Artificial intelligence-driven metabolomics of retinal nerve fibre layer to profile risks of mortality and cardiometabolic diseases

Abstract

Retinal nerve fibre layer (RNFL) is a non-invasive structural biomarker of cardiometabolic health, yet its biological underpinnings remain unknown. Here, we integrate advanced retinal optical biopsy and artificial intelligence (AI) algorithms with two complementary metabolomic assays across ethnically diverse cohorts to elucidate the metabolic basis underlying RNFL degeneration and its link to cardiometabolic disease (CMD) in Western cohort and Eastern cohort (Guangzhou Diabetic Eye Study, GDES). We identify 26 metabolic biomarkers significantly associated with RNFL thickness, most of which (ranging from 19 to 26) are linked to HDL composition and lipid transport, mediating a substantial proportion of the RNFL-CMD association (e.g., 63.7% for type 2 diabetes and 44.7% for myocardial infarction). AI-driven RNFL metabolic state model stratifies CMD risk with up to 21.8-fold enrichment between risk deciles and augments prediction while translating into clinical utility across genetic and demographic strata, particularly within socially vulnerable populations. This integrated approach highlights RNFL metabolic states as a shared basis underlying retinal-cardiometabolic connections and as early indicators that inform equitable CMD management.

Fig. 1. Overview of the study design.

a The eligible study population was categorized into three distinct groups: population-I, for identifying RNFLT metabolic states; population-II, for unravelling revelations on CMD outcomes; and population-III, for independent concept validation. b To identify RNFLT metabolic states, we conducted retinal scanning and utilized two complementary metabolomic assays. Genotyping was performed to assess genetic susceptibility in individuals. ML algorithms were employed for comprehensive model construction and evaluation. c For CMD outcome risk modelling, the study populations were randomized into training and testing sets. Dataset balancing techniques were applied before feature selection and model training. d The outcomes examined in this study include incident T2D, myocardial infarction, heart failure, stroke, all-cause mortality, and CMD mortality. e Distinct risk stratification and improved predictability and clinical utility were observed for all studied outcomes. f Special attention was given to extending the benefits to women and socially vulnerable communities. g Comprehensive sets of predictors commonly used in the CMD primary prevention were incorporated as benchmark models. Parts of panels a–d and f were created from BioRender (https://BioRender.com/yu7axft) and Flaticon (https://flaticon.com). RNFLT retinal nerve fibre layer thickness, CMD cardiometabolic disease, UKB UK Biobank, GDES Guangzhou Diabetes Eye Study, T2D type 2 diabetes, FGCRS Framingham General Cardiovascular Risk Score, SCORE2 Systematic Coronary Risk Evaluation 2, WHO-CVD World Health Organization Cardiovascular Disease, AHA-ASCVD American Heart Association-Atherosclerotic Cardiovascular Disease, UKPDS UK Perspective Diabetes Study, NZ-DCS New Zealand Diabetes Cohort Study, WAN Wan’s model, BMI body mass index, eGFR estimated glomerular filtration rate, HDL-c high-density lipoprotein cholesterol.

Fig. 2. RNFLT metabolic state profile stratifies CMD outcome risk.

a Attribution of metabolic biomarkers to the outcome-specific ML-driven RNFLT metabolic state (n = 17,203). Individual attributions are aggregated by percentiles, with each dot representing one percentile. The distance of a dot from the circular baseline reflects the strength of the absolute attribution for that percentile. Deviations towards the centre and periphery indicate negative and positive contributions, respectively. Brick colours denote outcomes, while dot colours represent the normalized values for each metabolic biomarker. b–g Cumulative event rates throughout the observation period for CMD outcomes (n = 17,203), stratified by RNFLT metabolic state quantiles. Data are presented as observed event frequencies with 95% CIs shown as shading derived from survival proportions. Red represents the top decile, yellow represents the middle and blue represents the bottom. Illustrations of uniform risk scales are provided in Supplementary Figs. S1 and S2. Source data are provided as a Source Data file. RNFLT retinal nerve fibre layer thickness, ML machine learning, CMD cardiometabolic disease, HDL high-density lipoprotein, VLDL very low-density lipoprotein, T2D type 2 diabetes.

