human plasma) included on the filtration plate. Filtration was accomplished using Supelco 96‐Well Protein Precipitation Filter Plate (Sigma Aldrich 55263‐U) paired with PlatePrep 96‐well Vacuum Manifold (Sigma Aldrich 57192‐U). Samples were then delivered to the mass spectrometry ionization source by high‐pressure liquid chromatography (HPLC), using a Shimadzu HPLC, a SIL‐30ACMP 6‐MTP autosampler, and Nexera LC‐30AD HPLC pumps. Targeted mass spectrometry assays used a Shimadzu 8060 triple‐quadrupole instrument equipped with an electrospray ionization source used in both positive and negative modes. Spectral Peak Data Processing After review and verification of the accuracy of peak integrations, measured areas for each compound were normalized for within‐batch drift by fitting a weighted scatterplot smoothing function generated from pooled quality control samples run between every 10th analytical sample.[53] ^15 , [54]^16 Drift correction was performed using the R LOESS statistical package. To account for variation in extraction efficiency, samples were scaled using the response of a suite of isotopically labeled internal standards spiked into each sample at the beginning of the extraction procedure. Variations in overall instrument sensitivity between analytical batches were corrected by transforming raw data so that the mean within‐batch areas for each compound were equivalent. Statistical Analysis Metabolomic data were corrected for sample osmolarity and then normalized, with the rare missing value replaced by half the minimum detectable level. No additional preprocessing of metabolomic data was performed. Using paired t‐test analysis, metabolite concentrations at each time point were compared using a false discovery rate of 5% (Benjamini‐Hochberg correction), which also established the P‐value cutoff for significance. Metabolites that had significant differential expression between various time points were further divided into subcolumns based on increase or decrease in levels. Additional features for association analysis were generated from metabolite concentrations, which included ratios between time points 1 and 2, 1 and 3, and 2 and 3, along with a cumulative value of all 3 time points. With the number of patients, metabolites measured at each time point, and the additional features derived between time points, this was a complex data set. Principal component (PC) analysis was performed to determine association between metabolites, generated features, and outcomes. PC analysis permitted simplification of the data set by reducing the dimensionality (ie, summarizing) the variations in the data points into smaller segments, which are the PCs.[55] ^17 To investigate the contribution of the different PCs to the various clinical outcomes, we used logistic regression models for binary outcomes. To develop the base clinical model, clinical factors with known associations with surgical outcomes were evaluated for inclusion. Preliminary factors group assignment, steroid administration, cardiopulmonary bypass duration, and similar surgical factors ultimately included surgical center, preoperative prostaglandin E‐1 (PGE) exposure, Society of Thoracic Surgeons category, preoperative lactate level, and gestational age at birth. PGE was included as a clinical variable because of the extensive hormonal,[56] ^18 renal,[57] ^19 hematologic,[58] ^20 and other systemic effects of PGE,[59] ^21 in addition to the associated high‐risk physiologies that require PGE infusions for ductal patency. The following performance metrics were determined for the base clinical model (including only adjustment factors): Akaike Information Criterion (AIC) for binary and continuous outcomes and Wald X^2 and c‐statistic for continuous outcomes. A second model was then created, by taking the base model followed by stepwise selection of the first 6 PC loads (based on examination of the scree plot), with P≤0.05 to include and P>0.05 to discard. Performance metrics were repeated for this second model (base model+PCs). A final, third, model was created to determine associations between individual metabolites (or features derived from them) and outcomes. A multistep approach for model building was used, starting with narrowing the list of potential