Abstract 12941: Using Deep Learning to Predict the Framingham Risk Score From Brain Magnetic Resonance Imaging

• Published in: Circulation (New York, N.Y.) N.Y.), 2019-11, Vol.140 (Suppl_1 Suppl 1), p.A12941-A12941
• Main Authors: Bermudez Noguera, Camilo, Held, Lori, Resnick, Susan, Wells, Quinn, landman, bennett
• Format: Article
• Language: English
• ISSN: 0009-7322; 1524-4539
• DOI: 10.1161/circ.140.suppl_1.12941

IntroductionCardiovascular disease (CVD) risk factors are increasingly recognized to adversely impact neurological function. However, it is unclear how CVD risk factors affect brain morphology. Identifying brain structural changes associated with cardiovascular health may provide insights into cognitive changes associated with heart disease. Advances in quantitative neuroimaging and machine learning methods, such as deep learning, have made possible the discovery of imaging biomarkers associated with disease at scale.HypothesisWe hypothesized that application of deep learning algorithms to a large library of clinical neuroimaging would discover biomarkers of CVD directly from imaging data with limited human biases.MethodsWe calculated the Framingham Heart Study Risk Score (FRS) in 305 subjects from the Baltimore Longitudinal Study of Aging. The mean FRS was 5.20 +/- 0.31 (range 0.5 -30). Subjects also underwent T1-weighted brain magnetic resonance imaging (MRI) at the time of FRS assessment. A fully convolutional neural network was trained on the raw brain MRI to predict the FRS score.ResultsThe neural network predicted the FRS with a mean absolute error of prediction of 2.71 ± 1.65 points. The correlation coefficient between the true FRS score and the predicted FRS score was 0.713 (Figure 1). A linear fit between the true FRS and predicted FRS returned an R value of 0.629.ConclusionsThese data suggest there are biomarkers of CVD in structural T1-weighted brain MRI that are detectable using machine learning. Subjects with large errors in prediction may represent different degrees of neurological involvement seen in imaging that are not captured by the FRS. These quantitative imaging biomarkers may improve detection of CVD and elucidate mechanisms linking CVD to neurologic disfunction.