Artificial intelligence-enhanced electrocardiography derived body mass index as a predictor of future cardiometabolic disease

Obesity is a rapidly growing public health challenge and a major contributor to cardiometabolic disease. Body Mass Index (BMI) is widely used to define obesity and estimate risk, but it is an insensitive measure of visceral adiposity and fails to capture fat distribution and total fat mass, which are important determinants of cardiometabolic risk. There is a need for an accurate, accessible biomarker that reflects obesity-related cardiometabolic risk across the population and can capture downstream physiological effects, including subclinical disease, to enable primary prevention. Deep learning applied to ECG data has shown strong diagnostic and predictive capabilities for future cardiac diseases and can capture information related to non-cardiac conditions including metabolic and liver disease. Obesity is associated with ECG changes (e.g., reduced precordial voltage, T wave flattening, QT prolongation). Building on prior work such as AI-ECG-predicted age where delta age relates to mortality, the authors hypothesised that ECG signals contain information about obesity and cardiometabolic risk that can be extracted via deep learning. They trained an AI-ECG to predict BMI from ECG alone and introduced delta-BMI (AI-ECG-predicted BMI minus measured BMI) as a novel biomarker to identify individuals at risk of future cardiometabolic disease, exploring its phenotypic, genetic, metabolomic, proteomic associations, and ECG morphological correlates.

The study situates itself within growing literature on AI-ECG for prediction of future cardiac events and non-cardiac conditions. Prior work documents obesity-related ECG remodeling (reduced precordial voltages, T-wave flattening, QT prolongation). The concept of "delta age" from AI-ECG predicted age has been associated with mortality, motivating the analogous delta-BMI metric. The authors compare their BMI prediction performance favorably to previous ECG-based BMI models (e.g., Ryu et al., reported R² = 0.279) and discuss other BMI prediction approaches using complex data modalities, such as brain MRI (r ≈ 0.68), psychological variables (r ≈ 0.81), and large-scale proteomics (r ≈ 0.89), emphasizing ECG’s accessibility for clinical translation.

Study design and cohorts: A retrospective study deriving an AI-ECG BMI model in a secondary care dataset (Beth Israel Deaconess Medical Center, BIDMC) and externally validating in a population-based volunteer cohort (UK Biobank, UKB). BIDMC included 1,163,401 ECGs from 189,540 patients; 512,950 ECGs from 114,415 subjects had paired BMI and were used for model development. The dataset was patient-level split into training/validation/holdout (60%/10%/30%). UKB provided 42,386 subjects with digital ECGs and measured BMI for external validation; a 10% validation subset was used to estimate bias-correction coefficients, with the remaining 90% (n = 38,148) as holdout for analyses. ECG preprocessing: 12-lead ECGs were bandpass filtered (0.5–100 Hz), notch filtered (60 Hz), resampled to 400 Hz, and zero-padded to 4096 samples per lead for 10-second recordings. Model architecture and training: A residual neural network (ResNet) adapted from Ribeiro et al. was trained to predict continuous BMI using 10-second 8-lead ECGs (lead III and augmented leads omitted as linear combinations of leads I and II). The final layer is linear for continuous BMI output. Internal validation was performed on the 30% holdout (152,166 ECGs from 34,325 patients). Bias correction: Because delta-BMI (AI-ECG BMI minus measured BMI) was negatively correlated with measured BMI (r = −0.695), bias correction was applied using linear regression between raw delta-BMI and measured BMI in a validation subset, then adjusting predicted BMI in the holdout by subtracting the intercept and dividing by the slope. This was performed analogously in UKB (10% validation used for correction, excluded from downstream analyses). Outcomes and follow-up: Primary outcome was incident cardiometabolic disease (composite of type 2 diabetes mellitus, hypertension, lipid disorders). Secondary outcomes were incident T2DM, hypertension, and lipid disorders (definitions in Supplementary Table 3). Follow-up was censored at event or last contact. Statistical evaluation: Model performance metrics included Pearson correlation, R², and MAE with 95% CIs via bootstrap. Prognostic analyses used Cox regression adjusted for measured BMI, age, and sex, with delta-BMI evaluated as tertiles (bottom ≤ −3.74; middle −3.74 to 2.44; top > 2.44) and as a continuous variable. Stratified Cox analyses by BMI categories (18.5–24.9, ≥25, ≥30) and sex were performed; BMI < 18.5 was excluded due to small numbers. Incremental predictive utility was assessed via likelihood ratio tests (LRT), continuous net reclassification improvement (NRI), and changes in concordance index (C-index). PheWAS and phenotypic associations: In BIDMC, a disease PheWAS (1,408 phecodes with at least 100 cases) used univariate logistic regressions of delta-BMI adjusted for BMI, sex, age, age² with Bonferroni correction. In UKB, a phenome-wide analysis (1,368 phenotypes across biomarkers, imaging, and physical measures) correlated phenotypes with delta-BMI with similar adjustments and Bonferroni correction. Metabolomics (MWAS): UKB Nightingale NMR data were processed (quality control, standardization). Univariate correlations of metabolites with delta-BMI (adjusted for BMI, sex, age, age²) identified significant variables (n = 136 of 168). Stability selection with LASSO (1,000 iterations, 80% subsampling) identified 14 stable metabolites, which were analyzed via multivariable linear regression to estimate contributions to delta-BMI variability. Proteomics (PWAS): UKB Pharma Proteomics Project (Olink Explore 3072) data were processed (imputation and scaling). Univariate associations of proteins with delta-BMI (adjusted for BMI, sex, age, age²) identified 100 significant proteins; stability selection with LASSO identified 39 stable proteins, followed by multivariable linear regression to quantify contributions to delta-BMI variability. Genetics (GWAS): UKB participants of European ancestry with ECGs (n = 27,988) were included after genetic QC. Delta-BMI was rank-based inverse normal transformed. FastGWA MLM (GCTA) with covariates (age, sex, height, BMI, assessment center, 10 PCs) was used; genome-wide significance threshold p < 5×10⁻⁸. Gene-based testing used MAGMA (18,882 protein-coding genes; significance threshold p < 2.65×10⁻⁶). SNP-based heritability was estimated via GREML (GCTA). Explainability: A convolutional variational autoencoder (VAE) was trained on median ECG beats (BRAVEHEART extraction) to obtain latent factors. An XGBoost model used VAE latent factors to estimate AI-ECG BMI; feature importance was assessed via SHAP. Latent traversals visualized ECG morphology effects. Correlations between latent factors and ECG parameters (QRS axis, heart rate, PR, QRS duration, etc.) were analyzed in both cohorts.