associated features for each outcome. Features with a P<0.10 (indicating a potential association between the outcome and the features) were identified for each outcome, followed by a bootstrap resampling approach (500 samples, including between 20% and 80% of the total observations) to identify features with high reliability. Features with scores >50% were considered highly reliable. For the next step, we selected features with reliability scores >30%. Thereafter, the final regression models were built using backward elimination (P<0.05 to remain) and included the clinical adjustment variables, the first 6 PC loads, and all metabolites or features thereof that met the above‐mentioned reliability criteria. In some cases where 2 highly correlated features were both candidates for inclusion in the multivariable model, an intermediate step was added to compare the model with either feature, and the feature resulting in the best model fit (based on a lower model AIC) was selected. Metabolite Enrichment Disease and Pathway‐Associated Analysis Final metabolites from the models or their features that improved an outcome were then evaluated using Metaboanalyst 5.0 for disease and metabolic pathways.[60] ^22 Findings were then compared for relevance to the outcome assessed and prior studies linking to diseases or pathways. RESULTS Subject Demographics Samples from a total of 154 participants were available; of those, 5 subjects were excluded because of a missing sample for at least 1 of the time points. The final cohort included 149 patients who had metabolomic data available at all 3 time points. Demographics and baseline characteristics are shown in Table [61]1. The cohort consisted of 89 male patients (60%), and patients fit into following diagnostic categories: 27 (18%) biventricular physiology and arch hypoplasia, 37 (25%) transposition of the great arteries, 40 (27%) biventricular physiology not previously listed, and 45 (30%) with single‐ventricle physiology. Most (79%) remained on PGE before surgery, and 87% had PGE exposure at some point before surgery. This is expected as most congenital heart disease that requires surgery in the neonatal period has a ductal‐dependent physiology. Median age at surgery was 7 days (interquartile range, 5–10 days). A total of 55 patients (37%) experienced the composite morbidity/mortality trial outcome, 34 (23%) experienced the cardiac composite, 66 (44%) experienced AKI, 29 (19%) experienced elevated lactate level, and 38 (26%) experienced hepatic injury. Table 1. Cohort Characteristics and Clinical Status Patient characteristics Values, mean±SD or N (%) Patient characteristics (cont.) Values, mean±SD or N (%) Total (N) 149 Surgical outcomes Site 2 (vs site 1) 38 (26) Low cardiac output syndrome 69 (46) Methylprednisolone arm 71 (48) Mechanical cardiac support 3 (2) Increase inotropic support by 100% 50 (34) Characteristics/clinical history New inotropic agents 40 (27) Age, d 8.7±5.3 Highest IS at 36 h 17.1±4.7 Gestational age at birth, wk 38.9±1.2 Highest VIS at 36 h 19.2±6.9 Sex (male) 89 (60) Highest lactate level at 36 h (mmol/L) 4.0±2.4 Diagnosis Ventilation time, d 7.7±15.1 Biventricular with arch hypoplasia 27 (18) ICU length of stay, d 17.0±20.6 TGA 37 (25) Hospital length of stay, d 30.4±39.1 All other biventricular disorders 40 (27) SV 45 (30) Mortality/morbidity composite outcomes 55 (37) Death 7 (5) Preoperative status Cardiac arrest 9 (6) Intubation before surgery 36 (24) ECMO 7 (5) PGE within 24 h of surgery 117 (79) Renal 6 (4) PGE ever before surgery 130 (87) Hepatic 38 (26) MOD ever before surgery 11 (7) Lactate level 28 (19) Highest creatinine (mg/dL) 0.8±0.5 Cardiac composite outcome 34 (23) Highest lactate level (mmol/L) 3.5±2.7 AKI 66 (44) Lactate level on day of surgery (mmol) 1.1±0.43 Corrective procedure 90 (60) STAT category 1–3 32 (21) 4 77 (52) 5 40 (27) CPB time, min 74.8±42.5 DHCA, min 9.9±65.1 [62]Open in a new tab AKI indicates acute kidney injury; CPB, cardiopulmonary bypass; DHCA, deep hypothermic circulatory arrest; ECMO, extracorporeal membrane oxygenation; ICU, intensive care unit; IS, inotropic score; MOD, multiorgan dysfunction; PGE, prostaglandin E‐1; STAT, Society of Thoracic Surgeons; SV, single ventricle; TGA, transposition of the