Metabolic Trajectories During Surgical Stress in Patients Undergoing Cardiac Surgery

Authors: Dohyun Ku, Jackson Bartelt, Jenna Feeley, M. Citlalli Perez-Guzman, Lizda Guerrero-Arroyo, Amalia Abraham, Saaki Kollipara, Nicolas Gonzalez, Sabeena Usman, Helaina E. Huneault, Andrea Corujo-Rodriguez, Georgia M. Davis, Richard G. Kibbey, Dean P. Jones, Thomas R. Ziegler, Michael Halkos, Arshed A. Quyyumi, M. Ryan Smith, Jing Li, Francisco J. Pasquel

Abstract

Stress hyperglycemia (SH) during acute illness is linked to adverse surgical outcomes, yet the accompanying metabolic perturbations are incompletely characterized. We profiled longitudinal metabolic changes in adults without diabetes undergoing cardiac surgery to identify pathways associated with perioperative SH (defined as point of care glucose ≥140 mg/dL on ≥3 readings or ≥180 mg/dL once). Blood was collected at baseline before surgery (T0) and at 2 h (T1), 24–48 h (T2), and 72–96 h (T3) after surgical initiation. High resolution metabolomics (LC–MS) was integrated with continuous glucose monitoring and inflammatory/cardiac biomarkers. At T0, several pathways were associated with subsequent SH, including bile acid metabolism, the carnitine shuttle, and fatty acid oxidation, suggesting preoperative metabolic susceptibility. In longitudinal analyses, participants who developed SH showed coordinated postoperative changes with significant enrichment of pathways not evident at baseline, C21 steroid hormone biosynthesis, glycerophospholipid metabolism, and glycosphingolipid (ceramide) metabolism, consistent with lipid remodeling and inflammatory signaling during surgical stress. Individuals with SH also exhibited higher inflammatory biomarker levels (high sensitivity C reactive protein and soluble urokinase plasminogen activator receptor). A machine learning model using early metabolomic features predicted SH with an area under the receiver operating characteristic curve of 0.86. These findings highlight distinct preoperative and perioperative metabolic trajectories associated with SH and implicate established dysglycemia-related pathways, as well as stress-induced pathways in perioperative metabolic dysregulation. Pathway enrichment analyses were exploratory and hypothesis-generating; validation in larger cohorts and assessment of implications for clinical outcomes are warranted.

INTRODUCTION

Glucose is the most routinely assayed metabolite in clinical medicine, and it’s widely used to gauge metabolic stability in patients with and without diabetes. However, during acute physiological stress such as major surgery, hyperglycemia may signal a much deeper biochemical disruption. Perioperative glucose elevations are a recognized hallmark of surgical stress and correlate with higher rates of surgical-site infection, impaired wound healing, prolonged hospital stay, and death.^1^ However, the broader metabolic derangements that accompany glucose excursions, as well as the mechanistic links between hyperglycemia and poor outcomes, remain inadequately defined.

Coronary-artery bypass grafting (CABG) is a well-standardized, high-intensity surgical procedure that evokes reproducible metabolic and inflammatory responses. Stress hyperglycemia (SH) occurs commonly in CABG patients and is consistently associated with adverse events,^2^ yet the metabolic pathways driving this response, and their clinical relevance remain largely unknown. Since CABG follows a predictable operative timeline, it offers a unique in-vivo model for studying the temporal evolution of stress-induced metabolic changes.

Recent advances in high-resolution metabolomics, continuous glucose monitoring (CGM), and integrative bioinformatics enable simultaneous tracking of thousands of metabolites alongside real-time glycemic profiles. Here, we applied longitudinal plasma metabolomics to detect changes in individual metabolites, CGM to characterize glucose changes, and targeted biomarker assays to determine if metabolites and biomarkers fluctuated with surgical stress and SH among adults without diabetes undergoing CABG. Using pathway enrichment analysis and machine learning classification, we defined metabolic trajectories linked to perioperative SH and delineated the biochemical networks that couple stress-induced glucose rises to inflammatory and metabolic pathways.

RESULTS

Figure 1 provides an overview of the study design and analytical workflow. Further methodological details are provided in the Methods section.

