Obesity is a major global health concern, significantly contributing to the rising prevalence of cardiometabolic diseases. Body Mass Index (BMI), while commonly used to assess obesity and associated risks, has limitations. It's an insensitive measure of visceral adiposity and doesn't account for body fat distribution, crucial factors in determining cardiometabolic risk. This inadequacy highlights the need for a more accurate and accessible biomarker that can effectively capture the risk of cardiometabolic diseases linked to obesity, even in subclinical stages, enabling proactive preventive interventions. Recent advancements in deep learning applied to electrocardiograms (ECGs) have demonstrated remarkable diagnostic and predictive capabilities in identifying various cardiac and non-cardiac conditions. Obesity is known to cause alterations in cardiac electrophysiology, leading to observable changes in ECGs, including reduced voltage in precordial leads, T-wave flattening, and QT interval prolongation. Leveraging the potential of AI-enhanced ECGs (AI-ECGs), this study explores whether deep learning can extract valuable information from ECGs to predict BMI and assess cardiometabolic risk. Inspired by the success of AI-ECG predicted age, where the difference between predicted and chronological age (“delta age”) is linked to mortality, this research proposes “delta-BMI” (the difference between AI-ECG predicted BMI and measured BMI) as a novel biomarker for cardiometabolic risk. This approach aims to identify individuals at risk of future cardiometabolic diseases, including diabetes, hypertension, and lipid disorders. The study further investigates the underlying biological pathways associated with delta-BMI by examining phenotypic, genetic, metabolomic, and proteomic associations, and uses a variational autoencoder to analyze ECG morphological changes associated with AI-ECG BMI predictions. The overall goal is to establish a non-invasive method for improved cardiometabolic risk stratification.

The existing literature highlights the limitations of BMI as a sole indicator of cardiometabolic risk. Studies have shown that BMI fails to capture the critical aspects of visceral adiposity and fat distribution, which are strongly linked to the development of cardiometabolic diseases. While BMI provides a general indication of weight status, it lacks the precision needed for accurate risk stratification. The application of AI to ECG analysis has emerged as a promising field, with numerous studies showcasing its capability in predicting future cardiac events and identifying various cardiac abnormalities. AI-ECG algorithms have proven to be adept at extracting subtle features from ECGs that might not be detectable by traditional methods. However, research exploring the use of AI-ECG to predict BMI and assess cardiometabolic risk beyond traditional measures is relatively limited. Previous studies have demonstrated that AI can extract information from ECGs related to other non-cardiac conditions, such as metabolic disorders and liver disease, laying the groundwork for the current investigation.

This study employed a two-cohort approach, utilizing the Beth Israel Deaconess Medical Center (BIDMC) cohort as the derivation cohort and the UK Biobank (UKB) cohort for external validation. **BIDMC Cohort:** This secondary care dataset comprised 512,950 12-lead ECGs from 114,415 subjects with paired BMI data. The dataset was split into training (60%), validation (10%), and holdout test sets (30%). **UK Biobank Cohort:** This healthy volunteer cohort provided digital ECGs and BMI measurements from 42,386 subjects, offering an independent validation set. **AI-ECG BMI Model:** A residual neural network (ResNet) architecture was used to train the AI-ECG model to predict BMI from the ECG data. The model's performance was evaluated using Pearson correlation coefficient, R², and mean absolute error (MAE). **Delta-BMI Calculation:** Delta-BMI was calculated as the difference between AI-ECG predicted BMI and measured BMI. Cox regression analyses were performed to assess the association between delta-BMI (categorized into tertiles and as a continuous variable) and the risk of future cardiometabolic disease (a composite of type 2 diabetes mellitus, hypertension, and lipid disorders), adjusting for measured BMI, age, and sex. **Biological Plausibility Analyses:** To explore the underlying biological mechanisms, the study conducted phenome-wide association studies (PheWAS) in both cohorts, examining associations between delta-BMI and a wide range of clinical phenotypes. Metabolomic and proteomic profiling analyses were performed using data from the UK Biobank to identify metabolites and proteins associated with delta-BMI. A genome-wide association study (GWAS) in the UK Biobank investigated genetic associations with delta-BMI. Finally, a variational autoencoder was used to analyze ECG morphological changes associated with AI-ECG BMI predictions. **Statistical Analyses:** Various statistical methods were employed, including Pearson correlation, R², MAE, Cox regression, likelihood ratio tests, net reclassification indices, and concordance indices, to assess the model's performance and the associations between delta-BMI and cardiometabolic outcomes.

