Moral approvals
For the Beth Israel Deaconess Medical Middle (BIDMC), cohort ethics evaluate and approval had been supplied by the Beth Israel Deaconess Medical Middle Committee on Scientific Investigations, IRB protocol # 2023P000042.
The UK Biobank has approval from the North West Multi-Centre Analysis Ethics Committee as a Analysis Tissue Financial institution (utility IDs 48666, 47602).
ECG datasets
As this was a retrospective research, no a priori pattern dimension calculations had been carried out. Lacking knowledge was dealt with by complete-case evaluation.
(i) The BIDMC cohort
The Beth Israel Deaconess Medical Middle (BIDMC) cohort is a secondary care dataset consisting of routinely collected knowledge from Boston, USA. Topics over 16 years previous with a legitimate ECG carried out from 2014 to 2023 had been included. Prior ECGs again to 2000 had been included for these topics. BMI was derived from contemporaneous weight measurements acquired inside a 30-day window of every ECG recording, complemented by peak assessments carried out inside 1 yr of the respective ECG. Diagnostic Worldwide Classification of Ailments (ICD) codes had been used to find out illness standing. Topics had been censored on the time of end result or final in-person hospital contact.
(ii) The UK Biobank Cohort
The UK Biobank is a longitudinal research of over 500,000 volunteers aged 40–69 on the time of enrolment in 2006–201061. At baseline evaluation, members supplied info on well being and life-style through questionnaire, had bodily measures taken (together with peak, weight, and blood stress), and donated samples of blood urine and saliva. A subgroup of members was invited again for subsequent visits for added investigations, together with cardiac magnetic resonance imaging (MRI), mind MRI, and digital ECGs. 42,386 topics with digital ECGs taken on the occasion 2 go to had been accessible for evaluation. The collected knowledge embody medical, metabolomic, proteomic, and genomic knowledge, and had been linked to most cancers and demise registries, hospital admissions, and first care data. There may be proof of wholesome volunteer choice bias18. Outcomes had been linked to most cancers and demise registry knowledge, hospital admissions, and first care data. Detailed phenotyping utilizing the cardiac MRI knowledge has been beforehand described62,63.
ECG pre-processing
The 12-lead ECGs from each cohorts had been pre-processed utilizing a bandpass filter at 0.5–100 Hz, a notch filter at 60 Hz, and re-sampled to 400 Hz. Zero padding was used to attain a sign with 4096 samples for every lead for a ten s recording.
Mannequin growth
The mannequin was derived utilizing the BIDMC cohort. The BIDMC cohort consists of 1,163,401 ECGs from 189,540 sufferers, of which 512,950 (44.1%) ECGs from 114,415 topics had paired BMI knowledge accessible. To forestall knowledge leakage, the dataset was divided by affected person ID utilizing a 60/10/30 break up into derivation, validation, and holdout check units, respectively. The ECG-AI mannequin, which employed a ResNet structure tailored from Ribeiro et al.17, was skilled utilizing 10-second 8-lead ECGs. Particularly, lead III and the augmented leads had been omitted from the mannequin as they’re linear mixtures of leads I and II. The structure incorporates a linear operate within the closing layer for BMI prediction. The mannequin was internally validated utilizing the 30% holdout check set (152,166 ECGs from 34,325 sufferers).
Exterior validation
We externally validated the mannequin utilizing the UKB dataset. The unique 42,386 ECGs had been divided into validation and holdout check units utilizing a break up ratio of 10/90%. The validation set was used to derive the bias-correction coefficients, which had been then used to regulate the BMI estimates of the holdout check set. The holdout check set (n = 38,148) was used for subsequent downstream analyses.
End result definition
BMI for the BIDMC cohort was calculated utilizing affected person peak inside a yr and weight inside 30 days of ECGs. For the UKB cohort, BMI was measured concurrently with ECGs. The first survival evaluation end result was incident cardiometabolic illness. The secondary outcomes had been incident sort 2 diabetes, hypertension, and lipid issues (see end result definitions in Supplementary Desk 3). Comply with-up durations had been adjusted by censoring at lack of follow-up or occasion incidence.
