Integrating multiple brain imaging modalities does not boost prediction of subclinical atherosclerosis in midlife adults

Abstract

Background: Human neuroimaging evidence suggests that cardiovascular disease (CVD) risk may relate to functional and structural features of the brain. The present study tested whether combining functional and structural (multimodal) brain measures, derived from magnetic resonance imaging (MRI), would yield a multivariate brain biomarker that reliably predicts a subclinical marker of CVD risk, carotid-artery intima-media thickness (CA-IMT).

Methods: Neuroimaging, cardiovascular, and demographic data were assessed in 324 midlife and otherwise healthy adults who were free of (a) clinical CVD and (b) use of medications for chronic illnesses (aged 30-51 years, 49% female). We implemented a prediction stacking algorithm that combined multimodal brain imaging measures and Framingham Risk Scores (FRS) to predict CA-IMT. We included imaging measures that could be easily obtained in clinical settings: resting state functional connectivity and structural morphology measures from T1-weighted images.

Results: Our models reliably predicted CA-IMT using FRS, as well as for several individual MRI measures; however, none of the individual MRI measures outperformed FRS. Moreover, stacking functional and structural brain measures with FRS did not boost prediction accuracy above that of FRS alone.

Conclusions: Combining multimodal functional and structural brain measures through a stacking algorithm does not appear to yield a reliable brain biomarker of subclinical CVD, as reflected by CA-IMT.

Keywords: Brain imaging; Cardiovascular disease; Intima-media thickness; Prediction stacking.

Fig. 1

A) Left panel shows CA-IMT acquisition using ultrasound. Middle and right panels show example ultrasound images with the CA-IMT indicated. B) Raincloud plot showing distribution of CA-IMT (mm) in our sample. C) Raincloud plot showing distribution of FRS in our sample. D) Scatterplot showing the linear regression of FRS on CA-IMT. Line of best fit shown in blue. CA-IMT = carotid artery intima-media thickness, FRS = Framingham Risk Score.

Fig. 2

Prediction stacking model schematic, with linear SVR and linear regression used in the unimodal predictions and random forest used in the multimodal prediction. FC = functional connectivity, SVR = support vector regression, LR = linear regression, RF = random forest.

Fig. 3

Scatterplot showing correlation between participants’ chronological age and predicted brain age according to multimodal model. Blue line represents the line of best fit.

Fig. 4

For all panels, blue bars show single channel predictions of CA-IMT. Yellow bars show channel combination predictions that include only brain measures. Green bars show channel combinations predictions that include FRS. Error bars indicated 95% confidence intervals (calculated using 1000 bootstrap iterations). Channel combinations are indicated numerically with 1 = resting-state FC, 2 = cortical SA, 3 = cortical thickness, 4 = subcortical volume, 5 = FRS. Median values for the Monte Carlo simulation for single channel and every possible channel combination prediction of mean CA-IMT: A) Pearson correlation coefficient, r, B) RMSE (horizontal dotted line represents the standard deviation of CA-IMT in our sample, 0.084 mm), C) coefficient of determination, and D) Bayesian information criterion. CA-IMT = carotid-artery intima-media thickness, FRS = Framingham Risk Score, FC = functional connectivity, SA = surface area.