Machine learning reveals distinct neuroanatomical signatures of cardiovascular and metabolic diseases in cognitively unimpaired individuals

Abstract

Comorbid cardiovascular and metabolic risk factors (CVM) differentially impact brain structure and increase dementia risk, but their specific magnetic resonance imaging signatures (MRI) remain poorly characterized. To address this, we developed and validated machine learning models to quantify the distinct spatial patterns of atrophy and white matter hyperintensities related to hypertension, hyperlipidemia, smoking, obesity, and type-2 diabetes mellitus at the patient level. Using harmonized MRI data from 37,096 participants (45-85 years) in a large multinational dataset of 10 cohort studies, we generated five in silico severity markers that: i) outperformed conventional structural MRI markers with a ten-fold increase in effect sizes, ii) captured subtle patterns at sub-clinical CVM stages, iii) were most sensitive in mid-life (45-64 years), iv) were associated with brain beta-amyloid status, and v) showed stronger associations with cognitive performance than diagnostic CVM status. Integrating personalized measurements of CVM-specific brain signatures into phenotypic frameworks could guide early risk detection and stratification in clinical studies.

Fig. 1. Development of SPARE-CVM models.

Cardiovascular and metabolic risk factors (CVMs) contribute to distinct alterations in brain structure, potentially leading to diverse patterns of brain change. The current study leverages supervised machine learning (ML) to derive the phenotypic expression of CVMs from structural brain magnetic resonance images (sMRI). T1-weighted and T2-weighted FLAIR MRI from the multi-study iSTAGING dataset underwent rigorous preprocessing, segmentation, quality control (QC), and harmonization across study sites to obtain region of interest (ROI) brain volumes and white matter hyperintensity (WMH) volumes, respectively. These imaging features serve as inputs to the ML models. To establish ground truth labels for the ML models, clinical measurements, laboratory tests, and self-reported medical history were standardized to determine the presence or absence of five key CVMs: hypertension, hyperlipidemia, smoking, obesity, and type 2 diabetes mellitus. The ML models identify distinct Spatial Patterns of Abnormalities on sMRI associated with each CVM, summarized as individualized phenotypic expression scores (SPARE-CVMs), reflecting the influence of individual CVMs on brain structure.

Fig. 2. CVM co-occurrence and multi-morbidity influence SPARE-CVM profiles across phenotypic dimensions.

The co-occurrence of cardiovascular and metabolic risk factors (CVMs) is prevalent in the general population, contributing to heterogeneity among participants and impacting the spatial patterns of abnormality related to CVMs (SPARE-CVMs). Center: A Venn diagram illustrates the CVM co-occurrence patterns observed in the training dataset. To visualize the influence of single- and multi-morbidity, three-dimensional projections of SPARE-CVMs are presented (A–D). Each ellipsoid within these projections represents the SPARE-CVM scores closest to the mean for a specific combination of CVM statuses. Participants with CVM− status for all three CVMs (All−) exhibit the lowest SPARE-CVMs, while those with CVM+ status for all three CVMs (All+) demonstrate the highest SPARE-CVMs. Notably, participants with CVM+ status in only one of the three CVMs show elevated scores specifically in the corresponding SPARE-CVM dimension, reflecting the distinct contributions of individual CVMs. Source data are provided as a Source Data file.

Fig. 3. SPARE-CVMs capture distinct spatial sMRI patterns of CVMs.

Volumes of gray matter (GM), white matter (WM), and white matter hyperintensities (WMH) showed significant associations with SPARE-CVMs (p < 0.001, Bonferroni corrected for multiple comparisons). Regional associations between GM volumes and SPARE-CVMs are visualized using 3D surface maps, displaying regression coefficients derived from two-sided multiple linear regression analyses. Lobar associations between WM and WMH volumes and SPARE-CVMs are visualized using glass brain plots, also displaying regression coefficients from two-sided multiple linear regression analyses. Hot colors (red) indicate positive associations (higher volumes associated with higher SPARE-CVMs), while cold colors (blue) indicate negative associations (lower volumes associated with higher SPARE-CVMs). SPARE Spatial Patterns of Abnormality Related to, GM Gray Matter, WMH White Matter Hyperintensities, L Left hemish.

Fig. 4. SPARE-CVMs detect brain patterns more effectively across the clinical stages.

SPARE-CVMs exhibit enhanced sensitivity to CVM-related brain changes compared to conventional volumetric measures and machine-learning-based imaging markers for Alzheimer’s disease and brain age. A A heatmap displays Cohen’s d effect sizes for each imaging marker (columns) at differentiating CVM+ and CVM− participants for the corresponding CVM (rows). The highest effect sizes for each CVM were observed in one-to-one correspondences between the SPARE model and the target CVM (outlined in blue). B SPARE-CVM indices show clear separability across clinical stages when compared to imaging markers which had small effect sizes in (A). Sub-clinical (undiagnosed) stages, defined by continuous clinical measures, were excluded from the training dataset (see “Methods”). C SPARE-CVM scores were significantly higher at sub-clinical stages, as demonstrated by: (i–iii) Linear regression for categorical variables (with participants without CVM [CVM−/Normal] as the reference group) for SPARE-Hypertension, SPARE-Hyperlipidemia, and SPARE-Diabetes, visualized as boxplots (median and interquartile range). (iv) Linear regression for SPARE-Smoking (with participants who never smoked [Never/0 years] as the reference group), visualized as a line plot (mean ± 95% confidence interval). (v) Pearson’s correlation for the correlation between SPARE-Obesity and body mass index (BMI), visualized in a scatterplot. All linear regressions were two-sided. Sample sizes, uncorrected p values, and confidence intervals for the statistical tests are provided in the Source Data file. SPARE Spatial Patterns of Abnormality Related to, AD Alzheimer’s Disease, CVM Cardiovascular and Metabolic Risk Factors, GM Gray Matter, WMH White Matter Hyperintensities.

Fig. 5. SPARE-CVMs detect strong CVM effects in mid-life.

Peak effect sizes of SPARE-CVMs were observed in individuals aged 45–65 years, with a subsequent decline in effect sizes at older ages. Source data are provided as a Source Data file.

Fig. 6. SPARE-CVMs exhibit stronger associations with cognitive performance than CVM labels.

The results of the two-sided multivariate regression models predicting cognitive performance in UKBiobank using SPARE-CVMs (left) or CVM status (right), while adjusting for the confounding covariates of study, age, sex, and number of years of education, are shown below. Regression coefficients are represented by circles, with error bars indicating 95% confidence intervals. SPARE-CVM scores were adjusted for potential confounders, namely age, sex, and DLICV, prior to the regression analysis. Participants without CVM (CVM−) were the reference group in the regression models for diagnostic status. Odds ratio for smokers (SM+) for P-Mem is not shown on the CVM-status graph because it fell outside the axis limits (OR = 5.5, CI limits = −0.5,11.9, p > 0.05). For the Sample sizes, p values, and confidence intervals are provided in the Source Data file. DSST Digit symbol substitution test, TMT−A/B Trail-making test −A/B; P-Mem Prospective memory. False discovery rate corrected p values are indicated by: *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < <0.001. Associations with p values > 0.05 are not marked. For similar analyses on other cohorts in iSTAGING, please refer to Supplementary Fig. 12.