great arteries; and VIS, vasopressor/inotropic score. Differential Expression of Metabolites at Time Points Of the 193 metabolites examined, 40 were detected and quantifiable in all 149 participants. The other 153 metabolites were not consistently identified or quantified. Results from paired t‐test analysis comparing metabolite concentrations at various time points are shown in Table [63]2. A total of 34 metabolites changed significantly between all measured time points. The general pattern was either an immediate postoperative decrease or increase, with a trend back toward the preoperative baseline. There were many exceptions to this pattern, including methyladenosine, which showed a delayed increase at the 12‐hour post‐op; leucine and nicotinamide, which had the inverse finding of a delayed decrease at the 12‐hour post‐op time point; arginine, isoleucine, methionine, methionine sulfoxide, ornithine, serine, and glycocholic acid, all of which continued to decrease at the time points measured; and paraxanthine, indoxyl sulfate, and hippuric acid, which increased across all time points. Table 2. Metabolites With Statistically Significant Differences Between Different Time Points Preoperative vs T2 T2 vs T3 Preoperative vs T3 Arginine* Arginine* Arginine* Aspartic acid* Isoleucine* Asparagine Carnitine* Leucine* Aspartic acid* Creatinine* Methionine* Carnitine* Glutamine* Methionine sulfoxide* Creatinine Glutamic acid* Nicotinamide* Glutamine* Histidine* Ornithine* Glutamic acid* 4‐Hydroxyproline* Proline* Histidine* Isoleucine* Gluconic acid* 4‐Hydroxyproline* Lysine* Galactitol* Isoleucine* Methionine* 4‐Pyridoxic acid* Leucine* Methionine sulfoxide* Glycocholic acid* Lysine* Ornithine* Sn‐glycero‐3‐phosphochocholine* Methionine* Phenylalanine* 1‐Methyladenosine† Methionine sulfoxide* Proline* N‐acetylserine† Nicotinamide* Serine* Histidine† Ornithine* Tryptophan* Homoserine† Phenylalanine* Homoserine* Carnitine† Proline* 5‐Oxoproline* Creatinine† Serine* 3‐Methoxytyrosine* Asparagine† Tryptophan* N‐acetylserine* 5‐Oxoproline* Glycocholic acid* 3‐Methoxytyrosine* Gluconic acid^† Glycocholic acid* N‐acetyltryptophan^† Gluconic acid† Paraxanthine^† Paraxanthine† Galactitol^† Galactitol† Indoxyl sulfate^† 1‐Methyladenosine† Sn‐glycero‐3‐phosphocholine^† Indoxyl sulfate† Hippuric acid^† Sn‐glycero‐3‐phosphocholine† Hippuric acid† N‐acetylserine† [64]Open in a new tab For metabolites with statistically significant differences, the fold change cutoff was 0.5/2.0 (P‐value cutoff, 0.0125). * indicates significant decrease; †, significant increase. T2 indicates immediate postoperative; and T3, 12 hours postoperative. Association of Metabolites to Clinical Outcomes The first 6 components of the PC analysis performed on metabolites and features individually had loads >3% and together explained 51% of the cumulative variance in metabolic compounds (detailed features of PCs in Table [65]S2). In univariate analysis (Table [66]3), PC1 alone was significantly associated with the composite morbidity/mortality, cardiac composite, and hepatic injury outcomes. PC1 and PC3 were associated with acute kidney injury, and PC1 and PC6 were associated with increasing lactate level. Of the PC loads, PC1 was of special interest as 14 of the 20 metabolites within this group were preoperative (Table [67]S2). Table 3. Association of PCs With Clinical Outcomes With Univariate Analysis Outcome PC[68]^* OR (95% CI) P value Composite: mortality/morbidity PC1 1.21 (1.10–1.33) <0.001 Composite: cardiac[69]^† PC1 1.14 (1.05–1.25) 0.003 Hepatic PC1 1.27 (1.14–1.41) <0.001 Lactate level PC1 1.13 (1.03–1.24) 0.01 PC6 1.25 (1.01–1.54) 0.04 AKI PC1 1.06 (1.01–1.13) 0.03 PC3 1.09 (1.01–1.16) 0.02 [70]Open in a new tab AKI indicates acute kidney injury; OR, odds ratio; and PC, principal component. ^* Significant PC. ^^† Only includes death, extracorporeal membrane oxygenation, cardiac arrest, and lactate level outcomes. A base clinical model (adjusted for enrolling site, PGE use, Society of Thoracic Surgeons category, presurgery lactate level, and gestational age at birth) was significantly associated with composite morbidity/mortality, cardiac composite, and hepatic injury outcomes, but not AKI (Table [71]4). Of note, an additional analysis including