Fig. 3. Improvement in predictive performance across baseline models and genetic susceptibility incorporating RNFLT metabolic state.

a–f Comparison of model performance, including the Age&Sex model, established models, and models incorporating RNFLT metabolic states to predict CMD outcomes, estimated using AUPRC (n = 17,203). Data are presented as estimated coefficients (dots) with 95% CIs indicated by error bars. Different colours denote distinct models, with horizontal dashed lines indicating the performance benchmarks set by the Age&Sex model and four established models. g Comparison of the added benefits derived from integrating RNFLT metabolic states into established models for predicting CMD outcomes across varying genetic susceptibility (n = 17,203). Data are presented as estimated coefficients (dots) with 95% CIs indicated by error bars. Source data are provided as a Source Data file. CMD cardiometabolic disease, AUPRC area under precision-call curve, RNFLT retinal nerve fibre layer thickness, RNFLT MET RNFLT metabolic state, T2D type 2 diabetes, FGCRS Framingham General Cardiovascular Risk Score, SCORE2 Systematic Coronary Risk Evaluation 2, WHO-CVD World Health Organization Cardiovascular Disease, AHA-ASCVD American Heart Association-Atherosclerotic Cardiovascular Disease.

Fig. 4. Model calibration and clinical utility incorporating RNFLT metabolic state.

a–f Calibration for Age&Sex (dotted line), the combination of Age&Sex with RNFLT metabolic states (sky blue), FGCRS (dashed line), and the combination of FGCRS with RNFLT metabolic states (red) in predicting CMD outcomes (n = 17,203). Flesh indicates optimal calibration. g–l Net benefit of clinical utility standardized by endpoint prevalence (n = 17,203), with horizontal dotted grey lines indicating ‘treat none’ and vertical solid grey lines indicating ‘treat all’. Shaded areas represent the incremental benefit of integrating RNFLT metabolic states into the Age&Sex model and FGCRS, respectively. Source data are provided as a Source Data file. RNFLT retinal nerve fibre layer thickness, RNFLT MET RNFLT metabolic state, FGCRS Framingham General Cardiovascular Risk Score, T2D type 2 diabetes.

Fig. 5. Comparison of absolute performance and benefit in predicting CMD outcomes across different demographic groups.

a–f Comparison across sexes (n = 9380 for females and n = 7823 for males). g–l Comparison across socioeconomic statuses (n = 8587 for high deprivation and n = 8591 for low deprivation). m–r Comparison across educational attainment (n = 6939 for university and n = 10,038 for non-university). Data are presented with bars representing the performance and benefits of various demographic groups, with 95% CIs indicated by error bars. Each bar represents a model-level C-index or ΔC-index calculated from all individuals within the corresponding subgroup. Source data are provided as a Source Data file. RNFLT MET RNFLT metabolic state, FGCRS Framingham General Cardiovascular Risk Score, T2D type 2 diabetes, TDI Townsend deprivation index.

Fig. 6. Extrapolation in the GDES.

a Landscape of the RNFLT metabolic state profile for cardiovascular disease captured by complementary LC–MS assays (n = 1286). Individual metabolite attributions are aggregated by percentiles, with each dot representing one percentile. The distance of a dot from the circular baseline reflects the strength of the absolute attribution for that percentile. Deviations towards the centre and periphery signify negative and positive contributions. Dot colours indicate the normalized values for each metabolite. b–e Comparison of predictability (b and d) and clinical utility (c and e) between established models and models incorporating RNFLT metabolic states for predicting cardiovascular disease across varying genetic susceptibility (n = 1286) (d and e). Data are presented as estimated performance for different models and genetic susceptibility contexts with 95% CIs indicated by error bars. Shaded areas illustrate the incremental net benefit of incorporating RNFLT metabolic states into established models. f–h Comparison of performance for predicting cardiovascular diseases across different demographic groups (n = 1286): sex (f), income (g), and educational attainment (h). Colours denote the absolute performance and benefits of various demographic groups. Source data are provided as a Source Data file. GDES Guangzhou Diabetic Eye Study, RNFLT MET RNFLT metabolic state, FGCRS Framingham General Cardiovascular Risk Score, UKPDS UK Perspective Diabetes Study, NZ-DCS New Zealand Diabetes Cohort Study, WAN Wan’s model.