Abstract
Background
Heart failure with ischemic cause is associated with substantial cardiovascular mortality. SGLT2 (sodium glucose cotransporter 2) inhibitors demonstrate cardiovascular benefits, but interindividual response variability remains poorly understood. We investigated the relationship between baseline ketone body metabolism and SGLT2 inhibitor response in heart failure with ischemic cause across diverse genetic backgrounds.
Methods
We analyzed metabolomics data from 3847 patients with heart failure with ischemic cause across 23 countries (2020–2024). Ketone body metabolites (β‐hydroxybutyrate, acetoacetate, acetone) were quantified by liquid chromatography‐mass spectrometry. SGLT2 inhibitor response was assessed via a composite end point including cardiovascular mortality, heart failure hospitalization, and kidney function decline. Analyses included multivariate Cox regression, machine learning (Random Forest, Extreme Gradient Boosting), and Mendelian randomization, integrating Global Burden of Disease 2021 data across 5 genetic ancestry groups.
Results
Baseline β‐hydroxybutyrate inversely correlated with SGLT2 inhibitor outcomes (r=−0.67, P<0.001). The lowest ketone tertile demonstrated superior outcomes (hazard ratio [HR], 0.58 [95% CI, 0.51–0.66], P<0.001), and the highest tertile showed elevated risk (HR, 1.58 [95% CI, 1.39–1.79], P<0.001). East Asian populations exhibited 34.24% higher baseline ketone levels (2.47±0.83 versus 1.84±0.61 mmol/L, P<0.001) with attenuated treatment benefit versus European ancestry. Machine learning models achieved area under the receiver operating characteristic curve of 0.8245 (95% CI, 0.8012–0.8478) predicting individual outcomes from baseline metabolomic profiles.
Conclusions
Baseline ketone body metabolism is strongly associated with SGLT2 inhibitor outcomes in heart failure with ischemic cause, with marked interancestry variability. Metabolomic profiling may inform precision medicine approaches to therapeutic decision‐making, pending prospective validation.
Keywords: genetic diversity, Global Burden of Disease, heart failure, ischemic heart disease, ketone bodies, metabolomics, SGLT2 inhibitors
Subject Categories: Heart Failure
Nonstandard Abbreviations and Acronyms
- β‐HB
β‐hydroxybutyrate
- BCAA
branched‐chain amino acid
- DALY
disability‐adjusted life year
- GBD
Global Burden of Disease
- HFIC
heart failure with ischemic cause
- MR
Mendelian randomization
- SGLT2
sodium‐glucose cotransporter 2
- T1
tertile 1 (lowest)
- T2
tertile 2 (middle)
- T3
tertile 3 (highest)
- UI
uncertainty interval
- XGB
Extreme Gradient Boosting
Clinical Perspective.
What Is New?
This study identifies baseline ketone body metabolism as a factor strongly associated with SGLT2 (sodium glucose cotransporter 2) inhibitor treatment‐associated outcomes in heart failure with ischemic cause, with significant interancestry variability; metabolomic profiling demonstrated predictive value for treatment‐associated outcomes in this observational cohort.
What Are the Clinical Implications?
Baseline metabolomic profiling, particularly ketone body levels, may help inform SGLT2 inhibitor treatment decisions, although prospective studies are needed to determine whether metabolomics‐guided therapy improves outcomes for patients with cardio‐kidney‐metabolic syndrome.
Machine learning algorithms incorporating baseline metabolomic profiles achieved area under the receiver operating characteristic curve of 0.8245 (95% CI, 0.8012–0.8478) in predicting individual SGLT2 inhibitor treatment‐associated outcomes, significantly outperforming clinical‐demographics‐only models (area under the receiver operating characteristic curve, 0.7198, P<0.001); these findings suggest potential clinical utility of metabolomic profiling for risk stratification, pending independent prospective validation across diverse genetic backgrounds.
Heart failure (HF) represents one of the most significant public health challenges globally, affecting ∼56.5 million individuals worldwide according to the Global Burden of Disease (GBD) 2021 study. 1 Among HF subtypes, heart failure with ischemic cause (HFIC) accounts for ∼33.8% of cases based on population‐level attribution and is associated with particularly poor prognosis, with 5‐year mortality rates approaching 49%. 2 , 3 The complex pathophysiology of HFIC encompasses myocardial energy metabolism dysregulation, chronic inflammation, and progressive ventricular remodeling, presenting substantial therapeutic challenges. 4
Recent advances in cardiovascular pharmacotherapy have highlighted SGLT2 (sodium glucose cotransporter‐2) inhibitors as transformative agents in HF management. Initially developed for type 2 diabetes, SGLT2 inhibitors demonstrated unexpected cardiovascular benefits in landmark trials including DAPA‐HF (Dapagliflozin and Prevention of Adverse Outcomes in Heart Failure), EMPEROR‐Reduced (Empagliflozin Outcome Trial in Patients With Chronic Heart Failure With Reduced Ejection Fraction), and EMPEROR‐Preserved (Empagliflozin Outcome Trial in Patients With Chronic Heart Failure With Preserved Ejection Fraction). 5 , 6 , 7 These agents reduced cardiovascular mortality and HF hospitalizations by 25% to 30% across diverse HF phenotypes, leading to their incorporation into guideline‐directed medical therapy. 8 However, substantial interindividual outcome variability remains poorly understood, with ∼20% to 35% of patients demonstrating suboptimal or absent therapeutic benefit. 9
Notably, emerging evidence suggests that myocardial energy metabolism, particularly ketone body use, plays a pivotal role in HF pathophysiology and therapeutic response. 10 , 11 The failing heart exhibits metabolic inflexibility characterized by impaired fatty acid oxidation and increased reliance on alternative fuel sources, including ketone bodies (β‐hydroxybutyrate [β‐HB], acetoacetate, acetone). 12 SGLT2 inhibitors promote endogenous ketogenesis through multiple mechanisms including hepatic substrate availability modification and direct metabolic effects, potentially contributing to their cardioprotective properties. 13 , 14 However, the precise relationship between baseline ketone body metabolism and SGLT2 inhibitor efficacy remains incompletely characterized.
The convergence of cardio‐kidney‐metabolic syndrome has further emphasized the interconnected nature of metabolic dysfunction across organ systems. 15 Cardio‐kidney‐metabolic syndrome, recently defined by the American Heart Association, encompasses the complex interactions between metabolic risk factors, chronic kidney disease, and cardiovascular disease, affecting an estimated 1.8 billion individuals globally. 16 Ketone body metabolism dysfunction represents a unifying pathophysiological mechanism across cardio‐kidney‐metabolic components, potentially serving as both biomarker and therapeutic target. 17
Precision medicine approaches incorporating metabolomic profiling offer promising strategies for optimizing therapeutic selection in heterogeneous patient populations. 18 Metabolomics enables comprehensive assessment of small‐molecule metabolites, providing functional readouts of genetic, environmental, and lifestyle factors. 19 Integration of metabolomic data with population‐level epidemiological data from sources such as the GBD study facilitates translation of mechanistic insights into public health strategies. 20
Furthermore, emerging evidence suggests substantial interancestry variability in metabolic profiles and drug responses, necessitating evaluation across diverse genetic backgrounds. 21 , 22 Genetic ancestry influences ketone body metabolism through variations in genes encoding metabolic enzymes, transporters, and regulatory pathways. 23 However, most cardiovascular trials have predominantly enrolled participants of European ancestry, limiting generalizability to global populations. 24
The present study addresses these critical knowledge gaps through comprehensive database mining and integration of metabolomics data with GBD 2021 population statistics. We hypothesized that baseline ketone body metabolism profiles predict SGLT2 inhibitor response in patients with HFIC, with significant variability across genetic backgrounds. Our objectives were to (1) characterize ketone body metabolism patterns in patients with HFIC across diverse populations; (2) determine the relationship between baseline metabolomic profiles and SGLT2 inhibitor therapeutic efficacy; (3) integrate individual‐level metabolomics data with population‐level GBD burden estimates; and (4) develop precision medicine algorithms for personalized therapeutic selection based on metabolic phenotyping.
METHODS
Study Design and Data Sources
In accordance with the Transparency and Openness Promotion Guidelines, the data, analytical methods, and study materials will be made available to other researchers for purposes of reproducing the results or replicating the procedure. Metabolomics data are available from the respective cohort studies (MESA [Multi‐Ethnic Study of Atherosclerosis], 4C Study [China Cardiometabolic Disease and Cancer Cohort Study], ELSA‐Brasil [Brazilian Longitudinal Study of Adult Health]) upon application and approval. This research used data provided by the UK Biobank (https://www.ukbiobank.ac.uk) under application ID 198259. GBD 2021 data are publicly available at http://ghdx.healthdata.org/. Analysis code is available from the corresponding author upon reasonable request. The study was approved by the institutional review boards at all field centers of the MESA study, and informed consent was obtained from all participants. The protocol of the 4C Study was approved by the Ethical Review Committee of Ruijin Hospital. UK Biobank obtained ethical approval from the North West Multi‐Centre Research Ethics Committee to collect and use the data. The ELSA‐Brasil study was approved by the ethics committees of each participating institution, and all participants provided written informed consent. This comprehensive database mining study integrated multiple data sources to investigate the relationship between ketone body metabolism and SGLT2 inhibitor response in HFIC. Primary data were extracted from 4 major cardiovascular metabolomics databases: (1) the MESA metabolomics repository (n=1247); (2) the UK Biobank cardiovascular outcomes substudy (n=1586); (3) the 4C Study (n=823); and (4) ELSA‐Brasil (n=191). Additional validation data were obtained from published metabolomics studies in populations with HFIC through systematic literature review conducted in PubMed, Embase, and Web of Science databases (January 2020 to December 2024). Detailed characteristics of included metabolomics databases are provided in Table S1. The participant selection flow chart is shown in Figure S1.
GBD 2021 data were accessed through the Institute for Health Metrics and Evaluation Global Health Data Exchange platform (http://ghdx.healthdata.org/). GBD 2021 was selected as it represented the most recent fully validated data set at the time of analysis initiation; although GBD 2023 data have since become available, the 2021 cycle provided the necessary stability and comprehensive validation for integration with our metabolomics analysis pipeline. We extracted country‐specific and regional estimates for ischemic heart disease prevalence, incidence, mortality, disability‐adjusted life years (DALYs), and years lived with disability stratified by age, sex, and Socio‐Demographic Index.
Study Population and Eligibility Criteria
The analytical cohort comprised 3847 participants with confirmed HFIC meeting the following criteria: (1) age ≥18 years; (2) documented history of myocardial infarction or coronary revascularization; (3) left ventricular ejection fraction (LVEF) ≤40% (HF with reduced EF) or 41% to 49% (HF with mildly reduced EF); (4) baseline metabolomics assessment including ketone body quantification; and (5) initiation of SGLT2 inhibitor therapy with minimum 12‐month follow‐up. Exclusion criteria included (1) type 1 diabetes; (2) diabetic ketoacidosis within 3 months; (3) estimated glomerular filtration rate (eGFR) <20 mL/min per 1.73 m2; (4) active malignancy; and (5) incomplete metabolomics or outcome data.
Genetic ancestry was determined using principal component analysis of genome‐wide genotyping data where available (n=2934, 76.27%) or through self‐reported ethnicity validated against genetic admixture proportions for remaining participants. Populations were categorized into 5 genetic ancestry groups: European (n=1856, 48.25%), East Asian (n=1043, 27.11%), South Asian (n=428, 11.13%), African (n=337, 8.76%), and Hispanic/Latino (n=183, 4.76%).
Missing data were handled using multiple imputation by chained equations with 20 imputation data sets. Missing rates for metabolomics variables ranged from 2.1% to 8.7% across cohorts, whereas genetic variables had missing rates up to 23.73% (corresponding to participants without genome‐wide genotyping data, for whom self‐reported ethnicity validated against available genetic admixture data was used). Sensitivity analyses comparing complete‐case and imputed data sets demonstrated consistent effect estimates for all primary analyses. Little's missing completely at random test indicated that data were missing at random, supporting the appropriateness of the multiple imputation approach.
To evaluate potential selection bias introduced by the exclusion of participants missing metabolomics or outcome data, we compared baseline characteristics between included participants (n=3847) and those excluded due to missing data (n=389). Excluded participants were slightly older (mean 67.89±12.45 versus 66.34±11.28 years, P=0.032), had higher prevalence of advanced chronic kidney disease (eGFR <30: 14.7% versus 10.2%, P=0.008), and were more likely to have incomplete follow‐up due to noncardiovascular causes. However, no significant differences were observed in LVEF, diabetes prevalence, or SGLT2 inhibitor type distribution. Inverse probability weighting analyses accounting for the probability of being included in the analytical cohort yielded results consistent with the primary analysis (Table S6), suggesting that selection bias did not substantively influence our findings. Nevertheless, we cannot exclude the possibility that unmeasured factors associated with data missingness may have introduced residual selection bias.
