Skip to content

Multimodal Cardiovascular Risk Prediction

Summary

Multimodal cardiovascular risk prediction combines conventional clinical variables with metabolomic biomarkers and polygenic risk scores. The approach tests whether molecular and inherited information add complementary discrimination and calibration beyond established clinical models. Ritchie, Jiang, Pennells, et al. (2026) provide the largest population-health assessment to date of this question: in 297,463 UK Biobank participants (8,919 incident CVD cases over a median 10.0-year follow-up), adding 11 clinical biomarkers, NMR metabolomic biomarker scores, and CHD/stroke polygenic risk scores to the ESC-recommended SCORE2 model each individually improved risk discrimination, and combining all three sources of information yielded the largest gain (ΔC-index 0.024) — with population-scale modeling suggesting this could increase the number of CVD events prevented per 100,000 screened from 229 to 413 without materially changing the number of statins prescribed per event prevented.

Model Design

The reported study evaluated clinical, metabolomic and polygenic predictors jointly rather than treating each as a replacement for conventional risk factors. This structure is important because genetic scores capture lifelong inherited susceptibility, metabolites reflect current physiology, and clinical variables summarize established risk burden.[1]

Incremental performance must be judged using external validation, calibration and clinically meaningful reclassification, not association strength alone.

Study Design and Cohort

  • Population: 297,463 UK Biobank participants aged 40–69, eligible for SCORE2 primary CVD risk assessment (no prior atherosclerotic CVD, diabetes, chronic kidney disease, familial hypercholesterolemia, or lipid-lowering treatment at baseline), drawn from 502,207 total UK Biobank enrollees.
  • Outcome: first-onset CVD (fatal/non-fatal, matching the SCORE2 working group's ICD-10 definition spanning ischaemic heart disease, cerebrovascular disease, heart failure, arrhythmia, and related fatal events) within 10 years; 8,919 cases recorded across 2,897,497 person-years, median follow-up 10.0 years.
  • Discovery/replication split: a 167,517-participant discovery cohort (5,096 cases, NMR data released July 2023) and a 128,946-participant replication cohort (3,823 cases, early-access NMR data from Q3 2024), randomized by UK Biobank independent of phenotype.
  • Biomarker panels tested: 249 NMR metabolomic biomarkers and 28 clinical chemistry biomarkers, each individually screened for FDR-significant improvement over SCORE2 (sex-stratified Cox models, biomarker as covariate with SCORE2 as offset), then combined into NMR metabolomic biomarker scores (machine-learning-trained) and a multivariable clinical biomarker model.
  • PRS: previously validated PGS Catalog scores for coronary heart disease (PGS000018) and ischaemic stroke (PGS000039).
  • Seven compared models: SCORE2 alone, plus SCORE2 + {clinical biomarkers | NMR scores | PRS | clinical+PRS | NMR+PRS | NMR+clinical | NMR+clinical+PRS}, all fit in discovery and evaluated in replication.

Primary Results

  • Baseline: SCORE2 alone achieved C-index=0.719.
  • Individual additions: clinical biomarkers ΔC-index=0.014 (95% CI 0.012–0.015); NMR metabolomic scores ΔC-index=0.010 (0.009–0.012); PRS ΔC-index=0.009 (0.008–0.011) — each independently significant.
  • Combined (clinical + NMR + PRS): the largest gain, ΔC-index=0.024 (0.022–0.027), roughly matching the sum of the individual gains and indicating the three modalities carry substantially non-redundant information.
  • Reclassification: net case reclassification of 16.66% (15.50–17.81%) under ESC 2021 categorical risk thresholds when all three biomarker sources were added to SCORE2.
  • Population health modeling: applying the combined biomarker panel to targeted re-screening of SCORE2-medium-risk individuals was projected to increase CVD events prevented from 229 to 413 per 100,000 screened (ΔCVD prevented = 184, 95% CI 174–194), while keeping the number of statins prescribed per CVD event prevented essentially unchanged — i.e., the improvement in detection did not come at the cost of substantially more unnecessary prescribing.

Key Implementations

  • CardiOmicScore: a UK Biobank-scale multitask deep learning framework that adds proteomics (2,920 proteins) alongside metabolomics (168 metabolites) to predict six cardiovascular diseases (CAD, stroke, heart failure, atrial fibrillation, peripheral artery disease, venous thromboembolism) rather than one. Its proteomic score (ProScore) and metabolomic score (MetScore) both add significant discrimination over clinical-only Cox models (ΔC-index up to 0.102), with proteomics contributing more than metabolomics across all six diseases — extending this concept's clinical-metabolomic-polygenic framing to a broader, multi-outcome, multi-omics setting.

Citations

[1] Ritchie, S.C., Jiang, X., Pennells, L., Xu, Y., Coffey, C., Liu, Y., Gibson, J.T., Surendran, P., Karthikeyan, S., Lambert, S.A., Danesh, J., Butterworth, A.S., Wood, A., Kaptoge, S., Di Angelantonio, E., & Inouye, M. (2026). Combined clinical, metabolomic, and polygenic scores for cardiovascular risk prediction. European Heart Journal, 47, 1861–1873. [2] Luo, Y., Zhang, N., Yang, J., Cui, M., Tsoi, K.K.F., Lip, G.Y.H., Liu, T., & Zhang, Q. (2026). AI-based multiomics profiling reveals complementary omics contributions to personalized prediction of cardiovascular disease. Nature Communications, 17, 2269.. Source paper: s41467-026-68956-6.pdf