Participant Characteristics

A total of 48 participants with metabolomics data were included in the analysis [mean age = 64.0 ± 10.1 years; 42% male; body mass index (BMI) = 28.3 ± 4.5 kg/m^2^]. Among these participants, 28 (58%) developed SH. Supplementary Table 1 presents the demographic characteristics of study participants.

Metabolites at Baseline (before Surgery) Associated with Stress Hyperglycemia

At baseline, two-sample t-tests were performed to compare SH and non-SH participants, identifying 602 and 769 significant putative metabolites using reverse-phase chromatography (C18) and Hydrophilic Interaction Liquid Chromatography (HILIC), respectively. Subsequent pathway enrichment analysis identified 22 significant metabolic pathways via C18 and 17 via HILIC (see Supplementary Table 2).

We applied a machine learning pipeline to predict SH using baseline metabolite levels identified from pathway enrichment analysis. This pipeline with logistic regression demonstrated the best predictive performance (AUC of 0.79, accuracy of 0.75, sensitivity of 0.79, and specificity of 0.70). ROC curves for all four methods are shown in Figure 2A. Comparative performance results for all models are presented in Supplementary Table 3. Figure 2B presents the top 20 metabolites ranked by feature importance in the pipeline. Detailed information on m/z values, retention times, annotations, and abbreviations is provided in Supplementary Table 4.

Significant Pathways Associated with Time Above Range (>140 mg/dL) and Glycemic Variability (Coefficient of Variation)

We investigated temporal metabolomic patterns in the whole cohort using all four time 24 hours before surgery (T0), 2 hours after surgery initiation (T1), 24–48 hours after surgery initiation (T2), and 72–96 hours after surgery initiation (T3). To identify metabolic pathways associated with glycemic variability and hyperglycemic exposure during the 48 hours following surgery initiation, we performed clinical difference analyses using two continuous coefficient of variation (CV) and time above range (TAR, >140 mg/dL). To obtain complete CGM traces for the 48-hour after surgery, we imputed missing readings and calibrated CGM values with concurrent POCT results. Calibration results and examples are provided in the Supplementary Section. From the resulting data, we calculated CV and TAR, each used as a continuous interacting predictor with time in the nonlinear mixed effects model (NMEM) for every metabolite across four time points. Figure 3A reveals a clear separation between groups, with participants experiencing SH consistently showing higher glucose levels in the perioperative period. This difference gradually diminishes until approximately 40 hours after surgery initiation.

After applying subsequent pathway enrichment analysis, we identified several putative metabolites associated with CV (1,628 in C18 and 1,667 in HILIC) and TAR (991 in C18 and 1,115 in HILIC). Pathways with the highest association with glycemic variability by CV included glycosphingolipid metabolism, glycerophospholipid metabolism, purine metabolism, vitamin B9 (folate) metabolism, glycine, serine, alanine and threonine metabolism, glycosphingolipid metabolism, and phosphatidylinositol phosphate metabolism. Supplementary Table 5 shows all significant pathways associated with CV along with overlap and pathway size. Pathways with the highest association with TAR included pyrimidine metabolism, pentose phosphate pathway, pentose and glucuronate interconversions, vitamin B9 (folate) metabolism, aspartate and asparagine metabolism, glycerophospholipid metabolism, glycosphingolipid metabolism, carnitine shuttle, and fatty acid activation. Supplementary Table 6 shows all significant pathways associated with TAR including overlap and pathway size. Figure 2C shows pathway enrichment analyses showing the most strongly associated pathways with glycemic variability (CV) and time spent above 140 mg/dL (TAR).

Dominant Temporal Metabolic Trajectories during Surgical Stress

Based on all four time points in the whole cohort, NMEM identified 5,123 and 6,500 significant mass spectrometry features showing significant temporal variation using C18 and HILIC, respectively. Several metabolic pathways (e.g. arachidonic acid metabolism, C21-steroid hormone biosynthesis and metabolism, glycerophospholipid metabolism, carnitine shuttle, bile acid metabolism) were identified via C18 and HILIC (Supplementary Table 7 shows all significant pathways including significance and overlap/pathway size). Figure 2C shows pathway enrichment analyses showing the most strongly associated pathways with surgical stress in the whole cohort.