steroid versus placebo administration was also run, without changes in results, so this factor was not included in further analysis. To determine if metabolite levels could improve the base clinical model performance, the 6 top PCs were incorporated into a regression analysis with the base clinical model (base model+metabolite PCs) and compared with the base clinical model using AIC, Wald X^2, and c‐statistic as performance metrics (Table [72]4 and Table [73]S3). For comparison, a lower AIC and higher c‐statistic indicate improved model fit. Both the base clinical model and the base model+metabolite PCs were significant in their associations to the outcomes. However, the second model had improved model fit, as indicated in the difference in AIC and c‐statistic. The addition of PC1 significantly improved the clinical model performance for the composite morbidity/mortality outcome, composite cardiac outcome, and hepatic injury (PC1 and PC3 for acute kidney injury and PC1 and PC6 for lactate level). Table 4. Model Comparison of Associations Between PC Loads and Binary Outcomes With Multivariate Regression Analysis Outcome Base model[74]^* Base model+PC Base model+individual metabolites[75]^† AIC Wald χ^2 P value C‐statistic AIC Wald χ^2 P value C‐statistic Significant PC AIC Wald χ^2 P value C‐statistic Significant PC Composite: mortality/morbidity 158.3 33.5 <0.001 0.828 137.3 38.2 <0.001 0.879 PC1 90.0 27.6 0.006 0.967 None Composite: cardiac[76]^‡ 140.2 24.7 <0.001 0.805 129.2 29.0 <0.001 0.850 PC1 113.0 30.4 <0.001 0.905 None Hepatic 150.0 23.1 <0.001 0.788 121.7 32.5 <0.001 0.884 PC1 99.5 32.1 <0.001 0.936 None Lactate level 123.8 32.2 <0.001 0.817 116.6 27.3 <0.001 0.845 PC1 93.5 27.1 0.001 0.930 None AKI 210.9 5.3 0.38 0.607 204.6 13.6 0.06 0.672 PC1, PC3 182.0 29.6 0.003 0.811 PC6 [77]Open in a new tab AIC indicates Akaike Information Criterion; AKI, acute kidney injury; and PC, principal component. ^* Adjusted for site, preoperative prostaglandin E‐1, Society of Thoracic Surgeons category, preoperative lactate value, and gestational age at birth. ^^† Individual metabolites, as listed in Table [78]S4. ^^‡ Only includes death, extracorporeal membrane oxygenation, cardiac arrest, and lactate level outcomes. Association of Individual Metabolites With Outcomes Given the significance of the association between clinical outcomes and metabolite PC loads, and improvement in base model+metabolite PCs compared with the clinical base model alone, we further investigated the association between individual metabolites (or features derived from them) and outcomes. As shown in Table [79]5, individual metabolites and features were significantly associated with outcomes. Individual performance metrics for the base model+individual metabolites are detailed in Table [80]S4. Significant individual metabolites improved (lower AIC and higher c‐statistic) the clinical model for the composite morbidity/mortality outcome, composite cardiac outcome, hepatic injury, and lactate level outcomes (Table [81]4). Association of individual metabolites with outcomes was not improved by including PC loads in the model for the composite morbidity/mortality outcome, composite cardiac outcome, hepatic injury, and lactate level outcomes. Table 5. Significant Individual Metabolites or Metabolite Features Associated With an Outcome Composite: morbidity/mortality Composite: cardiac Composite: hepatic Composite: lactate level AKI t1 Ornithine t1 Phenylalanine t1 Methionine t3 Cystine t3 4‐Hydroxyproline t3 3‐Methoxytyrosine r21 Homoserine Cum proline t3 proline t3 leucine t1 Carnitine t1 Aspartic acid t3 Tyrosine t1 Methionine r31 Galactitol t3 Proline t3 Leucine t1 Ornithine t1 Isoleucine t1 Aspartic acid t3 Alanine t3 Glutamic acid t3 4‐Hydroxyproline t3 Serine [82]Open in a new tab AKI indicates acute kidney injury; Cum, cumulative concentration from all time points; r21, ratio between time points 1 and 2; r31, ratio between time points 1 and 3; t1, time point 1 (preop); t2, time point 2 (immediate post‐op); t3, time point 3 (12 hours post‐op). Metabolite Enrichment Pathway and Disease‐Associated Analysis For the individual metabolites or features associated with clinical outcomes, we explored the metabolic pathways (Kyoto Encyclopedia of Genes