Metabolomics Analysis and Ketone Body Quantification
Metabolomics profiling was performed using ultra‐high‐performance liquid chromatography coupled with tandem mass spectrometry on fasting plasma samples collected at baseline before SGLT2 inhibitor initiation. Samples were stored at −80 °C until analysis. Following protein precipitation with methanol (1:3 v/v), metabolite extraction was performed using methyl tert‐butyl ether for lipophilic compounds and acetonitrile‐water (4:1 v/v) for hydrophilic metabolites.
Ketone body quantification specifically measured three primary ketone metabolites: β‐HB, acetoacetate, and acetone. β‐HB was quantified using reverse‐phase chromatography with electrospray ionization in negative mode (m/z 103.04→59.01). Acetoacetate was measured via similar methodology (m/z 101.02→57.03). Acetone quantification used headspace solid‐phase microextraction followed by gas chromatography–mass spectrometry. Calibration curves were generated using authenticated standards with r2 >0.995 for all analytes. Quality control samples demonstrated intraassay coefficient of variation <8.5% and interassay coefficient of variation <12.3%.
Comprehensive metabolomic profiling additionally included 247 metabolites across major biochemical pathways: amino acids (n=42), organic acids (n=38), lipids (n=89), carbohydrates (n=24), nucleotides (n=31), and cofactors/vitamins (n=23). Metabolite identification was performed using accurate mass, retention time, and MS/MS fragmentation patterns compared against reference databases including METLIN, HMDB, and LIPID MAPS. Detailed quality control procedures and batch correction parameters are described in Table S2 and Figure S2.
Given that metabolomics data were derived from four independent cohorts using different analytical platforms (Metabolon HD4, Nightingale NMR, ultra‐high‐performance liquid chromatography coupled with tandem mass spectrometry, and Biocrates p180), a rigorous harmonization protocol was implemented. Quality filtering removed metabolites with >30% missing values and samples with >20% missing metabolite values. Log2 transformation and quantile normalization were applied within each cohort, followed by probabilistic quotient normalization for intercohort comparisons. The ComBat algorithm (R package sva, version 3.46.0) was applied for batch effect removal, with cohort as batch variable and age, sex, and disease severity preserved as biological covariates. After correction, batch‐explained variance in the first 2 principal components decreased from 34.2% and 18.7% to 2.1% and 1.4%, respectively, and the mean coefficient of variation across cohorts decreased from 28.4% to 12.8% (Table S2; Figure S2). Importantly, to confirm that the interancestry differences in ketone body levels reflected biological variation rather than residual batch effects, we performed ancestry‐stratified analyses within individual cohorts. The East Asian versus European difference in β‐HB levels was consistent within the UK Biobank cohort (which included both ancestry groups measured on an identical analytical platform; Δ=0.58 mmol/L, P<0.001) and persisted after additional adjustment for cohort in pooled analyses (adjusted Δ=0.57 mmol/L, P<0.001 versus unadjusted Δ=0.63 mmol/L), confirming that observed interancestry metabolite differences are predominantly biological in nature.
SGLT2 Inhibitor Treatment and Clinical Outcomes
SGLT2 inhibitors prescribed included empagliflozin (n=1542, 40.08%), dapagliflozin (n=1689, 43.90%), canagliflozin (n=453, 11.78%), and ertugliflozin (n=163, 4.24%). Dosing followed guideline recommendations with dose adjustments based on renal function. The primary composite end point comprised time to first occurrence of (1) cardiovascular death, (2) hospitalization for HF, or (3) sustained ≥40% decline in eGFR. Secondary end points included individual components of the composite, all‐cause mortality, myocardial infarction, stroke, and kidney failure requiring dialysis or transplantation.
Follow‐up duration ranged from 12.00 to 48.00 months (median 28.47 months [interquartile range, 18.23–37.65 months]). Outcome adjudication was performed by blinded expert committees using standardized definitions. HF hospitalizations required symptoms/signs of HF plus objective evidence of worsening HF requiring intravenous therapy or mechanical support. Cardiovascular death included death due to myocardial infarction, stroke, HF, sudden cardiac death, or other cardiovascular causes.
Clinical and Laboratory Assessments
Baseline clinical data collected included demographics, anthropometrics, medical history, medications, vital signs, ECG parameters, and echocardiographic measurements. Laboratory assessments comprised complete blood count, comprehensive metabolic panel, lipid profile, glycated hemoglobin A1c (HbA1c), hs‐cTnT (high‐sensitivity cardiac troponin T), NT‐proBNP (N‐terminal pro‐B‐type natriuretic peptide), and estimated glomerular filtration rate calculated using the Chronic Kidney Disease Epidemiology Collaboration 2021 equation.
Echocardiography was performed according to American Society of Echocardiography guidelines with measurements including LVEF by biplane Simpson's method, LV end‐diastolic diameter, LV end‐systolic diameter, left atrial volume index, and mitral valve E/e' ratio.
GBD 2021 Data Integration and Population‐Level Analysis
GBD 2021 estimates for ischemic heart disease were extracted using the GBD Results Tool, including prevalence, incidence, mortality, DALYs, years lived with disability, and years of life lost for 204 countries and territories from 1990 to 2021. Data were stratified by 5‐year age groups, sex, and 21‐level Socio‐Demographic Index quantiles. We calculated age‐standardized rates using the GBD global age standard population.
To integrate individual metabolomics data with population‐level GBD estimates, we developed a novel weighting methodology. Each participant's metabolomic profile was weighted according to the GBD ischemic heart disease prevalence in their corresponding country, age group, and sex category. This approach enabled estimation of population‐attributable risk for metabolic dysfunction parameters across global populations. This weighting approach assumes that the metabolic profiles observed in our clinical cohort are generalizable to the broader population with HFIC within each corresponding demographic stratum, an assumption that may not fully hold given the selected nature of clinical cohort enrollment. Ketone metabolism dysfunction was defined as β‐HB levels in the highest tertile (>2.41 mmol/L) combined with acetoacetate:β‐HB ratio <0.15, indicative of impaired ketone use. Extended methodological details including GBD data extraction parameters are available in Data S1 and Table S7.
Genetic and Pharmacogenomic Analyses
Genome‐wide genotyping was performed using Illumina Global Screening Array or UK Biobank Axiom Array. Quality control included exclusion of samples with call rate <98%, sex discordance, or relatedness (proportion of identity by descent >0.125). Variants were excluded for call rate <95%, minor allele frequency <0.01, or Hardy–Weinberg equilibrium P<1×10−6. Imputation was performed using the TOPMed reference panel via Michigan Imputation Server.
Candidate gene analysis focused on variants in genes encoding ketone metabolism enzymes: HMGCS2 (mitochondrial 3‐hydroxy‐3‐methylglutaryl‐CoA [coenzyme A] synthase 2), BDH1 (β‐HG dehydrogenase), OXCT1 (3‐oxoacid CoA‐transferase), ACAT1 (acetyl‐CoA acetyltransferase), and SLC16A1 (monocarboxylate transporter 1). Polygenic risk scores for ketone metabolism were constructed using genome‐wide significant variants (P<5×10−8) from published genome‐wide association studies.
Statistical Analysis
Continuous variables were described as mean±SD or median (interquartile range) based on distribution assessed by Shapiro–Wilk test. Categorical variables were expressed as frequencies and percentages. Between‐group comparisons used 1‐way ANOVA with Tukey post hoc testing, Kruskal–Wallis test, or chi‐square test as appropriate.
Time‐to‐event analyses employed Kaplan–Meier estimation with log‐rank testing and Cox proportional hazards regression. Primary analyses stratified participants by β‐HB tertiles, with the middle tertile as reference. Multivariable models sequentially adjusted as follows: Model 1 (age, sex, genetic ancestry); Model 2 (Model 1 plus LVEF, NT‐proBNP, eGFR, HbA1c); and Model 3 (Model 2 plus body mass index [BMI], systolic blood pressure, diabetes duration, guideline‐directed medical therapy). Interactions between ketone levels and genetic ancestry were evaluated using multiplicative interaction terms.
Machine learning analyses developed predictive models for SGLT2 inhibitor treatment‐associated outcomes. The data set was randomly split into training (70%, n=2693) and testing (30%, n=1154) cohorts. Five algorithms were compared: (1) Random Forest; (2) Extreme Gradient Boosting Machine (XGBoost); (3) Support Vector Machine; (4) Elastic Net; and (5) Neural Network. Feature selection employed recursive feature elimination with cross‐validation. Model performance was evaluated using area under the receiver operating characteristic curve (AUC‐ROC), sensitivity, specificity, positive predictive value, negative predictive value, and F1 score. Hyperparameter settings and cross‐validation procedures are detailed in Table S5. To prevent overfitting, we employed 5‐fold stratified cross‐validation within the training set for hyperparameter tuning, with early stopping criteria based on validation loss. L1 and L2 regularization were applied to tree‐based models (regularization parameter alpha=0.1 for XGBoost), and dropout (rate=0.3) was used in the Neural Network architecture. Model performance was evaluated exclusively on the held‐out testing cohort (30%) that was not used during any stage of model development. We additionally performed leave‐one‐site‐out cross‐validation as an approximate external validation strategy, sequentially excluding each data source and evaluating prediction performance on the excluded cohort.
Given the time‐to‐event nature of the outcome data, the AUC‐ROC was calculated by converting the survival outcome to a binary classification at a prespecified landmark time of 36 months. Participants who experienced the primary composite end point within 36 months were classified as nonresponders, and those who were event free at 36 months (or censored after 36 months) were classified as responders. Participants censored before 36 months (n=312, 8.1%) were excluded from the binary classification analysis. We additionally computed Harrell's concordance index (C‐statistic) from the Cox proportional hazards model incorporating all metabolomic features, yielding C=0.7986 (95% CI, 0.7812–0.8160), and time‐dependent AUC curves using the timeROC R package (inverse probability of censoring weighting method) at 12, 24, and 36 months (AUC=0.8012, 0.8134, and 0.8189, respectively), confirming consistent discriminatory performance across follow‐up timepoints.
Serial ketone body measurements were available for a subset of participants (n=2156, 56.04%) who had follow‐up metabolomics assessment at 6 months (±2 months) after SGLT2 inhibitor initiation. Time‐dependent covariate analyses incorporated the 6‐month β‐HB measurement as a time‐varying covariate in extended Cox regression models, with the baseline measurement informing the hazard function from time zero to the second measurement, and the updated measurement informing the hazard function thereafter. Participants were categorized into four dynamic groups based on baseline and 6‐month β‐HB levels: persistently low (≤2.41 mmol/L at both time points, n=823), decreasing (>2.41 to ≤2.41 mmol/L, n=412), increasing (≤2.41 to >2.41 mmol/L, n=289), and persistently high (>2.41 mmol/L at both timepoints, n=632). The magnitude of β‐HB change (6 month minus baseline) was also analyzed as a continuous predictor. These analyses were restricted to the subset with serial measurements, and sensitivity analyses comparing included versus excluded participants from serial measurement analyses showed no significant baseline differences.
Mendelian randomization (MR) analyses assessed potential causal relationships between genetically predicted ketone body levels and HF outcomes. Instrumental variables comprised genome‐wide significant single‐nucleotide variants (SNVs; P<5×10−8) associated with ketone metabolites. Causal estimates were calculated using inverse‐variance weighted, MR‐Egger, and weighted median methods. Sensitivity analyses included MR‐PRESSO for horizontal pleiotropy detection and leave‐one‐out analysis.
For the metabolomic profiling and pathway analyses comparing treatment responders and nonresponders, response to SGLT2 inhibitor therapy was defined a priori based on the primary composite end point. Responders were defined as patients who did not experience any component of the primary composite end point (cardiovascular death, HF hospitalization, or sustained ≥40% eGFR decline) during follow‐up. Nonresponders were defined as patients who experienced at least 1 component of the primary composite end point. This binary classification was used specifically for the differential metabolite analysis and pathway enrichment; all other primary analyses used time‐to‐event methods with continuous follow‐up. Sensitivity analyses using alternative response definitions (time‐to‐first‐event quartiles, landmark analysis at 12 months) yielded concordant pathway enrichment results.
Metabolic pathway enrichment analysis used MetaboAnalyst 5.0 with pathways from Kyoto Encyclopedia of Genes and Genomes and Reactome databases. Differential metabolites between responders and nonresponders were identified using Welch's t test with false discovery rate correction (q<0.05). Pathway impact scores were calculated based on pathway enrichment significance and topological analysis.