All putative metabolites from significant pathways (n=313) were combined, and metabolite trajectories were extracted from the NMEM full model in the whole cohort statistical analysis, identifying a single, predominant “early-spike/rapid-recovery” trajectory that accounted for 35.8% of all profiled compounds (86 detected on the C18 column and 74 on the HILIC column). Metabolite intensities in this cluster rose sharply from baseline, peaking 2 hours after surgery, followed by rapid decline to baseline or slightly below within 24–48 hours, with levels remaining stable through 72–96 hours, shown in Figure 3B. Consistent with the dominant clustering pattern, the Figure 3C heatmap highlights metabolites that undergo significant up- and down-regulation at two hours after surgery initiation.

Dominant Metabolomic Temporal Patterns from Significant Pathways Associated with Hyperglycemia

To determine whether metabolomic temporal patterns varied by glycemic status, we applied NMEM using SH as a binary group variable. This analysis identified 1,733 significant putative compounds via C18 and 1,353 via HILIC. Subsequent pathway enrichment analysis revealed several significant metabolic pathways in both C18 and HILIC, as shown in Supplementary Table 8 and Figure 2C.

Multiple putative compounds were detected, resulting in the identification of 78 metabolites via C18 and 39 via HILIC. A density-based spatial clustering of applications with noise (DBSCAN) algorithm identified a dominant temporal pattern in participants with SH that encompassed 62.4% of the identified metabolites, containing 48 metabolites from C18 and 25 metabolites from HILIC. The pattern also showed significant metabolite level changes at T1 (2 hours) with a subsequent return to the original level at T2 (72–96 hours). In contrast, participants without SH did not exhibit a clear dominant pattern, as illustrated in Figure 3D.

Metabolomic Changes during Surgery Differentiate Groups with and without Stress Hyperglycemia

To identify metabolites that can distinguish SH versus non-SH participants, both the whole cohort and group difference analyses were subjected to further comparative analysis. Two sample t-tests were performed to compare the changes in metabolite levels between T0 and T1, representing the transition from the baseline to early intraoperative phase (during CABG) in participants with and without SH.

From the whole cohort analysis (without SH as a binary group variable) out of 160 putative metabolites, 22 achieved statistical significance (Figure 4A), with all metabolites showing greater absolute changes (baseline to 2 hours) among participants with SH. Similarly, group difference analysis of 73 metabolites identified 17 statistically significant metabolites (Figure 4B), of which 15 exhibited greater absolute changes in the SH group. Complete metabolite details, including m/z values, retention times, annotations, abbreviation, and clinical relevance, are provided in Tables 1 and 2, respectively. These findings indicate that specific metabolite changes between T0 and T1 may serve as predictive biomarkers for SH development.

Metabolic Signature Predictive of Stress Hyperglycemia

Based on our previous findings, we developed machine learning models to predict SH using these metabolomic changes between T0 and T1 as input features. Model performance, evaluated via LOOCV, is summarized in Supplementary Table 3. The machine learning pipeline with logistic regression achieved the highest performance, with an AUC of 0.857, accuracy of 0.833, sensitivity of 0.893, and specificity of 0.750 (Figure 4C). Figure 4D displays the top 20 putative metabolites predictive of SH ranked by feature importance. One of these metabolites was among the 22 statistically significant metabolites identified in the whole cohort analysis, and three were among the 17 significant metabolites from the group difference analysis. This overlap shows convergence between statistical analyses and machine learning approaches in identifying key metabolic markers. Detailed m/z values, retention times, annotations, and abbreviations are provided in Supplementary Table 9.