Bias correction
Delta-BMI was negatively correlated with BMI (Pearson r = −0.695, p ≈ 0.0), thus people with increased BMI demonstrated decrease delta-BMI. This phenomenon, often known as the ‘regression dilution’ impact noticed in earlier fashions predicting organic age, underscores the necessity for bias correction to mitigate measurement error. Primarily based on earlier research33,64,65,66, we addressed the correlation between delta-BMI and BMI within the holdout set by becoming a linear regression between the uncooked uncorrected delta-BMI and measured BMI within the validation dataset. Then, we adjusted the anticipated BMI within the holdout set by subtracting the intercept and dividing it by the slope of the linear regression mannequin66. This was repeated for the UKB cohort, utilizing 10% of the dataset as a validation cohort to derive the regression mannequin coefficients. This 10% validation cohort was excluded from subsequent downstream analyses. See Supplementary Fig. 1 for extra particulars.
Survival evaluation and statistical analyses
The AI-ECG BMI mannequin was assessed utilizing Pearson correlation, R2, and MAE, with 95% CIs derived from 1000 bootstrap iterations on 80% of the holdout datasets. To guage the prognostic significance of delta-BMI within the BIDMC and UK Biobank cohorts, we carried out Cox regression, adjusting for BMI, age, and intercourse, and reported HRs and 95% CIs. Delta-BMI was analysed each as steady and categorical variables. For the specific evaluation, delta-BMI was stratified into three tertiles: backside (≤−3.74), center (−3.74 to 2.44), and high (>2.44). Cox regression analyses had been additionally carried out for stratified intercourse and BMI teams (18.5–24.9, ≥25, ≥30). On account of inadequate numbers, BMI n = 794, UKB: n = 258). In line with latest proof towards the need of proportional hazards testing in well-powered medical datasets, we didn’t assess this assumption, treating the Cox mannequin’s hazard ratio as a mean estimate over the follow-up interval67.
Subsequent, we carried out an exploratory evaluation to judge the incremental predictive utility of delta-BMI in predicting future cardiometabolic illness and its elements. Firstly, we carried out probability ratio checks (LRT) to gauge the importance of mannequin enchancment upon incorporating delta-BMI. Secondly, we computed the continual internet reclassification index (NRI) to quantify the extent to which delta-BMI enhances danger stratification. Lastly, we examined adjustments within the concordance index upon the inclusion of delta-BMI into the Cox mannequin, offering insights into its discriminatory energy.
Phenome-wide affiliation research (PheWAS)
We carried out a phenome-wide affiliation research (PheWAS) within the BIDMC cohort to find out delta-BMI’s illness associations, utilizing univariate logistic regressions on 1408 illness phecodes. Equally, within the UKB cohort, we analysed 1368 medical phenotypes consisting of affected person measurements, surveys and investigations, with delta-BMI, utilizing univariate correlations. Changes had been made for BMI, intercourse, age, and age2, in addition to Bonferroni correction to account for a number of testing.
Genome-wide affiliation research (GWAS)
To determine genetic associations with delta-BMI, we carried out a genome-wide affiliation research (GWAS) within the UKB. Utilizing linear regression, delta-BMI was adjusted for the next covariates: age on the imaging go to, intercourse, peak, BMI, imaging evaluation centre, and the primary 10 genetic principal elements. The UKB members included within the genetic analyses had been chosen for European ancestry, missingness charge of SNPs p −8 or
The GWAS was carried out with the FastGWA MLM applied by the Genome-wide Advanced Trait Evaluation (GCTA) software program utilizing a genetic relationship matrix (GRM) to regulate for inhabitants construction68. The delta-BMI distributions had been normalised by rank-based inverse regular remodel previous to the evaluation. Age, intercourse, peak, BMI, the UKB evaluation centre, and the primary 10 genetic principal elements had been included as covariates. We report the SNPs which had been recognized by the traditional genome-wide significance threshold of p-value −8.
The genetic variance defined by genome-wide SNPs (SNP-based heritability) was calculated with the genomic-relatedness-based restricted most probability (GREML) evaluation utilizing the GCTA software program69. Genetic correlation was calculated with the bivariate GREML evaluation methodology70. The QQ plots for GWAS abstract statistics and gene-based check are proven in Supplementary Fig. 5.