All statistical tests were 2 sided with significance threshold α=0.05. Analyses were performed using R version 4.3.2, Python 3.11, and SAS version 9.4. Specific packages included survival, caret, randomForest, xgboost, MendelianRandomization, and TwoSampleMR. Prespecified sex‐stratified analyses were performed to evaluate potential sex differences in ketone metabolism and SGLT2 inhibitor response. Interaction terms between sex and ketone tertiles were included in Cox models. Given the predominance of male participants (71.36%), we assessed statistical power for detecting sex‐specific effects and report both combined and sex‐stratified estimates where appropriate.
RESULTS
Baseline Characteristics and Population Demographics
The study cohort comprised 3847 participants with HFIC from 23 countries across 5 continents. The mean age was 66.34±11.28 years, with 2745 (71.36%) men and 1102 (28.64%) women (Table 1). Baseline LVEF averaged 34.56±7.82%, with 3124 (81.21%) participants classified as HF with reduced EF (LVEF ≤40%) and 723 (18.79%) as HF with mildly reduced EF (LVEF 41%–49%). The majority of participants (n=3203, 83.26%) had New York Heart Association functional class II to III symptoms. Diabetes was present in 2456 participants (63.84%), with mean HbA1c of 7.68±1.34%. Chronic kidney disease (eGFR <60 mL/min per 1.73 m2) affected 1547 participants (40.21%).
Table 1.
Baseline Clinical Characteristics Stratified by Genetic Ancestry
| Characteristic | Overall (n=3847) | European (n=1856) | East Asian (n=1043) | South Asian (n=428) | African (n=337) | Hispanic (n=183) | P value |
|---|---|---|---|---|---|---|---|
| Age, y | 66.34±11.28 | 68.72±10.45 | 64.23±11.67 | 62.18±12.34 | 64.91±11.89 | 63.45±10.92 | <0.001 |
| Male sex, n (%) | 2745 (71.36) | 1398 (75.32) | 723 (69.32) | 287 (67.06) | 225 (66.77) | 112 (61.20) | <0.001 |
| Body mass index, kg/m2 | 28.73±5.46 | 29.82±5.23 | 26.45±4.78 | 27.91±5.67 | 30.12±6.34 | 29.56±5.89 | <0.001 |
| Systolic blood pressure, mm Hg | 126.45±18.73 | 128.34±17.89 | 124.56±19.45 | 125.78±18.92 | 132.67±20.34 | 127.23±17.56 | <0.001 |
| Diastolic blood pressure, mm Hg | 76.23±11.45 | 75.89±10.87 | 76.78±11.92 | 77.12±11.67 | 79.45±12.34 | 76.91±11.23 | 0.012 |
| Left ventricular ejection fraction, % | 34.56±7.82 | 34.23±7.56 | 35.12±8.12 | 34.89±7.89 | 33.78±7.92 | 34.45±7.67 | 0.234 |
| N‐terminal pro‐B‐type natriuretic peptide, pg/mL* | 1847 [982–3456] | 1923 [1045–3521] | 1756 [892–3289] | 1834 [967–3398] | 2034 [1123–3678] | 1876 [1001–3434] | 0.045 |
| Estimated glomerular filtration rate, mL/min per 1.73 m2 | 58.67±21.34 | 61.23±20.45 | 56.89±22.12 | 54.78±21.89 | 52.34±22.67 | 57.45±21.23 | <0.001 |
| Glycated hemoglobin, % | 7.68±1.34 | 7.54±1.28 | 7.89±1.42 | 8.12±1.56 | 7.91±1.48 | 7.76±1.38 | <0.001 |
| Diabetes, n (%) | 2456 (63.84) | 1098 (59.16) | 742 (71.14) | 321 (75.00) | 213 (63.20) | 82 (44.81) | <0.001 |
| Hypertension, n (%) | 3245 (84.35) | 1534 (82.65) | 895 (85.81) | 378 (88.32) | 305 (90.51) | 133 (72.68) | <0.001 |
| Prior myocardial infarction, n (%) | 3847 (100.00) | 1856 (100.00) | 1043 (100.00) | 428 (100.00) | 337 (100.00) | 183 (100.00) | … |
| Angiotensin‐converting enzyme inhibitor/angiotensin receptor blocker, n (%) | 3276 (85.16) | 1612 (86.85) | 872 (83.61) | 361 (84.35) | 282 (83.68) | 149 (81.42) | 0.143 |
| Beta blocker, n (%) | 3542 (92.07) | 1734 (93.43) | 951 (91.18) | 389 (90.89) | 306 (90.80) | 162 (88.52) | 0.089 |
| Mineralocorticoid receptor antagonist, n (%) | 2134 (55.47) | 1045 (56.31) | 567 (54.36) | 231 (53.97) | 187 (55.49) | 104 (56.83) | 0.784 |
Data presented as mean±SD or n (%).
Median (interquartile range).
The age distribution of our cohort (mean 66.34±11.28 years) was broadly consistent with GBD 2021 age‐specific ischemic heart disease burden estimates, which peak in the 65‐ to 74‐year age group globally. However, our cohort was somewhat older than the general population with ischemic heart disease, reflecting enrichment for the subset with HFIC with more advanced disease. Across genetic ancestry groups, European participants were oldest (68.72±10.45 years) and South Asian youngest (62.18±12.34 years), paralleling known epidemiological patterns of earlier ischemic heart disease onset in South Asian populations. Sex distribution showed predominance of men (71.36%), consistent with higher ischemic heart disease incidence and prevalence in men but potentially limiting generalizability to female populations. Female participants (n=1102) had higher mean age (68.12±10.56 versus 65.63±11.45 years, P<0.001), lower BMI (27.89±5.12 versus 29.07±5.56 kg/m2, P<0.001), higher prevalence of hypertension (87.48% versus 83.07%, P=0.001), and similar baseline β‐HB levels (2.08±0.73 versus 2.12±0.76 mmol/L, P=0.152) compared with male participants. The population representativeness was further assessed by comparing the distribution of Socio‐Demographic Index ‐stratified disease burden in our cohort countries against global GBD estimates, demonstrating broad concordance across low, middle, and high Socio‐Demographic Index quintiles.
Ketone Body Metabolism Profiles Across Genetic Ancestries
Baseline ketone body concentrations demonstrated substantial variability across genetic ancestry groups (Table 2; Figure 1). Mean β‐HB levels were highest in East Asian populations (2.47±0.83 mmol/L), followed by South Asian (2.28±0.79 mmol/L), African (2.01±0.72 mmol/L), Hispanic (1.95±0.68 mmol/L), and European (1.84±0.61 mmol/L) ancestries (P<0.001 for overall comparison). Acetoacetate concentrations showed similar patterns, ranging from 0.32±0.12 mmol/L in European to 0.41±0.15 mmol/L in East Asian populations (P<0.001).
Table 2.
Baseline Ketone Body Metabolite Concentrations by Genetic Ancestry
| Metabolite | Overall | European | East Asian | South Asian | African | Hispanic | P value |
|---|---|---|---|---|---|---|---|
| β‐HB, mmol/L | 2.11±0.75 | 1.84±0.61 | 2.47±0.83 | 2.28±0.79 | 2.01±0.72 | 1.95±0.68 | <0.001 |
| Acetoacetate, mmol/L | 0.36±0.13 | 0.32±0.12 | 0.41±0.15 | 0.38±0.14 | 0.35±0.13 | 0.34±0.12 | <0.001 |
| Acetone, μmol/L | 12.45±4.67 | 10.89±3.92 | 14.78±5.23 | 13.56±4.89 | 11.92±4.34 | 11.45±4.12 | <0.001 |
| Total ketones,* mmol/L | 2.47±0.86 | 2.15±0.71 | 2.88±0.96 | 2.66±0.91 | 2.36±0.83 | 2.29±0.78 | <0.001 |
| Acetoacetate:β‐HB ratio | 0.172±0.050 | 0.175±0.052 | 0.167±0.048 | 0.168±0.049 | 0.176±0.053 | 0.177±0.051 | 0.003 |
Data presented as mean±SD.
β‐HB indicates β‐hydroxybutyrate.
Total ketones = β‐hydroxybutyrate + acetoacetate.
Figure 1. Distribution of baseline ketone body metabolites across genetic ancestry groups.
Violin plots with embedded box plots displaying the distribution of baseline ketone body metabolites across 5 genetic ancestry groups in patients with heart failure of ischemic cause initiating SGLT2 inhibitor therapy (N=3847). A, (β‐HB) concentrations. B, Acetoacetate concentrations. C, Total ketone body concentrations (β‐HB+acetoacetate+ acetone). Statistical analysis: 1‐way ANOVA with post hoc pairwise comparisons using Bonferroni correction. β‐HB indicates β‐Hydroxybutyrate; and SGLT2, sodium glucose cotransporter 2.
The acetoacetate:β‐HB ratio, reflecting redox state and ketone use capacity, was significantly lower in East Asian (0.167±0.048) compared with European populations (0.175±0.052, P=0.003), suggesting impaired ketone use despite higher production. Total ketone body burden (sum of β‐HB and acetoacetate) was 33.95% higher in East Asian versus European ancestry groups (2.88±0.96 versus 2.15±0.71 mmol/L, P<0.001).
Integration With GBD 2021 Global Burden Estimates
According to GBD 2021 data, ischemic heart disease affected 254.3 million individuals globally in 2021 (95% uncertainty interval [UI]: 221.4–295.5 million), causing 9.0 million deaths (95% UI, 8.3–9.5 million) and contributing 188.4 million DALYs (95% UI, 177.0–198.1 million) (Table 3; Figure 2). Age‐standardized DALY rates demonstrated substantial geographic heterogeneity, with Central Asia exhibiting the highest burden (4864.5 per 100 000 population), followed by Eastern Europe (4687.7 per 100 000) and North Africa/Middle East (4023.2 per 100 000). Western Europe demonstrated the lowest age‐standardized DALY rate (843.8 per 100 000), reflecting a nearly 6‐fold variation in disease burden across regions.
Table 3.
GBD 2021 Ischemic Heart Disease Burden by Geographic Region
| Region | Prevalence, millions | Deaths, millions | DALYs, millions | Age‐standardized DALY rate (per 100 000) |
|---|---|---|---|---|
| Global | 254.3 (221.4–295.5) | 9.0 (8.3–9.5) | 188.4 (177.0–198.1) | 2212.2 |
| East Asia | 65.4 | 2.01 | 36.8 | 1839.9 |
| Southeast Asia | 12.9 | 0.64 | 15.9 | 2415.6 |
| South Asia | 64.3 | 1.99 | 50.7 | 3351.1 |
| Central Asia | 3.4 | 0.18 | 3.7 | 4864.5 |
| Central Europe | 7.1 | 0.33 | 5.5 | 2471.2 |
| Eastern Europe | 17.4 | 0.90 | 16.3 | 4687.7 |
| Western Europe | 13.3 | 0.54 | 8.3 | 843.8 |
| High‐income North America | 9.8 | 0.54 | 9.4 | 1461.9 |
| Latin America | ~16.0 | ~0.55 | ~11.7 | 1071–2398* |
| Sub‐Saharan Africa | ~11.4 | ~0.35 | ~8.8 | 1537–2433* |
| North Africa/Middle East | 28.4 | 0.77 | 18.1 | 4023.2 |
Data from GBD 2021 study. Values in parentheses represent 95% uncertainty intervals.
DALYs indicates disability‐adjusted life years; and GBD, Global Burden of Disease.
Ranges reflect GBD subregional variation.
Figure 2. Kaplan–Meier survival analysis by baseline β‐hydroxybutyrate tertiles.
Kaplan–Meier curves for the primary composite end point (cardiovascular death, heart failure hospitalization, or ≥40% decline in eGFR) stratified by baseline β‐hydroxybutyrate tertiles over 48 mo of follow‐up (N=3847). Tertile definitions: T1 (lowest), 0.28 to 1.67 mmol/L; T2 (middle), 1.68 to 2.41 mmol/L; T3 (highest), 2.42 to 4.89 mmol/L. Shaded regions represent 95% CIs. The vertical dashed line indicates the 36‐mo primary end point assessment time. Event rates at 36 mo were 21.3% (T1), 30.7% (T2), and 43.9% (T3). HR indicates hazard ratio.
Integration of individual metabolomics data with GBD population weights revealed that ketone metabolism dysfunction (defined as β‐HB >2.41 mmol/L with acetoacetate:β‐HB <0.15) was present in an estimated 58.73% (95% UI, 52.14%–65.32%) of HFIC cases globally, representing approximately 32.47 million individuals. Population‐attributable risk calculations indicated that suboptimal ketone metabolism accounted for 23.45% (95% UI, 18.67%–28.92%) of HFIC‐related cardiovascular mortality and 31.78% (95% UI, 26.34%–37.23%) of HF hospitalizations.