Biomarker Analysis and Multilevel Associations with Stress Hyperglycemia

Temporal trends for four biomarkers were assessed using the same NMEM approach applied to the whole cohort statistical analysis for high-sensitivity C-reactive protein (hsCRP), soluble urokinase plasminogen activator receptor (SuPAR), brain-type natriuretic peptide (BNP), and high-sensitivity troponin (hsTnI). All biomarkers exhibited statistically significant temporal patterns, with all p-values less than 0.001. The estimated biomarker trajectories showed an overall increasing trend across the four time points, as shown in Figure 5A. Group difference statistical analysis was subsequently conducted to evaluate whether these temporal patterns differed by SH status. Statistical significance was observed for hsCRP (p<0.001) and SuPAR (p=0.014), while BNP (p=0.239) and hsTnI (p=0.065) showed non-significant trends. For hsCRP, participants with SH exhibited an increase between T0 and T1, while those without SH showed a delayed increase after T1, as shown in Figure 5B. For SuPAR, participants with SH had a steady increase from T0-T2 with a slight decrease at T3 (72–96 hours), whereas those without SH showed an upward trend after T1.

To explore biomarker-metabolite relationships, we mapped metabolomic communities in participants with and without SH. Participants with SH displayed substantially more metabolites associated with the inflammatory biomarkers hsCRP, SuPAR, and BNP compared to those without SH, as shown in Figure 5C. The strength of associations between each biomarker and metabolic pathways revealed distinct patterns between predominantly inflammatory biomarkers and BNP (Figure 5D). SuPAR showed the strongest associations with de novo fatty acid biosynthesis, folate, linoleate, and glycerophospholipids pathways. Amino acid metabolism (branched-chain amino acid degradation and the urea cycle) was strongly associated with BNP. De novo fatty acid biosynthesis, linoleate, amino acid metabolism (glycine, serine, alanine, threonine, glutamate, beta-alanine), and fatty acid activation pathway were most strongly associated with hsCRP.

DISCUSSION

This longitudinal analysis revealed distinct and dynamic metabolic perturbations associated with surgical stress in a well-defined cohort without diabetes undergoing cardiac surgery. By integrating high-resolution metabolomics, continuous glucose monitoring (CGM), and biomarker data, we identified coordinated changes across biochemical pathways linked to inflammation, lipid remodeling, amino acid metabolism, and bile acid signaling, changes that were more pronounced in individuals who developed SH. These findings provide mechanistic insights into the metabolic underpinnings of surgical stress and highlight potential biomarkers that may inform perioperative risk stratification and potential targets for future therapeutic strategies.

Baseline Pathway Differences and Predisposition to Stress Hyperglycemia

Several metabolic pathways were different at baseline in those who developed SH, including Carnitine shuttle, Branched-chain amino acid (BCAA) metabolism, Bile acid biosynthesis, Inflammatory signaling (arachidonic acid metabolism), Microbiome-related metabolism (propanoate metabolism). Many of these pathways have been previously linked to diabetes risk. Their presence preoperatively suggests that individuals predisposed to SH may have underlying metabolic profiles favoring dysglycemia, even before surgical stress. The clinical implications of these preoperative signatures, particularly in patients without diagnosed diabetes, warrant further study in the perioperative and post-discharge periods.

Perioperative Dynamics and Dominant Temporal Patterns

Across temporal analyses (whole cohort, CGM-integrated, and SH-stratified), pathways not evident at baseline, most prominently glycerophospholipid and glycosphingolipid metabolism, emerged after including perioperative timepoints. A characteristic “increased-during-surgery” profile was observed, marked by a sharp rise at T1 followed by partial resolution at T2–T3 Complementary clusters exhibited a “decreased-during-surgery” pattern, with T1 nadirs and rebound thereafter. These coordinated lipid trajectories mirror the perioperative glucose dynamics and are biologically consistent with membrane remodeling and tissue injury/repair. Notably, the amplitude and persistence of these changes were greater in participants who developed SH, with slower normalization and higher trajectories across time. Together, the magnitude and synchrony of these lipid-centered shifts suggest that a subset of patients is more metabolically vulnerable to stress-induced derangements during cardiac surgery.

Key Metabolites and Pathway-Specific Insights

Lipid-Based Inflammatory Signaling.

Multiple lipid-derived inflammatory pathways were dominant postoperatively. Arachidonic acid metabolism, prostaglandin formation, and leukotriene metabolism were upregulated in SH, consistent with prior evidence linking them to inflammation, insulin resistance, and poor glucose control.^3–6^ Sphinganine and sphinganine-1-phosphate, precursors in ceramide biosynthesis, were associated with SH, suggesting ceramide-mediated metabolic disruption as a potential mechanism for perioperative insulin resistance.^7,8^ Free fatty acids (FFAs) may further exacerbate this state by stimulating hepatic glucose production, increasing oxidative stress, and preventing insulin-mediated suppression of glycogenolysis.^9^

Amino Acid and Sulfur Pathways.