and Genomes, (KEGG)) associated and if disease associated. Significant false discovery rate (FDR; <0.05) KEGG pathways represented included aminoacyl‐tRNA biosynthesis, arginine biosynthesis, and metabolism of multiple amino acids. Using metabolite enrichment disease pathway analysis ([83]Figure and Table [84]6), individual metabolites were significantly (FDR, <0.05) associated with disease pathways, including seizures (FDR, 2.67×10^−6), early markers of myocardial injury (FDR, 0.00091), and heart failure (FDR, 0.041). Figure  . Metabolite enrichment disease pathway analysis. Figure   [85]Open in a new tab Metabolite enrichment disease pathway analysis using individual metabolites that improved on the base model (listed in S4). Disease associations with metabolites are in the figure listed above. Asterisks (*) indicate attention to cardiac diseases that are of special interest to this patient population. Table 6. Metabolite Disease Enrichment Analysis Metabolite set enrichment analysis Total Hits Raw P value FDR Perioperative metabolites associated with disease Acute seizures 14 7 9.78E‐09 2.67E‐06 L‐alanine, L‐methionine, L‐leucine, L‐isoleucine, L‐phenylalanine, L‐histidine, L‐tryptophan Different seizure disorders 24 8 2.83E‐08 3.87E‐06 L‐alanine, L‐histidine, L‐isoleucine, L‐leucine, L‐methionine, L‐phenylalanine, L‐threonine, L‐tryptophan, L‐tyrosine Refractory localization‐related epilepsy 10 5 2.26E‐06 0.000154 L‐lysine, L‐leucine, L‐isoleucine, L‐phenylalanine, L‐tyrosine Early markers of myocardial injury 14 5 1.69E‐05 0.00091 L‐alanine, L‐isoleucine, L‐leucine, L‐proline, L‐serine Heart failure 10 3 0.00212 0.0414 L‐isoleucine, L‐leucine, L‐tyrosine Myocardial infarction 4 2 1.0 0.104 L‐phenylalanine, L‐tyrosine [86]Open in a new tab FDR indicates false discovery rate. DISCUSSION Early identification of patients at risk for postoperative morbidity could improve outcomes. Clinical models alone using known, or suspected, risk factors, such as age, known comorbidities, weight, and complexity of surgery, have offered some insight into mortality risk, but may have no applicability to subtler organ‐specific injury. The findings of this study show a significant association with metabolites and multiple clinical outcomes, and metabolite data improved on a clinical‐only risk model. This study also suggests that metabolites may provide an actionable target to improve outcomes following congenital cardiac surgery. Prior studies comparing perioperative shifts in metabolic profiles and the relationship to postoperative outcomes have suggested the utility of incorporating metabolites into outcome prediction and/or risk stratification.[87] ^6 , [88]^7 These pediatric studies have been limited by small patient cohort sizes from single centers and a broad age range. Our work expands on these studies using a multicenter and larger cohort, with a specific focus on neonates requiring cardiac surgery with CPB. This study also benefits from an expanded targeted assay panel of 193 metabolites. These measures allowed us to explore the potential metabolic pathways more specific to neonatal cardiac surgery and recovery. In regard to the significant perioperative metabolite changes identified, these mostly represented aminoacyl‐tRNA biosynthesis, arginine biosynthesis, and metabolism of multiple amino acids pathways. Enrichment analysis revealed that many of the individual metabolites were previously identified in studies of cardiac dysfunction and seizures, reflecting both cardiac dysfunction[89] ^23 , [90]^24 , [91]^25 and neurologic injury,[92] ^26 which are common postoperative morbidities. Although seizures were not recorded in the primary trial, we have shown previously that the brain injury biomarker GFAP (glial fibrillary acidic protein) was increased postoperatively in this cohort and was associated with a worse Bayley Scale of Infant Development III score at 12 months of age,[93] ^27 providing evidence of at least postoperative subclinical brain injury in this cohort and consistent with findings of seizure‐associated metabolites. Three (L‐isoleucine, L‐leucine, and L‐tyrosine) of the 8 amino acids associated with seizures were also associated with acute myocardial injury and heart failure, which could reflect a more global injury phenotype as opposed to specific organ injury. Previously, dysregulation of glutathione and amino acid synthesis pathways was identified after congenital heart surgery and was confirmed in this study.