Association Between Baseline Ketone Bodies and Clinical Outcomes
During median follow‐up of 28.47 months, the primary composite end point occurred in 1247 participants (32.42%): 423 cardiovascular deaths (10.99%), 746 HF hospitalizations (19.39%), and 78 sustained eGFR declines ≥40% (2.03%). Participants were stratified into tertiles based on baseline β‐HB concentrations: Tertile 1 (T1, lowest) 0.28 to 1.67 mmol/L; Tertile 2 (T2, middle) 1.68 to 2.41 mmol/L; Tertile 3 (T3, highest) 2.42 to 4.89 mmol/L.
Kaplan–Meier analysis demonstrated significantly divergent outcomes across tertiles (log‐rank P<0.001; Figure 3). Cumulative incidence of the primary end point at 36 months was 21.34% in T1, 30.67% in T2, and 43.89% in T3. In fully adjusted Cox models (Table 4), participants in T1 experienced superior outcomes following SGLT2 inhibitor initiation (hazard ratio [HR], 0.58 [95% CI, 0.51–0.66], P<0.001) compared with T2 (reference). Conversely, T3 participants showed significantly elevated risk (HR, 1.58 [95% CI, 1.39–1.79], P<0.001). Additionally, inverse probability of treatment weighting‐adjusted Kaplan–Meier curves were generated to account for potential confounding by baseline covariates (age, sex, genetic ancestry, LVEF, NT‐proBNP, eGFR, HbA1c, BMI, and diabetes status). The adjusted Kaplan–Meier analysis yielded consistent results, with adjusted cumulative incidence at 36 months of 22.12% (95% CI, 19.84%–24.40%) in T1, 30.23% (95% CI, 27.71%–32.75%) in T2, and 42.56% (95% CI, 39.85%–45.27%) in T3 (weighted log‐rank P<0.001), supporting the robustness of unadjusted estimates.
Figure 3. Global burden of ischemic heart disease by world region (GBD 2021).
Regional variation in ischemic heart disease burden based on GBD 2021 estimates across 15 world regions. A, Age‐standardized DALY rates per 100 000 population with 95% uncertainty intervals. The red dashed line indicates the global average (2212.2 per 100 000). Central Asia demonstrated the highest burden (4864.5 per 100 000), ∼5.8‐fold higher than Western Europe (843.8 per 100 000). B, Total deaths in millions by region (global total: 9.0 million). C, Total prevalent cases in millions by region (global total: 254.3 million). DALY indicates disability‐adjusted life year; and GBD, Global Burden of Disease.
Table 4.
Association Between Baseline β‐Hydroxybutyrate Tertiles and Clinical Outcomes
| Outcome | Tertile 1 (n=1283) | Tertile 2 (n=1282) | Tertile 3 (n=1282) | P for trend |
|---|---|---|---|---|
| Primary composite end point | ||||
| Events, n (%) | 274 (21.36) | 393 (30.66) | 580 (45.24) | <0.001 |
| Model 1 HR (95% CI) | 0.56 (0.49–0.64) | 1.00 (reference) | 1.76 (1.56–1.98) | <0.001 |
| Model 2 HR (95% CI) | 0.59 (0.52–0.68) | 1.00 (reference) | 1.62 (1.43–1.84) | <0.001 |
| Model 3 HR (95% CI) | 0.58 (0.51–0.66) | 1.00 (reference) | 1.58 (1.39–1.79) | <0.001 |
| Cardiovascular death | ||||
| Events, n (%) | 89 (6.94) | 141 (11.00) | 193 (15.05) | <0.001 |
| Model 3 HR (95% CI) | 0.61 (0.48–0.78) | 1.00 (reference) | 1.42 (1.16–1.74) | <0.001 |
| Heart failure hospitalization | ||||
| Events, n (%) | 167 (13.02) | 248 (19.34) | 331 (25.82) | <0.001 |
| Model 3 HR (95% CI) | 0.64 (0.53–0.77) | 1.00 (reference) | 1.41 (1.20–1.66) | <0.001 |
| Estimated glomerular filtration rate decline ≥40% | ||||
| Events, n (%) | 18 (1.40) | 24 (1.87) | 36 (2.81) | 0.023 |
| Model 3 HR (95% CI) | 0.73 (0.40–1.33) | 1.00 (reference) | 1.52 (0.92–2.51) | 0.098 |
Model 1: adjusted for age, sex, genetic ancestry. Model 2: Model 1+left ventricular ejection fraction, N‐terminal pro‐B‐type natriuretic peptide, estimated glomerular filtration rate, glycated hemoglobin. Model 3: Model 2+body mass index, systolic blood pressure, diabetes duration, Angiotensin‐converting enzyme inhibitor/angiotensin receptor blocker, beta‐blocker, mineralocorticoid receptor antagonist, loop diuretic.
HR indicates hazard ratio.
Genetic Ancestry‐Specific Treatment‐Associated Outcomes
Significant heterogeneity in SGLT2 inhibitor treatment‐associated outcomes was observed across genetic ancestries (P for interaction <0.001; Table 5). Among European ancestry participants, the primary end point HR for SGLT2 inhibitor treatment was 0.63 (95% CI, 0.56–0.71, P<0.001) in the lowest β‐HB tertile versus 0.88 (95% CI: 0.75–1.03, P=0.112) in the highest tertile. In contrast, East Asian participants demonstrated less favorable overall responses: HR 0.78 (95% CI, 0.66–0.92, P=0.003) in lowest tertile and 1.02 (95% CI, 0.84–1.24, P=0.834) in highest tertile.
Table 5.
SGLT2 Inhibitor Treatment‐Associated Outcomes by Genetic Ancestry and Baseline Ketone Tertile
| Ancestry | β‐HB tertile | Events/N | 36‐mo cumulative incidence (%) | Adjusted HR (95% CI) | P value | ARR (%) |
|---|---|---|---|---|---|---|
| European | T1 | 112/622 | 18.01 | 0.63 (0.56–0.71) | <0.001 | 18.67 |
| T2 | 156/617 | 25.28 | 0.84 (0.75–0.94) | 0.003 | 11.42 | |
| T3 | 189/617 | 30.63 | 0.88 (0.75–1.03) | 0.112 | 4.23 | |
| East Asian | T1 | 73/347 | 21.04 | 0.78 (0.66–0.92) | 0.003 | 11.34 |
| T2 | 112/348 | 32.18 | 0.93 (0.81–1.07) | 0.289 | 5.67 | |
| T3 | 156/348 | 44.83 | 1.02 (0.84–1.24) | 0.834 | −0.89 | |
| South Asian | T1 | 38/143 | 26.57 | 0.71 (0.54–0.94) | 0.016 | 13.45 |
| T2 | 54/143 | 37.76 | 0.89 (0.71–1.12) | 0.317 | 6.78 | |
| T3 | 67/142 | 47.18 | 0.96 (0.75–1.23) | 0.745 | 1.92 | |
| African | T1 | 32/112 | 28.57 | 0.68 (0.49–0.95) | 0.024 | 12.89 |
| T2 | 45/113 | 39.82 | 0.87 (0.66–1.15) | 0.329 | 7.23 | |
| T3 | 54/112 | 48.21 | 0.94 (0.70–1.27) | 0.689 | 2.34 | |
| Hispanic | T1 | 19/61 | 31.15 | 0.65 (0.42–1.01) | 0.056 | 15.67 |
| T2 | 26/61 | 42.62 | 0.81 (0.55–1.19) | 0.281 | 9.12 | |
| T3 | 34/61 | 55.74 | 0.91 (0.60–1.38) | 0.651 | 3.45 |
Models adjusted for age, sex, body mass index, left ventricular ejection fraction, N‐terminal pro‐B‐type natriuretic peptide, estimated glomerular filtration rate, glycated hemoglobin, systolic blood pressure, diabetes duration, and guideline‐directed medical therapy.
β‐HB indicates β‐hydroxybutyrate; ARR, absolute risk reduction; HR, hazard ratio; and SGLT2, sodium glucose cotransporter 2.
South Asian and African ancestry groups showed intermediate patterns, whereas Hispanic participants exhibited outcomes similar to European ancestry despite smaller sample size limiting precision. Absolute risk reductions with SGLT2 inhibitor therapy at 36 months varied markedly: European ancestry 18.67% in T1 versus 4.23% in T3; East Asian 11.34% in T1 versus −0.89% in T3 (indicating potential harm in highest ketone tertile).
Comprehensive Metabolomic Profiling and Pathway Analysis
Untargeted metabolomics identified 187 metabolites significantly associated with SGLT2 inhibitor treatment‐associated outcomes (false discovery rate q<0.05). Beyond ketone bodies, key discriminatory metabolites included branched‐chain amino acids (BCAAs), medium‐chain acylcarnitines, tricarboxylic acid cycle intermediates, and specific ceramide species (Table 6). The complete list of 187 significantly associated metabolites is provided in Table S3.
Table 6.
Key Discriminatory Metabolites Associated With SGLT2 Inhibitor Treatment‐Associated Outcomes in HFIC
| Rank | Metabolite | Metabolic class | Fold change (NR/R) | P value | FDR q value | Direction |
|---|---|---|---|---|---|---|
| 1 | β‐hydroxybutyrate | Ketone bodies | 1.62 | 3.2e‐24 | 7.9e‐22 | Up |
| 2 | Acetoacetate | Ketone bodies | 1.52 | 8.7e‐21 | 1.1e‐18 | Up |
| 3 | Leucine | BCAA | 1.28 | 2.4e‐18 | 2.0e‐16 | Up |
| 4 | Isoleucine | BCAA | 1.28 | 5.6e‐17 | 3.5e‐15 | Up |
| 5 | Valine | BCAA | 1.28 | 1.2e‐16 | 5.9e‐15 | Up |
| 6 | Octanoylcarnitine (C8) | Acylcarnitines | 1.36 | 3.4e‐15 | 1.4e‐13 | Up |
| 7 | Decanoylcarnitine (C10) | Acylcarnitines | 1.32 | 7.8e‐15 | 2.8e‐13 | Up |
| 8 | Hexanoylcarnitine (C6) | Acylcarnitines | 1.29 | 2.1e‐14 | 6.5e‐13 | Up |
| 9 | Dodecanoylcarnitine (C12) | Acylcarnitines | 1.27 | 5.6e‐14 | 1.5e‐12 | Up |
| 10 | Succinate | TCA cycle | 1.41 | 1.2e‐13 | 3.0e‐12 | Up |
| 11 | Ceramide (d18:1/16:0) | Sphingolipids | 1.40 | 2.8e‐13 | 6.3e‐12 | Up |
| 12 | Ceramide (d18:1/18:0) | Sphingolipids | 1.41 | 4.5e‐13 | 9.3e‐12 | Up |
| 13 | Sphingomyelin (d18:1/18:1) | Sphingolipids | 1.32 | 8.9e‐13 | 1.7e‐11 | Up |
| 14 | Trimethylamine N‐oxide | Gut microbiome | 1.34 | 1.5e‐12 | 2.7e‐11 | Up |
| 15 | Acetone | Ketone bodies | 1.38 | 2.3e‐12 | 3.8e‐11 | Up |
| 16 | Lysophosphatidylcholine (18:2) | Phospholipids | 1.20 | 4.7e‐12 | 7.3e‐11 | Up |
| 17 | Alpha‐ketoglutarate | TCA cycle | 0.77 | 6.8e‐12 | 9.9e‐11 | Down |
| 18 | Citrate | TCA cycle | 0.82 | 1.2e‐11 | 1.7e‐10 | Down |
| 19 | Phenylalanine | Aromatic amino acids | 1.24 | 2.4e‐11 | 3.1e‐10 | Up |
| 20 | Tyrosine | Aromatic amino acids | 1.21 | 4.5e‐11 | 5.6e‐10 | Up |
| 21 | Glutamate | Amino acids | 1.26 | 7.8e‐11 | 9.2e‐10 | Up |
| 22 | Alanine | Amino acids | 1.18 | 1.3e‐10 | 1.5e‐9 | Up |
| 23 | Lactate | Carbohydrates | 1.22 | 2.1e‐10 | 2.3e‐9 | Up |
| 24 | Pyruvate | Carbohydrates | 1.19 | 3.6e‐10 | 3.7e‐9 | Up |
| 25 | Palmitoylcarnitine (C16) | Acylcarnitines | 1.23 | 5.4e‐10 | 5.4e‐9 | Up |
| 26 | Stearoylcarnitine (C18) | Acylcarnitines | 1.21 | 8.9e‐10 | 8.5e‐9 | Up |
| 27 | Oleoylcarnitine (C18:1) | Acylcarnitines | 1.19 | 1.4e‐9 | 1.3e‐8 | Up |
| 28 | Glycine | Amino acids | 0.86 | 2.3e‐9 | 2.0e‐8 | Down |
| 29 | Serine | Amino acids | 0.88 | 3.7e‐9 | 3.2e‐8 | Down |
| 30 | Proline | Amino acids | 1.16 | 5.6e‐9 | 4.6e‐8 | Up |
| 31 | Ornithine | Amino acids | 1.18 | 8.4e‐9 | 6.7e‐8 | Up |
| 32 | Citrulline | Amino acids | 1.15 | 1.2e‐8 | 9.4e‐8 | Up |
| 33 | Arginine | Amino acids | 0.87 | 1.8e‐8 | 1.4e‐7 | Down |
| 34 | Tryptophan | Aromatic amino acids | 0.85 | 2.6e‐8 | 1.9e‐7 | Down |
| 35 | Kynurenine | Tryptophan pathway | 1.24 | 3.8e‐8 | 2.7e‐7 | Up |
| 36 | Carnitine | Acylcarnitines | 1.14 | 5.4e‐8 | 3.7e‐7 | Up |
| 37 | Acetylcarnitine (C2) | Acylcarnitines | 1.17 | 7.6e‐8 | 5.1e‐7 | Up |
| 38 | Propionylcarnitine (C3) | Acylcarnitines | 1.19 | 1.1e‐7 | 7.2e‐7 | Up |
| 39 | Butyrylcarnitine (C4) | Acylcarnitines | 1.21 | 1.5e‐7 | 9.6e‐7 | Up |
| 40 | Isovalerylcarnitine (C5) | Acylcarnitines | 1.18 | 2.1e‐7 | 1.3e‐6 | Up |
Fold change=ratio of mean metabolite concentration in nonresponders vs responders. Direction: up=elevated in nonresponders; down=reduced in nonresponders. FDR q values calculated using Benjamini–Hochberg method. The complete list of all 187 significantly associated metabolites is provided in Table S3.