Methionine and cysteine metabolism, closely linked to homocysteine regulation, was altered. Elevated homocysteine is a cardiovascular risk factor, and changes post-CABG may reflect altered folate-dependent methylation.^10^ Glycine, alanine, and gluconeogenic amino acids may reflect enhanced hepatic glucose output.^11^ increased release of glucose from the liver has been shown to occur with elevated gluconeogenic amino acids, suggesting one way in which altered amino acid metabolism can lead to SH.^12^ In addition, alanine is a well-known substrate of gluconeogenesis, suggesting that elevated alanine levels may upregulate gluconeogenesis, leading to more glucose production and subsequently greater glucose levels.^12^

Bile Acid Signaling.

Bile acids were associated with SH. While bile acids can have protective metabolic effects via FXR and TGR5 signaling, higher levels in SH may reflect glucose-driven synthesis rather than a protective adaptation.^13,14^

Histidine Metabolism.

Lower levels of 4-imidazolone-5-propanoate in SH suggest increased conversion of histidine to imidazole propionate, a metabolite linked to impaired glucose metabolism.^15^

Biomarker–Metabolite Associations

To integrate biochemical changes with systemic signals, we examined correlations between key biomarkers and SuPAR was associated with de novo fatty acid biosynthesis, glycerophospholipid metabolism, and linoleate, consistent with inflammation and immune activation. SuPAR is linked to diabetes, cardiovascular disease, and adverse renal outcomes.^16,17^ BNP was linked to alterations in amino acid metabolism, particularly branched-chain amino acid degradation and urea cycle pathways, suggesting possible nitrogen handling dysregulation in the context of CABG and CAD. hsCRP had associations similar to SuPAR, reinforcing the central role of inflammation in SH pathophysiology. These biomarkers may serve as accessible surrogates for deeper metabolic changes and help stratify perioperative risk.

Interpretation in Context

The convergence of lipid-derived inflammatory pathways, amino acid dysregulation, and bile acid signaling highlights the multifactorial metabolic stress response to surgery. In SH, this response is amplified, possibly reflecting a combination a) pre-existing metabolic susceptibility (evident at baseline), b) surgical tissue injury and membrane remodeling, and c) inflammatory-hormonal activation (cortisol, catecholamines, cytokines). This metabolic fingerprint may also have implications for post-discharge diabetes risk assessments in susceptible individuals.

Strengths and Limitations

This study integrates metabolomics, continuous glucose monitoring, and biomarker data in a non-diabetic surgical cohort, offering a multidimensional view of metabolic responses to surgical stress. A multi-timepoint design enabled dynamic trajectory analysis, while advanced analytical methods captured coordinated pathway-level changes beyond individual metabolites.

Limitations include a relatively small sample size, which reduced statistical power—particularly for subgroup and pathway-level analyses—and may limit generalizability. Although CGM was calibrated to point-of-care testing, perioperative factors such as temperature, perfusion, and hemodynamic instability may have affected accuracy. Residual confounding from hormonal fluctuations, genetic variability, and differences in surgical severity cannot be excluded; notably, non-SH group surgeries may have been less severe than those in the SH group. Replicating the complexity of human surgical stress in controlled models also remains challenging. Finally, pathway enrichment analyses were exploratory, intended to generate mechanistic hypotheses; confirmation in larger, independent cohorts is needed to establish clinical relevance.

Conclusion

Patients who develop SH after CABG exhibit distinct baseline metabolic signatures, encompassing carnitine shuttle activity, branched-chain amino acid metabolism, bile acid biosynthesis, inflammatory processes, and microbiome-related pathways, that persist or intensify after surgery. Surgical stress further unmasks additional perturbations, particularly in glycerophospholipid and glycosphingolipid metabolism, which may reflect tissue injury. Integration of machine learning with pathway analysis offers a potential framework for precision medicine strategies to predict and mitigate SH. Given the close links between these metabolic stress signatures, inflammation, and insulin resistance, they may also serve as early indicators of postoperative diabetes risk, underscoring the need for validation in larger, independent cohorts.