[94] ^6 , [95]^7 Shifts in glutathione metabolism are relevant to our study because glutathione has a role in antioxidant defense, nutrient metabolism, and regulation of protein synthesis, cell proliferation, and apoptosis.[96] ^28 We found this compelling as neonates typically have limited glycogen and fat energy stores; therefore, they are more reliant on glucose‐derived energy. Feeding preoperatively is typically restricted in neonates with congenital heart disease, with dependence on intravenous nutrition in the form of dextrose‐containing fluids. Because of these feeding patterns, we postulate that our patients are at risk of even more poorly developed energy stores, which could explain glutathione dysregulation. For most outcomes measured, high PC1 was universally associated with poor outcomes. Analyzing the individual components of this PC, we were surprised to find many of the metabolites were preoperative values, with the remaining components being cumulative values. This suggests that these patients were at risk because of their preoperative physiological state, and not because of surgical factors (CPB duration, circulatory arrest, and cross‐clamp time). In the initial conceptualization of this study, we had hypothesized that the changes from surgery and CPB would be the most influential on patient outcomes. The recurring importance of PC1 strongly suggests that the patient's preoperative physiological state could be most important modification for improving morbidity and mortality.[97] ^23 , [98]^24 , [99]^25 , [100]^29 , [101]^30 As in the study by Davidson et al,[102] ^6 essential amino acids (leucine, methionine, and phenylalanine) were found to be associated with worse outcomes, in addition to conditionally essential (cystine and proline). The continued relevance of amino acids between our study and prior work supports the role that nutritional status plays in improving outcomes. Despite increased cohort size compared with prior works, this study remains limited by a relatively small cohort produced from only 2 participating sites. Statistical analysis relied heavily on PC analysis, which allowed for interpretation of complex data in this relatively small cohort. A more straightforward approach using regression analysis for each metabolite against each outcome would likely have missed significant findings because of the smaller sample size. The drawback is that a general metabolic profile can be associated with an outcome (a PC), without individual metabolites being as clearly linked. Other limitations include the primary trial enrolling criteria, and exclusion of younger neonates (aged <37 weeks) or older infants/children, which also left us unable to determine whether metabolite changes differ between these patient populations. These results may not be necessarily generalizable to nonneonates. Last, we used a targeted metabolomic analysis, which lacked a full range of metabolites, such as membrane and vasoactive lipids, which could potentially be informative. The association between metabolites and outcomes in neonatal cardiac surgery that we have demonstrated herein suggests an additional tool beyond clinical data to help identify patients at risk for poor outcomes. What we found most promising was the association between preoperative metabolic profiles and outcomes of interest, suggesting improvement in outcomes begins preoperatively. Further research would include validation studies, and metabolomics in expanded age groups, in addition to targeted analysis of identified pathways. The proposed mechanism of preoperative nutritional status and effect on operational outcomes would bear further investigation also. Sources of Funding This study was supported by The Children's Heart Foundation (Dr Everett). Disclosures None. Supporting information Tables S1–S4 [103]Click here for additional data file.^ (219.5KB, pdf) Supplemental Material is available at [104]https://www.ahajournals.org/doi/suppl/10.1161/JAHA.121.024996 For Sources of Funding and Disclosures, see page 9. REFERENCES