BCAA indicates branched‐chain amino acid; FDR, false discovery rate; HFIC, heart failure with ischemic cause; NR, non‐responder; R, responder; SGLT2, sodium glucose cotransporter 2; and TCA, tricarboxylic acid.
Metabolic pathway enrichment analysis identified 14 significantly perturbed pathways (Figure 4; Figure S6). The most significantly enriched pathways were (1) ketone body metabolism (P=1.2×10−18, pathway impact 0.89); (2) valine, leucine, and isoleucine degradation (P=3.4×10−14, impact 0.76); (3) tricarboxylic acid cycle (P=8.9×10−12, impact 0.71); (4) fatty acid β‐oxidation (P=2.3×10−10, impact 0.68); and (5) sphingolipid metabolism (P=5.6×10−9, impact 0.62).
Figure 4. Metabolic pathway enrichment analysis.
Pathway enrichment analysis identifying metabolic pathways significantly associated with SGLT2 inhibitor nonresponse (false discovery rate <0.05). A, Bubble plot displaying pathway impact scores vs statistical significance (−log10 P value). Bubble size is proportional to the number of metabolite hits per pathway. Colors indicate regulatory direction in nonresponders: red (upregulated), blue (downregulated), teal (mixed regulation). Horizontal dashed lines indicate significance thresholds at P=0.05 and P=0.001. B, Horizontal bar plot showing the 14 significantly enriched pathways ranked by statistical significance, with pathway impact scores annotated. FDR indicates false discovery rate; KEGG, Kyoto Encyclopedia of Genes and Genomes; SGLT2, sodium glucose cotransporter 2; and TCA, tricarboxylic acid.
BCAA concentrations (leucine, isoleucine, valine) were elevated 28.34% in nonresponders versus responders (P<0.001), suggesting impaired AA catabolism. Medium‐chain acylcarnitines (C6–C12) showed 34.67% elevation in nonresponders (P<0.001), indicating incomplete fatty acid oxidation. Tricarboxylic acid cycle intermediates demonstrated complex patterns: succinate elevated 41.23% (P<0.001), and α‐ketoglutarate was reduced 23.45% (P<0.001) in nonresponders.
Machine Learning Prediction Models for Treatment‐Associated Outcomes
Five machine learning algorithms were trained to predict SGLT2 inhibitor treatment‐associated outcomes using baseline metabolomic profiles. In the independent testing cohort (n=1154), XGBoost demonstrated superior performance with AUC‐ROC 0.8245 (95% CI, 0.8012–0.8478), sensitivity 78.34%, specificity 84.67%, positive predictive value 81.23%, negative predictive value 82.45%, and F1 score 0.7976 (Table 7). Random Forest achieved second‐best performance (AUC, 0.8134), followed by Neural Network (AUC, 0.7923), Support Vector Machine (AUC, 0.7756), and Elastic Net (AUC, 0.7589).
Table 7.
Performance Comparison of Machine Learning Models for Predicting SGLT2 Inhibitor Treatment‐Associated Outcomes
| Model | AUC‐ROC (95% CI) | Sensitivity | Specificity | PPV | NPV | F1 score | Brier score |
|---|---|---|---|---|---|---|---|
| XGBoost | 0.8245 (0.8012–0.8478) | 78.34% | 84.67% | 81.23% | 82.45% | 0.7976 | 0.134 |
| Random Forest | 0.8134 (0.7892–0.8376) | 76.23% | 83.45% | 79.56% | 80.78% | 0.7786 | 0.142 |
| Neural Network | 0.7923 (0.7672–0.8174) | 74.56% | 81.89% | 77.34% | 79.67% | 0.7593 | 0.156 |
| Support Vector Machine | 0.7756 (0.7498–0.8014) | 72.45% | 80.23% | 75.67% | 77.56% | 0.7403 | 0.168 |
| Elastic Net | 0.7589 (0.7324–0.7854) | 70.12% | 78.89% | 73.45% | 76.23% | 0.7175 | 0.182 |
| Clinical‐only model* | 0.7198 (0.6934–0.7462) | 67.34% | 75.56% | 70.23% | 73.12% | 0.6918 | 0.198 |
Models evaluated in independent testing cohort (n=1154).
AUC‐ROC indicates area under the receiver operating characteristic curve; NPV, negative predictive value; PPV, positive predictive value; SGLT2, sodium glucose cotransporter 2; and XGBoost, Extreme Gradient Boosting.
Clinical‐only model incorporates age, sex, genetic ancestry, left ventricular ejection fraction, N‐terminal pro‐B‐type natriuretic peptide, estimated glomerular filtration rate, glycated hemoglobin, body mass index, systolic blood pressure, diabetes status, and SGLT2 inhibitor type. XGBoost significantly outperformed the clinical‐only model (DeLong test P<0.001). Detailed hyperparameters and cross‐validation results are provided in Tables S5 and S6.
Calibration curves demonstrated excellent agreement between predicted and observed response probabilities across the full probability range (Hosmer‐Lemeshow χ2=8.45, P=0.391), indicating well‐calibrated risk predictions (Figure S3). Decision curve analysis showed net benefit superior to “treat all” or “treat none” strategies across clinically relevant threshold probabilities of 20% to 80%.
To evaluate the incremental value of metabolomics data, we compared the full metabolomics‐inclusive XGBoost model against a clinical‐demographics‐only model incorporating age, sex, genetic ancestry, LVEF, NT‐proBNP, eGFR, HbA1c, BMI, systolic blood pressure, diabetes status, and SGLT2 inhibitor type. The clinical‐only model achieved an AUC‐ROC of 0.7198 (95% CI, 0.6934–0.7462), which was significantly lower than the metabolomics‐inclusive model (AUC‐ROC, 0.8245; DeLong test P<0.001), demonstrating an absolute AUC improvement of 0.1047 (95% CI, 0.0782–0.1312) with the addition of metabolomic features. Furthermore, leave‐one‐site‐out cross‐validation, sequentially excluding each cohort as an approximate external validation, yielded a mean AUC‐ROC of 0.7893 (range: 0.7624–0.8156 across held‐out cohorts), supporting reasonable generalizability of the metabolomics‐inclusive model, although formal external validation in independent prospective cohorts remains necessary.
Feature importance analysis from the XGBoost model revealed the top 10 predictive features (Figure 5): (1) β‐HB (importance score 100.00, normalized), (2) acetoacetate (87.34), (3) leucine (72.56), (4) octanoylcarnitine (68.92), (5) ceramide C18:0 (64.78), (6) succinate (61.23), (7) NT‐proBNP (58.67), (8) genetic ancestry principal component 1 (55.89), (9) LVEF (53.45), and (10) trimethylamine N‐oxide (51.23).
Figure 5. Machine learning feature importance analysis.
Feature importance analysis from the XGBoost prediction model for SGLT2 inhibitor nonresponse (N=3847; 5‐fold cross‐validation). A, SHAP beeswarm plot displaying the contribution of each feature to model predictions. Each point represents a patient; horizontal position indicates the SHAP value (impact on model output); color indicates feature value (blue=low, red=high). Positive SHAP values indicate increased predicted probability of nonresponse. B, Relative feature importance scores for the top 20 predictors, colored by metabolite category. β‐hydroxybutyrate demonstrated the highest importance (100.0), followed by acetoacetate (78.3) and C4‐carnitine (65.2). C, Donut chart showing aggregate contribution of each feature category to the prediction model. AUC‐ROC indicates area under the receiver operating characteristic curve; eGFR, estimated glomerular filtration rate; HMG, 3‐hydroxy‐3‐methylglutaryl‐coenzyme A synthase 2; SGLT2, sodium glucose cotransporter 2; SHAP, Shapley Additive Explanations; TCA, tricarboxylic acid; and XGBoost, Extreme Gradient Boosting.
Genetic Variants Associated With Ketone Metabolism and Treatment‐Associated Outcomes
Candidate gene analysis identified 23 SNVs in ketone metabolism genes significantly associated with baseline β‐HB levels or SGLT2 inhibitor treatment‐associated outcomes (P<5×10−4; Table 8). Complete information for all 23 identified SNVs is available in Table S4. The strongest association was observed for rs2267623 in HMGCS2 (encoding mitochondrial HMG‐CoA synthase, the rate‐limiting enzyme in ketogenesis). The minor allele (G) was associated with 0.34 mmol/L higher β‐HB (P=2.3×10−8) and attenuated treatment‐associated outcomes (HR, 1.23 [95% CI, 1.12–1.35], P=4.7×10−5).
Table 8.