METHODS

Study Population

Participants comprised patients without diabetes (blood glucose < 126 mg/dL and HbA1c < 6.5%) undergoing CABG surgery at three academic hospitals (Emory University Hospital, Emory University Hospital Midtown, Grady Memorial Hospital) in Atlanta, Georgia. The study was approved by the Institutional Review Board at Emory University (Protocol IRB00097963). All participants provided written informed consent prior to enrollment.

Participants in this study were part of a nested case-control study linked to a clinical trial testing a GLP-1 receptor agonist in the perioperative period (NCT03743025). We collected the following variables during this 1) demographic characteristics (age, sex [male, female], race-ethnicity [White, Black, Hispanic, Others]); 2) clinical information (weight, BMI, smoking status [never smoked], cardiac ejection fraction, HbA1c, blood glucose at consent, APACHE II score, length of stay); 3) surgical information (type of surgery [open, robotic], cardiopulmonary bypass); 4) history of cardiovascular or other diseases (previous cardiac condition, hyperlipidemia, hypertension, infection); 5) medication exposure (vasopressors or inotropes, norepinephrine, vasopressin, dobutamine, milrinone).

Data Collection and Variable Specification

Blood samples were collected at four perioperative time points from T0 to T3. Participants who received dulaglutide before surgery (eight) and those who received steroids during surgery (six) were excluded. Statistical comparisons between groups were performed using the two-sample t-test for continuous variables and chi-squared tests for categorical variables. All data was securely stored in REDCap.

Point-of-Care Testing.

Accu-Chek Inform II glucometer (Emory University Hospital and Emory University Hospital Midtown) and Nova StatStrip (Grady Memorial Hospital) were used to measure glucose from capillary or arterial blood. Point-of-Care (POC) glucose values were checked during the perioperative period, both during and after surgery, based on hospital protocols or anesthesiology discretion.

Continuous Glucose Monitoring.

Factory-calibrated, blinded Dexcom G6-pro CGM devices were placed on participants prior to CABG surgery. The device measures glucose every five minutes for up to 10 days. The device was previously shown to have an overall mean absolute relative difference (MARD) of 12.8% relative to POC capillary glucose previously in inpatient settings.^18^

Biomarkers.

Plasma was collected to measure biomarkers. hsCRP was measured by FirstMark and suPAR was measured using suPARnostic kit from Virogates, Copenhagen, Denmark. Trends between biomarkers, metabolites, and SH were assessed by metabolome-wide association studies (xMWAS) and integrative network analysis.

Metabolomics.

Blood samples were collected at four time points, and high-resolution untargeted metabolomics analysis was performed at the Clinical Biomarkers Laboratory using established protocols with C18 under negative ionization mode (8,631 mass spectrometry features) and HILIC under positive ionization mode (10,919 mass spectrometry features).^19–21^ For data pre-processing, metabolites with zero values were imputed using half the minimum detected value for that metabolite. Metabolite intensities were log-transformed prior to statistical analysis.

Stress Hyperglycemia Definition

SH was defined as meeting either of the following criteria during the period from 4 to 28 hours after the start of surgery (24-hour duration): 1) three or more POCT measurements above 140 mg/dL, or 2) any single POCT measurement above 180 mg/dL. Perioperative insulin administration followed standard institutional protocols.

Statistical Analysis

Baseline Group Difference Analysis.

Baseline differences in metabolite levels between groups were assessed using two-sample t-tests. For each metabolite, a separate test was performed using baseline (T0) values to compare group means, where a significant p-value indicated that the metabolite showed a statistically significant difference at baseline.

Whole Cohort Analysis.

Temporal changes in metabolite levels were investigated using NMEM.^22^ For each metabolite, two alternative models were compared to determine whether the metabolite exhibited significant temporal variation (full model) or remained stable over time (null model). Specifically, the full model took the following yij=ftj+bi+ϵij, where yij denotes the metabolite level for participant i at time point j,ftj is a non-linear function of time modeled using natural splines with three degrees of freedom (excluding the intercept), biN0,σb2, is a participant-specific random intercept, and ϵijN0,σ2, is the residual error. In contrast, the null model assumes no temporal change of the metabolite by replacing ftj with a constant intercept β0. The full and null models were compared using likelihood ratio tests, where a significant p-value indicated that the metabolite exhibited a statistically significant temporal change.