The AI-ECG BMI model demonstrated strong performance in both the BIDMC and UK Biobank cohorts, achieving Pearson correlation coefficients of 0.65 and 0.62, and R² values of 0.43 and 0.39, respectively. Delta-BMI emerged as a significant predictor of future cardiometabolic disease, independently of measured BMI. In the BIDMC holdout dataset, individuals in the top tertile of delta-BMI showed a 15% increased risk of cardiometabolic disease (HR 1.15, p<0.001), with similar increases in risk observed for type 2 diabetes mellitus (HR 1.25, p<0.001), hypertension (HR 1.26, p<0.001), and lipid disorders (HR 1.21, p<0.001). The UK Biobank cohort showed even stronger associations, with a 58% increased risk of cardiometabolic disease (HR 1.58, p<0.001) and a more than twofold increased risk of type 2 diabetes mellitus (HR 2.28, p<0.001) in the top delta-BMI tertile. The inclusion of delta-BMI significantly improved model fit, reclassification, and discriminatory power for predicting cardiometabolic disease and its components in both cohorts. Phenotypic profiling revealed associations between delta-BMI and various cardiometabolic diseases, anthropometric measures of truncal obesity, and pericardial fat mass. Metabolic profiling identified significant associations with several metabolites, including valine, lipids in small HDL, and triglycerides. Proteomic analysis revealed associations with proteins such as syntaxin-3 and carnosine dipeptidase 1. GWAS identified associations with genes involved in regulating cardiovascular/metabolic traits, such as SCN10A, SCN5A, EXOG, and RXRG. Analysis using a variational autoencoder highlighted the importance of QRS axis, resting heart rate, PR interval, and QRS duration in AI-ECG BMI predictions.

This study successfully developed and validated an AI-ECG model capable of accurately predicting BMI, demonstrating its potential as a novel tool for assessing cardiometabolic risk. The identification of delta-BMI as a significant predictor of future cardiometabolic disease, independent of measured BMI, provides valuable additional prognostic information. The findings underscore the importance of considering factors beyond traditional BMI in risk assessment. The significant improvement in model performance by incorporating delta-BMI highlights its additive value in risk stratification. The consistent findings across two diverse cohorts, a secondary care dataset and a healthy volunteer cohort, enhance the generalizability of the results. The exploration of biological pathways through phenome-wide, metabolomic, proteomic, and genomic analyses contributes to a deeper understanding of the underlying mechanisms linking delta-BMI to cardiometabolic risk. These findings can inform the development of more targeted and effective preventive interventions. The model's explainability analyses provide further insights into the biological plausibility of the AI-ECG BMI predictions. The identified ECG morphological features related to delta-BMI are consistent with prior research on obesity-related cardiac changes. However, the study's limitations, such as the temporal gap between metabolomic/proteomic data collection and ECG acquisition in the UK Biobank, warrant consideration when interpreting the results.

This study demonstrates the utility of an AI-ECG model in accurately predicting BMI and introduces delta-BMI as a novel, non-invasive biomarker for cardiometabolic risk stratification. The model's strong performance and the significant association between delta-BMI and future cardiometabolic disease suggest its potential for improving clinical risk assessment and informing preventive strategies. Future research should focus on larger, more diverse populations to further validate these findings and explore the clinical implications of delta-BMI in various settings. Further investigation into the identified biological pathways could also lead to the development of targeted therapies for cardiometabolic disease.

The study acknowledges several limitations. The temporal gap between the collection of metabolomic and proteomic data and ECGs in the UK Biobank introduces potential noise into the analysis of relationships between these markers and electrocardiographic outcomes. The UK Biobank cohort's limited ethnic diversity and healthy volunteer bias may affect the generalizability of findings to more diverse populations. Further research in larger, more diverse datasets is needed to fully evaluate the implications of the findings, particularly those related to ethnic subgroups.