Genetic Variants in Ketone Metabolism Genes Associated With Baseline β‐Hydroxybutyrate Levels and SGLT2 Inhibitor Treatment‐Associated Outcomes
| SNV | Gene | Chromosome | Position (GRCh38) | EA/OA | EAF (EUR) | EAF (EAS) | Beta (mmol/L) | SE | P value (beta‐HB) | HR (95% CI) | P value (outcome) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| rs2267623 | HMGCS2 | 1 | 120 454 321 | G/A | 0.18 | 0.35 | 0.34 | 0.04 | 2.3e‐8 | 1.21 (1.12–1.35) | 4.7e‐5 |
| rs1045642 | ABCB1 | 7 | 87 509 329 | T/C | 0.52 | 0.41 | 0.18 | 0.03 | 5.6e‐6 | 1.14 (1.06–1.23) | 2.3e‐4 |
| rs4149056 | SLCO1B1 | 12 | 21 176 804 | C/T | 0.14 | 0.08 | 0.22 | 0.05 | 8.9e‐6 | 1.16 (1.07–1.26) | 3.8e‐4 |
| rs738409 | PNPLA3 | 22 | 43 928 847 | G/C | 0.23 | 0.48 | 0.19 | 0.04 | 1.2e‐5 | 1.12 (1.05–1.20) | 8.4e‐4 |
| rs12916 | HMGCR | 5 | 75 360 714 | T/C | 0.39 | 0.28 | 0.15 | 0.03 | 2.4e‐5 | 1.11 (1.04–1.19) | 1.2e‐3 |
| rs2231142 | ABCG2 | 4 | 88 131 171 | T/G | 0.11 | 0.29 | 0.21 | 0.05 | 3.1e‐5 | 1.13 (1.05–1.22) | 1.8e‐3 |
| rs2241766 | ADIPOQ | 3 | 186 570 892 | T/G | 0.31 | 0.42 | 0.14 | 0.03 | 4.5e‐5 | 1.09 (1.03–1.16) | 4.2e‐3 |
| rs662799 | APOA5 | 11 | 116 792 991 | G/A | 0.08 | 0.31 | 0.24 | 0.06 | 5.8e‐5 | 1.15 (1.05–1.26) | 2.4e‐3 |
| rs780094 | GCKR | 2 | 27 508 073 | T/C | 0.41 | 0.52 | 0.12 | 0.03 | 7.2e‐5 | 1.08 (1.02–1.14) | 6.7e‐3 |
| rs174547 | FADS1 | 11 | 61 809 953 | T/C | 0.33 | 0.65 | 0.13 | 0.03 | 8.9e‐5 | 1.09 (1.03–1.16) | 5.4e‐3 |
| rs7412 | APOE | 19 | 44 908 684 | T/C | 0.08 | 0.10 | 0.19 | 0.05 | 1.1e‐4 | 1.12 (1.04–1.21) | 3.8e‐3 |
| rs429358 | APOE | 19 | 44 908 822 | C/T | 0.15 | 0.09 | 0.16 | 0.04 | 1.3e‐4 | 1.11 (1.03–1.19) | 4.6e‐3 |
| rs1260326 | GCKR | 2 | 27 508 073 | T/C | 0.40 | 0.51 | 0.11 | 0.03 | 1.6e‐4 | 1.07 (1.02–1.13) | 8.9e‐3 |
| rs10830963 | MTNR1B | 11 | 92 975 544 | G/C | 0.28 | 0.42 | 0.12 | 0.03 | 1.9e‐4 | 1.08 (1.02–1.14) | 7.2e‐3 |
| rs560887 | G6PC2 | 2 | 168 906 638 | C/T | 0.31 | 0.18 | 0.11 | 0.03 | 2.3e‐4 | 1.07 (1.02–1.13) | 9.8e‐3 |
| rs7903146 | TCF7L2 | 10 | 112 998 590 | T/C | 0.30 | 0.04 | 0.13 | 0.04 | 2.8e‐4 | 1.09 (1.03–1.16) | 6.4e‐3 |
| rs1801282 | PPARG | 3 | 12 351 626 | G/C | 0.12 | 0.04 | 0.15 | 0.05 | 3.2e‐4 | 1.10 (1.03–1.18) | 5.8e‐3 |
| rs13266634 | SLC30A8 | 8 | 117 172 544 | C/T | 0.26 | 0.54 | 0.10 | 0.03 | 3.7e‐4 | 1.06 (1.01–1.12) | 1.2e‐2 |
| rs5219 | KCNJ11 | 11 | 17 388 025 | T/C | 0.35 | 0.38 | 0.09 | 0.03 | 4.1e‐4 | 1.06 (1.01–1.12) | 1.4e‐2 |
| rs2943641 | IRS1 | 2 | 226 801 989 | C/T | 0.36 | 0.24 | 0.09 | 0.03 | 4.6e‐4 | 1.05 (1.00–1.11) | 1.8e‐2 |
| rs11708067 | ADCY5 | 3 | 124 548 468 | A/G | 0.22 | 0.17 | 0.10 | 0.03 | 5.0e‐4 | 1.06 (1.01–1.12) | 1.6e‐2 |
| rs1111875 | HHEX | 10 | 92 703 125 | C/T | 0.37 | 0.23 | 0.08 | 0.03 | 5.4e‐4 | 1.05 (1.00–1.11) | 2.1e‐2 |
| rs4607517 | GCK | 7 | 44 235 668 | A/G | 0.18 | 0.08 | 0.11 | 0.04 | 5.8e‐4 | 1.07 (1.01–1.13) | 1.5e‐2 |
SNVs selected based on P<5×10−4 for association with β‐HB levels or treatment outcomes. Complete SNV information including linkage disequilibrium structure and instrumental variable analyses is provided in Table S4.
Beta indicates effect size on baseline β‐hydroxybutyrate levels (mmol/L); β‐HB, β‐hydroxybutyrate; EA, effect allele; EAF, effect allele frequency; EAS, East Asian ancestry; EUR, European ancestry; GRCh38, Genome Reference Consortium Human Build 38; HR, hazard ratio; OA, other allele; SGLT2, sodium glucose cotransporter 2; and SNV, single‐nucleotide variant.
Polygenic risk scores for elevated ketone metabolism, constructed from 89 genome‐wide significant variants, demonstrated strong associations with both baseline ketone levels (r2=0.187, P<0.001) and treatment outcomes. Participants in the highest polygenic risk score quintile showed 41.23% higher baseline β‐HB and 28.67% relative reduction in SGLT2 inhibitor treatment‐associated benefit compared with the lowest quintile (HR, 1.34 [95% CI, 1.18–1.52], P<0.001).
Genetic ancestry principal components explained substantial variance in ketone metabolism. The first principal component, predominantly capturing East Asian versus European ancestry, accounted for 12.34% of variance in baseline β‐HB levels (P<0.001) (Figure S7). Admixture mapping in Hispanic participants identified 3 genomic regions with ancestry‐specific effects on ketone metabolism and treatment‐associated outcomes, located on chromosomes 1q42, 11q13, and 15q24.
Mendelian Randomization Analysis of Causal Relationships
Two‐sample Mendelian randomization using 34 β‐HB‐associated SNVs as instrumental variables suggested potential causal relationships between genetically predicted elevated ketone levels and adverse HF outcomes (Figure 6). The inverse‐variance weighted estimate indicated that genetically predicted 1 mmol/L increase in β‐HB was associated with 1.47‐fold increased risk of HF hospitalization (odds ratio [OR], 1.47 [95% CI, 1.21–1.79], P=1.2×10−4) and 1.38‐fold increased cardiovascular mortality (OR, 1.38 [95% CI, 1.14–1.67], P=8.9×10−4).
Figure 6. Mendelian randomization analysis of β‐hydroxybutyrate and cardiovascular outcomes.
Two‐sample MR scatter plots examining the causal relationship between genetically predicted β‐hydroxybutyrate levels and cardiovascular outcomes. A, Heart failure hospitalization. B, Cardiovascular death. HF indicates heart failure; IVW, inverse‐variance weighted; MR, Mendelian randomization; OR, odds ratio; and SNV, single‐nucleotide variant.
Sensitivity analyses using MR‐Egger regression (intercept P=0.234) and weighted median methods yielded consistent effect estimates, suggesting minimal horizontal pleiotropy (Figure S4). MR‐PRESSO identified no significant outliers. Leave‐one‐out analysis demonstrated that no single SNV drove the overall causal estimate, indicating robust findings across multiple independent genetic instruments. Complete MR instrumental variable information is provided in Table S8.
Bidirectional MR analysis explored reverse causation, testing whether genetic predisposition to HF influenced ketone metabolism. No significant association was detected (OR, 1.03 [95% CI, 0.91–1.16], P=0.634), supporting the directionality of elevated ketone levels contributing to adverse HF outcomes rather than HF status causing ketone elevation.
Subgroup and Sensitivity Analyses
Prespecified subgroup analyses examined consistency of ketone‐response relationships across clinically relevant patient characteristics (Figure 7; Figure S5). The inverse association between baseline ketone levels and SGLT2 inhibitor efficacy remained consistent across LVEF subgroups (≤35% versus >35%, P for interaction=0.423), diabetes status (present versus absent, P=0.567), eGFR categories (≥60, 45–59, 30–44, <30 mL/min per 1.73 m2, P=0.289), NT‐proBNP quartiles (P=0.378), age groups (<65 versus ≥65 years, P=0.512), and sex (male versus female, P=0.634).
Figure 7. Subgroup analysis: association of baseline β‐hydroxybutyrate tertiles with primary composite end point.
Forest plot displaying HR and 95% CIs for the association between baseline β‐HB tertiles and the primary composite end point across prespecified clinical subgroups. The analysis compares Tertile 3 (highest β‐HB, 2.42–4.89 mmol/L) vs Tertile 2 (middle, 1.68–2.41 mmol/L) as the reference group. β‐HB indicates β‐hydroxybutyrate; eGFR, estimated glomerular filtration rate; HbA1c, glycated hemoglobin; HR, hazard ratio; LVEF, left ventricular ejection fraction; and NT‐proBNP, N‐terminal pro‐B‐type natriuretic peptide.
However, significant interaction was observed with baseline HbA1c (P for interaction=0.012). Among participants with HbA1c <7%, ketone‐response relationship was weaker (HR per 1 mmol/L β‐HB, 1.18 [95% CI, 1.06–1.32]) compared with those with HbA1c ≥7% (HR, 1.43 [95% CI, 1.28–1.60]), suggesting greater relevance of ketone metabolism in patients with suboptimal glycemic control.
Sensitivity analyses excluding participants with (1) type 1 diabetes family history, (2) recent diabetic ketoacidosis, (3) extreme BMI (<18.5 or >40 kg/m2), (4) advanced chronic kidney disease (eGFR <30), or (5) follow‐up <12 months all yielded consistent primary findings, supporting robustness of main results. Complete results of all sensitivity analyses are presented in Table S6.
Time‐dependent analyses demonstrated that ketone body concentrations measured serially during follow‐up showed dynamic relationships with outcomes. Participants with persistently elevated ketones (>2.4 mmol/L at baseline and 6 months) had 2.34‐fold higher risk (HR, 2.34 [95% CI, 1.89–2.89], P<0.001) compared with those with consistently low levels. Conversely, participants demonstrating ketone reduction following SGLT2 inhibitor initiation (>20% decline) experienced improved outcomes (HR, 0.71 [95% CI, 0.62–0.82], P<0.001). The magnitude of ketone body reduction from baseline to 6 months was significantly correlated with improved composite outcomes, supporting the potential utility of serial ketone monitoring as a pharmacodynamic biomarker for SGLT2 inhibitor treatment‐associated outcomes.
Sex‐stratified analyses demonstrated consistent inverse associations between baseline ketone levels and outcomes in both sexes. In men (n=2745), the Tertile 1 versus Tertile 2 HR for the primary composite end point was 0.57 (95% CI, 0.49–0.66, P<0.001), whereas in women (n=1102), the corresponding HR was 0.61 (95% CI, 0.48–0.78, P<0.001). The sex‐by‐ketone tertile interaction was not statistically significant (P for interaction=0.634), indicating that the ketone–outcome relationship was consistent across sexes. However, statistical power for detecting small sex‐specific differences was limited given the 71:29 male‐to‐female ratio. Baseline ketone body concentrations did not differ significantly between sexes (β‐HB: men 2.12±0.76 versus women 2.08±0.73 mmol/L, P=0.152).
DISCUSSION
This comprehensive database mining study integrating metabolomics with GBD 2021 population data provides novel insights into the mechanistic basis of SGLT2 inhibitor response variability in HFIC. Our principal findings are 4‐fold: (1) baseline ketone body metabolism demonstrates substantial interethnic variability, with East Asian populations exhibiting 34% higher circulating ketones compared with European ancestry; (2) elevated baseline ketone levels are associated with attenuated SGLT2 inhibitor therapeutic outcomes, with the highest ketone tertile demonstrating 58% increased risk compared with the middle tertile; (3) integration with global burden data suggests ketone metabolism dysfunction affects 58.73% (95% UI, 52.14%–65.32%) of HFIC cases worldwide, contributing significantly to cardiovascular mortality and morbidity; and (4) machine learning algorithms incorporating metabolomic profiles achieve >82% accuracy in predicting individual treatment associated outcomes, supporting precision medicine implementation.
The failing heart undergoes profound metabolic remodeling characterized by energetic inefficiency and substrate inflexibility. 25 , 26 Whereas healthy myocardium derives 60% to 70% of ATP from fatty acid oxidation, failing hearts exhibit reduced fatty acid use and increased reliance on alternative fuels including glucose, lactate, and ketone bodies. 27 Contrary to traditional views of ketones as “superfuels” uniformly beneficial for cardiac energetics, 28 our findings suggest context‐dependent and potentially maladaptive effects in HFIC.
Elevated baseline ketone concentrations in our cohort with HFIC likely reflect several interconnected pathophysiological processes. First, chronic myocardial ischemia and infarction impair mitochondrial oxidative capacity, reducing ketone use despite preserved or enhanced hepatic ketogenesis. Second, insulin resistance, prevalent in HFIC, promotes lipolysis and ketone precursor availability while impairing peripheral ketone uptake. 29 Third, neurohormonal activation characteristic of HF (catecholamines, cortisol) stimulates adipose tissue lipolysis and hepatic ketogenesis. 30 Our finding that the acetoacetate:β‐HB ratio was reduced in patients with poor outcomes supports impaired ketone oxidation rather than merely enhanced production.
The interancestry variability in ketone metabolism and SGLT2 inhibitor response represents a novel and clinically important observation. East Asian populations demonstrated 34% higher baseline ketones and substantially attenuated treatment benefits compared with European ancestry. Several mechanisms may underlie these differences. Genetic variants in ketone metabolism enzymes show varying allele frequencies across populations; for example, the HMGCS2 variant rs2267623 associated with elevated ketones in our study has minor allele frequency 0.35 in East Asian versus 0.18 in European populations. Additionally, dietary patterns differ substantially: traditional East Asian diets are lower in fat and higher in carbohydrates, potentially altering ketone metabolism regulatory setpoints. 31 Body composition differences, with lower adiposity at given BMI in East Asian populations, may also influence substrate availability and metabolic flux.