Group or Clinical Difference Analysis.

To assess whether temporal trajectories of metabolite levels were modified by a variable of interest Zi (binary group or continuous clinical measure), a NMEM was extended to include an interaction between time and Zi. The full model takes the following yij=ftj+gtj⋅Zi+bi+ϵij, where gtj is a non-linear function of time modeled using natural splines and gtj⋅Zi captures the interaction, allowing the time trajectory to vary by Zi. In contrast, the null model is excluded in this interaction. Likelihood ratio tests were used to compare models, with significant p-values indicating that temporal patterns differed by Zi.

Pathway Enrichment Analysis.

To identify biologically meaningful pathways associated with temporal changes in metabolite levels, pathway enrichment analysis was performed using Mummichog version 2.7.0.^23,24^ Separate analyses were conducted for C18 and HILIC datasets, with metabolites analyzed alongside their corresponding p-values derived from NMEM. The p-value cutoff was set to 0.05, and 1,000 permutations were used to assess statistical significance. The identified significant pathways provided insights into the underlying metabolic processes associated with the observed temporal metabolite dynamics.

To uncover common temporal trends among metabolites within significant biological pathways, we analyzed the subset of metabolites included in significant pathways identified through pathway enrichment analysis. Metabolites from both C18 and HILIC datasets were combined, and temporal trajectories were derived from NMEM-fitted values using the full model. These trajectories were clustered using the DBSCAN algorithm,^25^ with similarity defined as one minus the Pearson correlation coefficient between metabolite trajectories. The resulting clusters revealed dominant temporal patterns, reflecting groups of metabolites that exhibited similar dynamics over time.

Machine Learning Pipeline

A participant-level classifier was developed to predict SH development using metabolomic data. The initial feature set comprised empirical metabolites identified through Mummichog pathway enrichment 1,455 putative metabolites from C18 and 1,320 from HILIC (2,775 total metabolites). Model performance was evaluated using leave-one-out cross-validation (LOOCV) based on area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, and specificity.

For each LOOCV training set, recursive feature elimination (RFE) was employed,^26–29^ which iteratively removed 10% of features with the lowest RFE importance scores until 25 features remained. RFE importance scores were determined using ridge logistic regression coefficients to avoid overfitting. Partial least squares (PLS) regression was subsequently applied to derive a single composite feature, with age, sex, and BMI included as additional covariates, resulting in four predictors. Four classification algorithms were then evaluated based on these Ridge logistic regression, support vector machine (SVM) classifier, Random Forest, and Multilayer perceptron (MLP) classifier.

Feature importance for interpretation was calculated within the LOOCV framework. In each iteration, logistic coefficients were multiplied by the corresponding PLS coefficients to account for magnitude and direction. The absolute values of these combined coefficients were averaged across all LOOCV iterations, and the top 20 metabolites were selected based on these averaged absolute importance values.

Integrative Network Analysis of Clinical and Metabolic Data

To investigate metabolic pathways associated with SH development, data-driven integrative analysis was performed using xMWAS.^30^ Eight clinical covariates at T0 and T1 (age, sex, BMI, SuPAR, BNP, hsTnI, hsCRP, and SH status) were integrated with corresponding metabolic profiles using PLS canonical regression. A multilevel community detection algorithm identified connected groups of variables that exhibited strong intra-community connections while maintaining sparse connections with the broader network. Network edges were retained based on threshold criteria of the absolute correlation coefficient greater than 0.6 and p-value less than 0.05 as determined by Student’s t-test.