Model performance: In BIDMC holdout, AI-ECG BMI vs. measured BMI achieved r = 0.65 (95% CI 0.65–0.66), R² = 0.43 (0.42–0.43), MAE = 3.95 (3.93–3.97). In UKB, r = 0.62 (0.62–0.63), R² = 0.39 (0.38–0.40), MAE = 2.94 (2.91–2.96). Performance was better in females (higher R²) in both cohorts; in BIDMC, predictions were more accurate for African-American compared to Caucasian individuals and less accurate for Asian participants compared to Caucasians. Prognostic value of delta-BMI: Adjusted Cox models (for measured BMI, age, sex) showed higher risk in the top tertile of delta-BMI vs. bottom:

• BIDMC: cardiometabolic disease HR 1.15 (95% CI 1.08–1.23, p < 0.001); T2DM HR 1.25 (1.18–1.31, p < 0.001); hypertension HR 1.26 (1.19–1.33, p < 0.001); lipid disorders HR 1.21 (1.15–1.27, p < 0.001).
• UKB: cardiometabolic disease HR 1.58 (1.41–1.76, p < 0.001); T2DM HR 2.28 (1.76–2.96, p < 0.001); hypertension HR 1.54 (1.37–1.75, p < 0.001); lipid disorders HR 1.56 (1.32–1.85, p < 0.001). Associations held across BMI categories and sex; continuous delta-BMI was also significant. Incremental predictive utility: Adding delta-BMI significantly improved model fit (LRT) and reclassification (continuous NRI) for cardiometabolic disease and components in both cohorts. Discriminatory power (C-index) improved modestly but significantly; notable C-index gains in BIDMC for BMI > 25 and > 30 (AC-index ≈ 0.0098 and 0.0362). For T2DM, the largest improvements in model fit and NRI were observed. Among BIDMC outpatients with BMI 18.5–24.9, C-index improvement was 0.0239 (0.0053–0.0377); in UKB, 0.0377 (0.0036–0.0751). Delta-BMI improved model fit and reclassification for cardiometabolic disease in Caucasians and Hispanics and for T2DM in Caucasians and Asians (BIDMC). Phenotypic associations: In BIDMC PheWAS (1,408 phecodes), 55 (3.9%) reached Bonferroni significance, including T2DM, hypertension, hypertensive heart disease, insulin/oral hypoglycemic use, hyperlipidemia, hypercholesterolemia, liver disease, and ASCVD. In UKB (1,368 phenotypes), 231 (16.9%) significant associations included biomarkers (triglycerides, ALT, apolipoprotein B, high light scatter reticulocyte percentage), physical measures (abdominal fat ratio, diastolic BP, waist circumference), imaging (abdominal subcutaneous adipose tissue volume, bone mineral density), and cardiac MRI (pericardial fat mass). Negative associations were observed with sex hormone-binding globulin (SHBG) and HDL cholesterol. Metabolomics: Of 168 metabolites, 136 were significant in MWAS; 14 were stably selected by LASSO. Multivariable regression showed negative associations with glutamine, citrate, glycine, and cholesteryl esters in very large HDL; positive associations with omega-3 fatty acids, valine, glucose, glycoprotein acetyls, total lipids in small HDL, and triglycerides in very large VLDL. Metabolites explained R² = 0.046 of delta-BMI variability. Proteomics: Of 2,919 proteins, 100 were significant in PWAS; 39 were stably selected. Multivariable regression showed negative associations with CLPS (colipase), C9, ADIPOQ (adiponectin), HRC, and VCAN; positive associations with STX3 and CNDP1. Proteins explained R² = 0.098 of delta-BMI variability. Genetics: GWAS identified a significant locus near SCN10A and a borderline locus near CASC20. Gene-level associations included RXRG, EXOG, and SCN5A. SNP-heritability of delta-BMI was 0.109 (95% CI 0.083–0.135). Explainability: VAE-derived latent factors and XGBoost approximation (r ≈ 0.77–0.78; R² ≈ 0.61 in both cohorts) indicated that QRS axis (latent factors 50, 6), resting heart rate (factor 43), PR interval (factor 16), and QRS duration (factors 6, 31, 16) are important for AI-ECG BMI predictions, aligning with known obesity-related ECG changes.