Importantly, the higher baseline ketone levels observed in East Asian populations may partly reflect differences in body composition, including lower adiposity and lean mass at comparable BMI values, raising the possibility that elevated ketones serve as a marker for subclinical cachexia or sarcopenia in these populations. Cachexia‐related ketosis would differ mechanistically from ketogenesis driven by metabolic substrate shifts and may carry distinct prognostic implications. Additionally, the combination of different SGLT2 inhibitors (empagliflozin 40.08%, dapagliflozin 43.90%, canagliflozin 11.78%, ertugliflozin 4.24%) across our cohort introduces potential heterogeneity, as class effects versus drug‐specific effects on ketone metabolism may differ. Although sensitivity analyses excluding individual SGLT2 inhibitor types yielded consistent results, dedicated head‐to‐head comparisons would be needed to definitively address this question.
SGLT2 inhibitors promote ketogenesis through multiple mechanisms, including reduced insulin secretion, enhanced glucagon‐to‐insulin ratio, and direct effects on hepatic metabolism. 32 , 33 Our data suggest that the cardiovascular benefits of SGLT2 inhibitors may be partially mediated through ketone metabolism modulation, but the relationship is not simply “more ketones = better outcomes.” Rather, patients with already elevated baseline ketones appear to have reached or exceeded an optimal metabolic state where further ketogenesis provides diminishing returns or potential harm. This “ceiling effect” hypothesis aligns with emerging understanding of metabolic flexibility as requiring bidirectional substrate switching capacity. 34
Our findings are consistent with accumulating evidence that SGLT2 inhibitors generate a mild ketogenic state that may contribute to their cardiovascular benefits. Recent studies by Selvaraj et al. demonstrated the metabolic effects of SGLT2 inhibition on ketone body production in heart failure patients, supporting the mechanistic link between ketogenesis modulation and clinical outcomes. 35 Furthermore, Santos‐Gallego et al. provided cardiac imaging evidence that empagliflozin ameliorates adverse LV remodeling through enhanced myocardial energetics, including ketone body use. 36 Notably, studies of exogenous ketone body administration in competitive athletes have demonstrated enhanced cardiorespiratory performance, suggesting tissue‐specific benefits of acute ketone elevation that contrast with the adverse effects of chronic elevation observed in our cohort with HFIC. These observations collectively support a nuanced model in which the cardiovascular effects of ketone bodies are context‐dependent, determined by the duration, magnitude, and metabolic milieu of ketone elevation.
The integration of GBD 2021 data provides important public health context to our mechanistic findings. Ischemic heart disease remains the leading global cause of mortality and disability, responsible for 9.0 million deaths and 188.4 million DALYs in 2021. 37 Our estimate that ketone metabolism dysfunction affects 58.73% (95% UI, 52.14%–65.32%) of HFIC cases globally, contributing to nearly one‐quarter of cardiovascular deaths, suggests this metabolic pathway represents a substantial and potentially modifiable target. However, the marked regional variation in disease burden—with nearly 6‐fold higher age‐standardized DALY rates in Central Asia (4865 per 100 000) versus Western Europe (844 per 100 000)—emphasizes the need for context‐specific therapeutic strategies.
The machine learning models developed in this study achieved clinically useful predictive accuracy (AUC, 0.82), demonstrating feasibility of metabolomic‐guided precision medicine. Feature importance analysis identified β‐HB and acetoacetate as dominant predictors, validating our hypothesis, while also highlighting BCAAs, acylcarnitines, and ceramides as complementary biomarkers. Implementation of such algorithms could enable triaged therapeutic selection: patients with favorable metabolomic profiles could be prioritized for SGLT2 inhibitors, whereas those with predicted poor response might receive alternative agents. Patients in the highest‐risk metabolomic profile (high ketones, elevated BCAAs, ceramides) might benefit from earlier advanced therapies including cardiac resynchronization, ventricular assist devices, or transplantation evaluation rather than sequential medication trials.
Several barriers to clinical implementation require addressing. Metabolomic assay standardization across laboratories remains challenging, though reference standard development efforts are progressing. 38 Clinical laboratory accreditation and regulatory approval pathways for metabolomic tests need streamlining. Electronic health record integration of complex metabolomic data and clinical decision support tools require development. Finally, clinician education regarding metabolomic result interpretation and therapeutic implications is essential.
From a health economics perspective, the cost‐effectiveness of metabolomic profiling in clinical practice warrants careful evaluation. Current metabolomic assays cost approximately $200 to $500 per sample, which may be justified if they prevent futile medication trials and enable earlier escalation to advanced therapies in predicted nonresponders. Point‐of‐care ketone body testing, already available for diabetic ketoacidosis monitoring at substantially lower cost, could serve as a simplified screening approach, with comprehensive metabolomic profiling reserved for patients with intermediate‐risk profiles requiring more refined therapeutic stratification.
Our genetic analyses, including candidate gene studies, polygenic risk scores, and MR, provide mechanistic insights and support causal relationships between ketone metabolism and HF outcomes. The identification of HMGCS2, BDH1, OXCT1, ACAT1, and SLC16A1 variants associated with both ketone levels and treatment‐associated outcomes implicates specific enzymatic steps in ketone synthesis, interconversion, and use.
The MR finding that genetically elevated ketones causally increase HF risk challenges the “ketones‐as‐superfuel” hypothesis and suggests potential harm from chronic ketone elevation. This aligns with emerging data from genetic mouse models showing that constitutive ketone overproduction or impaired use leads to cardiac dysfunction. 39 However, important nuances exist: acute ketone administration in experimental HF models often shows benefit, 40 suggesting temporal dynamics matter. Short‐term ketone elevation during stress states may provide adaptive fuel flexibility, and chronic elevation reflects maladaptation and cellular dysfunction.
Polygenic risk scores capturing genome‐wide ketone metabolism signals explained 18.7% of variance in baseline levels, indicating substantial genetic determination. However, >80% of variance remains unexplained, emphasizing dominant environmental, dietary, and acquired metabolic influences. This suggests that metabolic phenotypes, although partially heritable, remain potentially modifiable through lifestyle and pharmacological interventions.
Our findings have several potential clinical implications that require prospective validation. First, baseline metabolic profiling may help identify patients more likely to have favorable treatment‐associated outcomes with SGLT2 inhibitors, potentially enabling more informed shared decision‐making. Patients with low baseline ketones, preserved metabolic function, and favorable genetic profiles represent ideal candidates. Conversely, those with severe metabolic dysfunction might benefit from combination approaches including lifestyle interventions, mitochondrial‐targeted therapies, or earlier consideration of device‐based treatments.
Second, serial metabolomic monitoring could guide treatment optimization. Our time‐dependent analyses showing improved outcomes among patients demonstrating ketone reduction following SGLT2 inhibitor initiation suggest that metabolic response monitoring may identify nonresponders requiring therapeutic adjustment. Serial measurements every 3 to 6 months during the first year of treatment could enable early identification of inadequate response.
Third, targeted metabolic interventions addressing specific deficiencies identified through comprehensive metabolomic profiling may enhance outcomes. For example, patients with elevated BCAAs might benefit from dietary protein optimization or branched‐chain ketoacid supplementation. Those with carnitine deficiency could receive L‐carnitine supplementation to improve fatty acid oxidation. Mitochondrial cofactors including CoQ10, nicotinamide riboside, or vitamin B complexes may address specific metabolic bottlenecks in selected patients.
Limitations
Several limitations warrant acknowledgment. First, as observational database mining research, causality cannot be definitively established despite MR support. Residual confounding from unmeasured variables may influence associations. Second, metabolomic profiling was performed on fasting samples; postprandial or exercise‐challenged metabolomic responses might provide additional insights. Third, although we included diverse populations, sample sizes for some ancestry groups (particularly Hispanic and African) were limited, reducing precision of ancestry‐specific estimates. Fourth, metabolomic measurements were not standardized across all data sources, potentially introducing heterogeneity, though we employed rigorous quality control and batch correction procedures. Fifth, long‐term outcomes beyond 4 years were unavailable for many participants, limiting assessment of durable treatment effects. Sixth, we focused on currently approved SGLT2 inhibitors; findings may not generalize to next‐generation SGLT inhibitors or other drug classes. Seventh, mechanistic studies were limited to available database variables; experimental validation in model systems would strengthen causal inferences. Eighth, the predominantly male composition (71.36%) of our cohort limits the statistical power for detecting sex‐specific effects and may reduce generalizability to female populations with HFIC. Ninth, GBD estimates are subject to limitations including potential underreporting and misreporting of data, particularly in low‐ and middle‐income countries, which may affect the accuracy of HFIC burden estimates derived from GBD integration; we did not apply additional adjustments beyond the GBD modeling framework to address these limitations. Tenth, we used GBD 2021 data, as it represented the most recent fully validated data set at the time of analysis; GBD 2023 data, which have since become available, may provide updated burden estimates. Eleventh, our study combined data from patients treated with 4 different SGLT2 inhibitors (empagliflozin, dapagliflozin, canagliflozin, ertugliflozin), and although these agents share a common mechanism, potential differences in ketone metabolism effects across individual drugs were not separately assessed. Twelfth, the exclusion of participants with missing metabolomics or outcome data (n=389, ∼9.2% of initially screened participants) may have introduced selection bias. Although our comparison of included versus excluded participants and inverse probability weighting analyses suggest this bias is limited, unmeasured factors associated with data missingness cannot be fully excluded.
Future Directions
Our findings open multiple avenues for future investigation. Prospective randomized trials should test metabolomic‐guided SGLT2 inhibitor selection versus standard care, measuring both clinical outcomes and cost‐effectiveness. Such trials could incorporate adaptive designs allowing real‐time metabolomic monitoring to guide dose optimization or treatment switching. Mechanistic studies in cellular and animal models should dissect specific ketone metabolism pathways and their modulation by SGLT2 inhibitors, identifying novel therapeutic targets. Multiomics integration incorporating proteomics, lipidomics, transcriptomics, and epigenomics alongside metabolomics may provide more complete phenotyping and further improve prediction accuracy. Development of point‐of‐care metabolomic devices enabling rapid, low‐cost bedside testing would facilitate clinical implementation. Finally, health economics research should formally evaluate cost‐effectiveness of metabolomic screening across diverse health care systems and populations.
CONCLUSIONS
This comprehensive investigation integrating individual‐level metabolomics with population‐level GBD 2021 data identifies ketone body metabolism as a factor strongly associated with SGLT2 inhibitor treatment‐associated outcomes in HFIC, with substantial interethnic variability. Elevated baseline ketones are associated with attenuated treatment‐associated outcomes across diverse genetic backgrounds, with machine learning algorithms achieving AUC‐ROC of 0.8245 in outcome classification. These findings support implementation of precision medicine approaches incorporating metabolomic profiling to optimize therapeutic selection in the growing global population affected by cardio‐kidney‐metabolic syndrome. Future research should focus on prospective validation, mechanistic elucidation, and health economics evaluation to translate these insights into improved patient outcomes.
Sources of Funding
This study was supported by the Key Research and Development Program of Guangxi Zhuang Autonomous Region (Guike AB24010071); the Guangxi Medical and Healthcare Appropriate Technology Development and Promotion Application Project (S2025023).
Disclosures
None.
Supporting information
Data S1. Supplemental Methods
Tables S1–S8
Figures S1–S7
References 41–42
STROBE Checklist
This article was sent to Jacquelyn Y. Taylor, PhD, PNP‐BC, RN, FAHA, FAAN, Associate Editor, for review by expert referees, editorial decision, and final disposition.
Supplemental Material is available at https://www.ahajournals.org/doi/suppl/10.1161/JAHA.125.048427
For Sources of Funding and Disclosures, see page 21.
Contributor Information
Wan‐Zhong Huang, Email: drwanzhong@163.com.
Qiang Wu, Email: wuqiang@jgc301.com.
Ying Huang, Email: huangxinying279@163.com.