Metabolome-wide affiliation research (MWAS)
Over 250,000 UKB members have been profiled by Nightingale Well being to acquire their nuclear magnetic resonance (NMR) EDTA plasma pattern metabolic biomarkers71. After extracting absolute measures and eliminating entries with over 40% missingness for rows and 20% for columns, we retained knowledge for 274,350 members from occasion 0. We subsequently adjusted and standardised NMR metabolite readings for spectrometer variations and excluded outliers exceeding 4 interquartile ranges71. The ultimate dataset, following filtering for full circumstances, included 22,322 UK Biobank members with each NMR metabolite profiles and 12-lead ECGs (UKB-NMRmet-ECG). To research the associations between delta-BMI and UKB NMR metabolites knowledge, we initially carried out univariate correlation analyses of particular person metabolites with delta-BMI, adjusted for BMI, intercourse, age, and age2, with a significance threshold of two.976×10−4. Secondly, we employed stability choice with the least absolute shrinkage and choice operator (LASSO) regression to determine steady predictors of delta-BMI (R package deal Sharp, model 1.4.572). The soundness choice with LASSO was carried out over 1000 iterations with subsampling of 80% of the full UKB-NMRmet-ECG dataset, with lambda constrained to 1 ×10−8 and 5 ×10−1. The calibration of π (choice proportion) and λ (L1 penalty issue) is proven in Supplementary Fig. 2. The mix of parameters ensuing within the highest stability rating was chosen (π = 0.650 and λ = 0.142), thus acquiring 14 stably chosen predictors. These predictors had been then utilized in a multivariate linear regression to disentangle the metabolomic associations with delta-BMI, as seen in Fig. 6c.
Proteomic-wide affiliation research (PWAS)
53,030 UKB members had their plasma proteomic profiles analysed by the Pharma Proteomics Undertaking36 at occasion 0 utilizing the antibody-based Olink Discover 3072 PEA, measuring 2923 protein analytes36. Observations with lower than 40% missingness and protein analytes with lower than 20% missingness had been retained, leaving 2919 analytes for additional evaluation. Lacking values had been imputed utilizing easy imply imputation and scaled (sci-kit-learn, model 1.1.1) (Supplementary Fig. 3). In complete, 3512 sufferers had each proteomic knowledge and 12-lead ECGs (UKB-PPP-ECG). To research the associations between delta-BMI and the UK Biobank plasma proteomic knowledge, we mirrored the metabolomic method. Initially, we carried out a univariate regression evaluation with delta-BMI as end result and protein analytes as predictors, adjusted for BMI, intercourse, age, and age2, with a significance threshold of 1.7123 ×10−5. Secondly, we carried out stability choice with LASSO regression to determine stably chosen protein predictors of delta-BMI. Variable choice utilizing the Sharp package deal was carried out over 1000 iterations with a subsampling of 80% of the full UKB-PPP-ECG dataset. The calibration of π and λ is depicted in Supplementary Fig. 4. The mix of parameters with the very best stability rating was chosen (π = 0.560 and λ = 0.107), thus figuring out 39 stably chosen predictors. The steady predictors had been then utilized in multivariate linear regression to disentangle the proteomic associations with delta-BMI, as seen in Fig. 7c.
Explainability
To grasp the ECG morphologies related to the AI-ECG BMI predictions, we skilled a VAE, as beforehand described73, utilizing median ECG beats. Median beats had been extracted utilizing the BRAVEHEART software program, as beforehand described74. The VAE was based mostly on a convolutional encoder/decoder structure, which was impressed by architectures beforehand used for ECG evaluation75,76. Particularly, the encoder comprised of six convolutional blocks of function extraction, adjusted for the median beat ECG sign. The decoder structure was designed as a symmetrically inverse community of the encoder. The full variety of parameters was 1,533,888 and 1,283,976 for the encoder and decoder, respectively. An in depth depiction of the networks’ structure could be seen in Supplementary Fig. 6.
Utilizing the identical knowledge break up as for the AI-ECG BMI mannequin, we then skilled an eXtreme Gradient Boosting (XGBoost) mannequin utilizing the VAE latent options with AI-ECG BMI because the output. Within the BIDMC holdout set, the XGBoost mannequin achieved a Pearson correlation coefficient of 0.77 (95% CI: 0.77–0.78) and R2 of 0.61 (95% CI: 0.60–0.61) (Supplementary Fig. 7a). The mannequin was then validated within the UK Biobank, the place it achieved a Pearson correlation coefficient of 0.78 (95% CI: 0.78–0.79) and R2 of 0.61 (95% CI: 0.600–0.62) (Supplementary Fig. 7b). The highest 5 most essential options, as assessed by SHAP values (Fig. 9a), had been visualised by latent traversals73 (Fig. 9b) and cross-correlated with identified ECG parameters (Fig. 9c, d).
Reporting abstract
Additional info on analysis design is on the market within the Nature Analysis Reporting Abstract linked to this text.