The AI-ECG model accurately predicts BMI from standard ECG signals and generalizes across distinct populations (secondary care and community volunteers), sexes, and major ethnic groups, with particularly stronger performance in females. Crucially, the delta-BMI measure provides additive prognostic information beyond measured BMI, identifying individuals at increased risk for future cardiometabolic disease (including T2DM, hypertension, and dyslipidemia) regardless of BMI category. Incorporating delta-BMI improves model fit and reclassification across cohorts and yields modest gains in discrimination, with notable benefits in subgroups such as higher BMI categories and normal-weight individuals. These findings suggest clinical utility for integrating AI-ECG BMI and delta-BMI into routine care, where ECGs are already collected, to enhance screening and monitoring (e.g., targeted blood pressure checks, HbA1c and lipid testing) and to support behavior change through personalized risk feedback. Biological and morphological analyses support plausibility: delta-BMI associates with truncal adiposity, pericardial fat, adverse lipid and metabolite profiles, and proteins involved in metabolic regulation, and shows genetic links to loci influencing ECG conduction and adiposity traits. Explainability via VAE/XGBoost links model predictions to ECG parameters (QRS axis, PR interval, heart rate, QRS duration), consistent with known obesity-related cardiac electrophysiology. Together, these results strengthen confidence in the model’s interpretability and its relevance to cardiometabolic risk pathways.

An AI-enhanced ECG model can accurately predict BMI using routine ECG data. Introducing delta-BMI (AI-ECG-predicted minus measured BMI) provides additional prognostic insight for future cardiometabolic disease, independent of measured BMI. Phenotypic, metabolomic, proteomic, and genetic analyses elucidate plausible biological pathways underpinning delta-BMI and its risk associations, while model explainability links predictions to interpretable ECG features. Clinically, these tools could augment screening strategies, enable more personalized risk assessment, and motivate lifestyle interventions. Future work should validate these findings in more ethnically diverse and representative populations, assess clinical implementation pathways, and evaluate prospective impact on outcomes.

There was an approximately 8-year temporal gap between the collection of metabolomic/proteomic data and ECG acquisition in UK Biobank, potentially introducing noise into biomarker-ECG relationships. The UK Biobank cohort’s healthy volunteer bias and limited ethnic diversity may restrict generalizability to broader populations. Subgroup analyses by ethnicity in BIDMC were limited by small sample sizes, warranting further evaluation in larger and more diverse datasets.