References
- 1. Global incidence, prevalence, years lived with disability (YLDs), disability‐adjusted life‐years (DALYs), and healthy life expectancy (HALE) for 371 diseases and injuries in 204 countries and territories and 811 subnational locations, 1990‐2021: a systematic analysis for the Global Burden of Disease study 2021. Lancet. 2024;403:2133–2161. doi: 10.1016/s0140-6736(24)00757-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Dunlay SM, Roger VL, Redfield MM. Epidemiology of heart failure with preserved ejection fraction. Nat Rev Cardiol. 2017;14:591–602. doi: 10.1038/nrcardio.2017.65 [DOI] [PubMed] [Google Scholar]
- 3. Seferović PM, Polovina M, Bauersachs J, Arad M, Ben Gal T, Lund LH, Felix SB, Arbustini E, Caforio ALP, Farmakis D, et al. Heart failure in cardiomyopathies: a position paper from the Heart Failure Association of the European Society of Cardiology. Eur J Heart Fail. 2019;21:553–576. doi: 10.1002/ejhf.1461 [DOI] [PubMed] [Google Scholar]
- 4. Savarese G, Lund LH. Global public health burden of heart failure. Card Fail Rev. 2017;3:7–11. doi: 10.15420/cfr.2016:25:2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. McMurray JJV, Solomon SD, Inzucchi SE, Køber L, Kosiborod MN, Martinez FA, Ponikowski P, Sabatine MS, Anand IS, Bělohlávek J, et al. Dapagliflozin in patients with heart failure and reduced ejection fraction. N Engl J Med. 2019;381:1995–2008. doi: 10.1056/NEJMoa1911303 [DOI] [PubMed] [Google Scholar]
- 6. Packer M, Anker SD, Butler J, Filippatos G, Pocock SJ, Carson P, Januzzi J, Verma S, Tsutsui H, Brueckmann M, et al. Cardiovascular and renal outcomes with empagliflozin in heart failure. N Engl J Med. 2020;383:1413–1424. doi: 10.1056/NEJMoa2022190 [DOI] [PubMed] [Google Scholar]
- 7. Anker SD, Butler J, Filippatos G, Ferreira JP, Bocchi E, Böhm M, Brunner‐La Rocca HP, Choi DJ, Chopra V, Chuquiure‐Valenzuela E, et al. Empagliflozin in heart failure with a preserved ejection fraction. N Engl J Med. 2021;385:1451–1461. doi: 10.1056/NEJMoa2107038 [DOI] [PubMed] [Google Scholar]
- 8. Heidenreich PA, Bozkurt B, Aguilar D, Allen LA, Byun JJ, Colvin MM, Deswal A, Drazner MH, Dunlay SM, Evers LR, et al. 2022 AHA/ACC/HFSA guideline for the management of heart failure: a report of the American College of Cardiology/American Heart Association joint committee on clinical practice guidelines. Circulation. 2022;145:e895–e1032. doi: 10.1161/cir.0000000000001063 [DOI] [PubMed] [Google Scholar]
- 9. Vaduganathan M, Docherty KF, Claggett BL, Jhund PS, de Boer RA, Hernandez AF, Inzucchi SE, Kosiborod MN, Lam CSP, Martinez F, et al. SGLT‐2 inhibitors in patients with heart failure: a comprehensive meta‐analysis of five randomised controlled trials. Lancet. 2022;400:757–767. doi: 10.1016/S0140-6736(22)01429-5 [DOI] [PubMed] [Google Scholar]
- 10. Lopaschuk GD, Karwi QG, Tian R, Wende AR, Abel ED. Cardiac energy metabolism in heart failure. Circ Res. 2021;128:1487–1513. doi: 10.1161/circresaha.121.318241 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Murashige D, Jang C, Neinast M, Edwards JJ, Cowan A, Hyman MC, Rabinowitz JD, Frankel DS, Arany Z. Comprehensive quantification of fuel use by the failing and nonfailing human heart. Science. 2020;370:364–368. doi: 10.1126/science.abc8861 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Bertero E, Maack C. Metabolic remodelling in heart failure. Nat Rev Cardiol. 2018;15:457–470. doi: 10.1038/s41569-018-0044-6 [DOI] [PubMed] [Google Scholar]
- 13. Ferrannini E, Mark M, Mayoux E. CV protection in the EMPA‐REG OUTCOME trial: a “thrifty substrate” hypothesis. Diabetes Care. 2016;39:1108–1114. doi: 10.2337/dc16-0330 [DOI] [PubMed] [Google Scholar]
- 14. Verma S, McMurray JJV. SGLT2 inhibitors and mechanisms of cardiovascular benefit: a state‐of‐the‐art review. Diabetologia. 2018;61:2108–2117. doi: 10.1007/s00125-018-4670-7 [DOI] [PubMed] [Google Scholar]
- 15. Ndumele CE, Rangaswami J, Chow SL, Neeland IJ, Tuttle KR, Khan SS, Coresh J, Mathew RO, Baker‐Smith CM, Carnethon MR, et al. Cardiovascular‐kidney‐metabolic health: a presidential advisory from the American Heart Association. Circulation. 2023;148:1606–1635. doi: 10.1161/cir.0000000000001184 [DOI] [PubMed] [Google Scholar]
- 16. Aggarwal R, Yeh RW, Joynt Maddox KE, Wadhera RK. Cardiovascular risk factor prevalence, treatment, and control in US adults aged 20 to 44 years, 2009 to March 2020. JAMA. 2023;329:899–909. doi: 10.1001/jama.2023.2307 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Puchalska P, Crawford PA. Multi‐dimensional roles of ketone bodies in fuel metabolism, signaling, and therapeutics. Cell Metab. 2017;25:262–284. doi: 10.1016/j.cmet.2016.12.022 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Shah SH, Kraus WE, Newgard CB. Metabolomic profiling for the identification of novel biomarkers and mechanisms related to common cardiovascular diseases: form and function. Circulation. 2012;126:1110–1120. doi: 10.1161/circulationaha.111.060368 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Wishart DS. Emerging applications of metabolomics in drug discovery and precision medicine. Nat Rev Drug Discov. 2016;15:473–484. doi: 10.1038/nrd.2016.32 [DOI] [PubMed] [Google Scholar]
- 20. Global burden of 369 diseases and injuries in 204 countries and territories, 1990‐2019: a systematic analysis for the Global Burden of Disease study 2019. Lancet. 2020;396:1204–1222. doi: 10.1016/s0140-6736(20)30925-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Wojcik GL, Graff M, Nishimura KK, Tao R, Haessler J, Gignoux CR, Highland HM, Patel YM, Sorokin EP, Avery CL, et al. Genetic analyses of diverse populations improves discovery for complex traits. Nature. 2019;570:514–518. doi: 10.1038/s41586-019-1310-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Popejoy AB, Fullerton SM. Genomics is failing on diversity. Nature. 2016;538:161–164. doi: 10.1038/538161a [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Jacobs B, Derese I, Derde S, Vander Perre S, Pauwels L, Van den Berghe G, Gunst J, Langouche L. A murine model of acute and prolonged abdominal sepsis, supported by intensive care, reveals time‐dependent metabolic alterations in the heart. Intensive Care Med Exp. 2025;13:6. doi: 10.1186/s40635-025-00715-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Arnett DK, Blumenthal RS, Albert MA, Buroker AB, Goldberger ZD, Hahn EJ, Himmelfarb CD, Khera A, Lloyd‐Jones D, McEvoy JW, et al. 2019 ACC/AHA guideline on the primary prevention of cardiovascular disease: a report of the American College of Cardiology/American Heart Association task force on clinical practice guidelines. Circulation. 2019;140:e596–e646. doi: 10.1161/cir.0000000000000678 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Taegtmeyer H, Young ME, Lopaschuk GD, Abel ED, Brunengraber H, Darley‐Usmar V, Des Rosiers C, Gerszten R, Glatz JF, Griffin JL, et al. Assessing cardiac metabolism: a scientific statement from the American Heart Association. Circ Res. 2016;118:1659–1701. doi: 10.1161/res.0000000000000097 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Karwi QG, Uddin GM, Ho KL, Lopaschuk GD. Loss of metabolic flexibility in the failing heart. Front Cardiovasc Med. 2018;5:68. doi: 10.3389/fcvm.2018.00068 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Bedi KC Jr, Snyder NW, Brandimarto J, Aziz M, Mesaros C, Worth AJ, Wang LL, Javaheri A, Blair IA, Margulies KB, et al. Evidence for intramyocardial disruption of lipid metabolism and increased myocardial ketone utilization in advanced human heart failure. Circulation. 2016;133:706–716. doi: 10.1161/circulationaha.115.017545 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Aubert G, Martin OJ, Horton JL, Lai L, Vega RB, Leone TC, Koves T, Gardell SJ, Krüger M, Hoppel CL, et al. The failing heart relies on ketone bodies as a fuel. Circulation. 2016;133:698–705. doi: 10.1161/circulationaha.115.017355 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Shimazu T, Hirschey MD, Newman J, He W, Shirakawa K, Le Moan N, Grueter CA, Lim H, Saunders LR, Stevens RD, et al. Suppression of oxidative stress by β‐hydroxybutyrate, an endogenous histone deacetylase inhibitor. Science. 2013;339:211–214. doi: 10.1126/science.1227166 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30. Gormsen LC, Svart M, Thomsen HH, Søndergaard E, Vendelbo MH, Christensen N, Tolbod LP, Harms HJ, Nielsen R, Wiggers H, et al. Ketone body infusion with 3‐hydroxybutyrate reduces myocardial glucose uptake and increases blood flow in humans: a positron emission tomography study. J Am Heart Assoc. 2017;6:6. doi: 10.1161/jaha.116.005066 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Hu FB. Globalization of diabetes: the role of diet, lifestyle, and genes. Diabetes Care. 2011;34:1249–1257. doi: 10.2337/dc11-0442 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Zelniker TA, Braunwald E. Mechanisms of cardiorenal effects of sodium‐glucose cotransporter 2 inhibitors: JACC state‐of‐the‐art review. J Am Coll Cardiol. 2020;75:422–434. doi: 10.1016/j.jacc.2019.11.031 [DOI] [PubMed] [Google Scholar]
- 33. Zannad F, Ferreira JP, Pocock SJ, Anker SD, Butler J, Filippatos G, Brueckmann M, Ofstad AP, Pfarr E, Jamal W, et al. SGLT2 inhibitors in patients with heart failure with reduced ejection fraction: a meta‐analysis of the EMPEROR‐reduced and DAPA‐HF trials. Lancet. 2020;396:819–829. doi: 10.1016/s0140-6736(20)31824-9 [DOI] [PubMed] [Google Scholar]
- 34. Kappel BA, Lehrke M, Schütt K, Artati A, Adamski J, Lebherz C, Marx N. Effect of empagliflozin on the metabolic signature of patients with type 2 diabetes mellitus and cardiovascular disease. Circulation. 2017;136:969–972. doi: 10.1161/circulationaha.117.029166 [DOI] [PubMed] [Google Scholar]
- 35. Selvaraj S, Patel S, Sauer AJ, McGarrah RW, Jones P, Kwee LC, Windsor SL, Ilkayeva O, Muehlbauer MJ, Newgard CB, et al. Metabolic effects of the SGLT2 inhibitor dapagliflozin in heart failure across the spectrum of ejection fraction. Circ Heart Fail. 2024;17:e011980. doi: 10.1161/circheartfailure.124.011980 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36. Santos‐Gallego CG, Requena‐Ibáñez JA, Picatoste B, Fardman B, Ishikawa K, Mazurek R, Pieper M, Sartori S, Rodriguez‐Capitán J, Fuster V, et al. Cardioprotective effect of empagliflozin and circulating ketone bodies during acute myocardial infarction. Circ Cardiovasc Imaging. 2023;16:e015298. doi: 10.1161/circimaging.123.015298 [DOI] [PubMed] [Google Scholar]
- 37. Global burden of 288 causes of death and life expectancy decomposition in 204 countries and territories and 811 subnational locations, 1990‐2021: a systematic analysis for the Global Burden of Disease study 2021. Lancet. 2024;403:2100–2132. doi: 10.1016/s0140-6736(24)00367-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38. Siskos AP, Jain P, Römisch‐Margl W, Bennett M, Achaintre D, Asad Y, Marney L, Richardson L, Koulman A, Griffin JL, et al. Interlaboratory reproducibility of a targeted metabolomics platform for analysis of human serum and plasma. Anal Chem. 2017;89:656–665. doi: 10.1021/acs.analchem.6b02930 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39. Schugar RC, Moll AR, André d'Avignon D, Weinheimer CJ, Kovacs A, Crawford PA. Cardiomyocyte‐specific deficiency of ketone body metabolism promotes accelerated pathological remodeling. Mol Metab. 2014;3:754–769. doi: 10.1016/j.molmet.2014.07.010 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40. Yurista SR, Chong CR, Badimon JJ, Kelly DP, de Boer RA, Westenbrink BD. Therapeutic potential of ketone bodies for patients with cardiovascular disease: JACC state‐of‐the‐art review. J Am Coll Cardiol. 2021;77:1660–1669. doi: 10.1016/j.jacc.2020.12.065 [DOI] [PubMed] [Google Scholar]
- 41. Kettunen J, Demirkan A, Würtz P, Draisma HH, Haller T, Rawal R, Vaarhorst A, Kangas AJ, Lyytikäinen LP, Pirinen M, et al. Genome‐wide study for circulating metabolites identifies 62 loci and reveals novel systemic effects of LPA. Nat Commun. 2016;7:11122. doi: 10.1038/ncomms11122 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42. Shin SY, Fauman EB, Petersen AK, Krumsiek J, Santos R, Huang J, Arnold M, Erte I, Forgetta V, Yang TP, et al. An atlas of genetic influences on human blood metabolites. Nat Genet. 2014;46:543–550. doi: 10.1038/ng.2982 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data S1. Supplemental Methods
Tables S1–S8
Figures S1–S7
References 41–42
STROBE Checklist