Manual Annotation of Metabolites

For metabolites not annotated by Mummichog, manual annotation was performed using the Human Metabolome Database (HMDB) version 5. m/z ratios were queried using the liquid chromatography-mass spectrometry (LC-MS) search feature with separate searches for positive and negative ionization modes. For positive mode, the following adducts were [M+H]^+^, [M+ACN+H]^+^, [M+ACN+Na]^+^, with additionally consideration of [M+NH4]^+^, [M+Na]^+^, [M+H−H2O]^+^, [M+K]^+^, and [M+H+Na]^+^. For negative mode, [M−H]^−^, [M−H2O−H]^−^, and [M+Cl]^−^ adducts were included. All searches employed a five-ppm mass tolerance between measured and theoretical m/z values. Annotation assignment followed a hierarchical prioritization scheme. First, molecules not endogenously produced by human or microbiome metabolism were given lower priority to focus on biologically relevant metabolites. Second, annotations with the smallest mass deviation (lowest delta m/z) received highest priority. Lastly, biological relevance and isotopic consistency of the same chemical formula were used to assign priority at the discretion of the researchers.

CGM Imputation and Calibration

To address the challenges of missing or inaccurate values in CGM data during the perioperative period,^31^ we developed an imputation and calibration protocol that integrates CGM and POCT measurements, which includes three steps.

Step Time Alignment.

Prior studies have reported a temporal lag between POCT and CGM, with POCT measurements typically preceding CGM readings.^32–34^ To account for this lag, Pearson correlations between CGM and POCT were calculated using time shifts ranging from 5 to 25 minutes. The time shift with the highest correlation was selected as the optimal lag for each participant. The CGM and POCT time series were then aligned based on this optimal lag.

Step Imputation.

Unlike conventional imputation methods that rely solely on observed CGM values to estimate missing data, our approach leveraged the availability of paired POCT and CGM measurements to improve imputation quality. Specifically, for time points where both CGM and POCT readings are available, we modeled their relationship using local linear regression and used this model to predict missing CGM values from corresponding POCT measurements. This enabled imputation based on a clinically verified glucose reference rather than simply relying on internal CGM trends. Specifically, at each time point t with a missing CGM value and a corresponding POCT, a local window Wt was defined, which consisted of nine nearby time points with available CGM-POCT pairs. Within this window, the following local regression was yCGM(s)=β0+β1⋅POCT(s)+ϵs,s∈Wt, where yCGM(s) and POCT(s) denote the CGM and POCT values at time s, respectively, and ϵs is the error term. This model allowed for the local and non-stationary relationship between CGM and POCT to be captured. To estimate β0 and β1, we proposed to solve a constrained optimization problem as minβ0,β1∑s∈WtlδyCGM(s)-β0-β1⋅POCT(s)subjecttoβ1≥0, where lδ(⋅) is the Huber loss function between the true and predicted CGM values.^35^ This loss function was chosen for its robustness to outliers, helping to mitigate the impact of occasional noise or unreliability in CGM readings. The non-negativity constraint on β1 reflects the expected positive association between CGM and POCT values. After solving the optimization, the estimated βˆ0 and βˆ1 were used to impute the missing CGM value at time t using the corresponding POCT value. For remaining CGM time points without paired POCT data, linear interpolation between neighboring CGM values was used to complete the time series.^36–38^

Step Calibration.

To further improve the accuracy of CGM data, we developed a calibration step to account for time-varying sensor behavior, such as signal drift or adaptation effects.^39,40^ For this, a separate local linear regression model was fitted to predict POCT values using CGM readings (both observed and imputed from Step 2) and the time since sensor insertion. Specifically, the following regression was POCT(s)=γ0+γ1⋅yCGM(s)+γ2⋅logTinsert(s)+ϵs,s∈Wt, where Tinsert(s) denotes the time since sensor insertion at time s and the logarithmic transformation was used to capture non-linear behavior related to sensor adaptation over time. To estimate γ0,γ1, and γ2, we proposed to solve a similar constrained optimization as in Step 2 as shown minγ0,γ1,γ2∑s∈WtlδPOCT(s)-γ0-γ1⋅yCGM(s)-γ2⋅logTinsert(s)subjecttoγ1≥0,

The estimated γˆ0,γˆ1, and γˆ2 were utilized to predict the POCT values, which served as calibrated CGM readings. This method adjusts CGM readings by accounting for both their instantaneous relationship with POCT and time-dependent sensor drift, enhancing calibration accuracy.

Supplementary Material