Project Summary
Covert cerebrovascular disease (CCD), comprising covert brain infarction (CBI) and white matter disease
(WMD), is commonly found incidentally on neuroimaging scans obtained in routine clinical care. CBI is
believed to be pathophysiologically similar to clinically-evident stroke, but with infarct locations in clinically less
articulate parts of the brain. Epidemiologic studies based on MRI-screened cohorts indicate that CBI is far
more common than clinically-evident stroke, and WMD is even more common. Both types of CCD are strong,
independent risk factors for future stroke and dementia. The American Heart Association/American Stroke
Association has identified CCD as a major priority for new studies on stroke prevention. Identifying high-risk
individuals prior to the onset of severe cognitive decline is a central goal in dementia prevention research.
However, application of insights from studies of MRI-screened cohorts to patients with incidentally-discovered
CCD is not straightforward. Patients are not screened but are selected for clinically-indicated neuroimaging;
real-world imaging is dominated by CT scans; neuroimaging interpretation and reporting are heterogeneous
and poorly standardized. Additionally, there are no ICD-9 codes for CBI or WMD, which also do not generally
appear in a patient’s problem list or in structured fields of electronic health records. Indeed, both patients and
their providers are often unaware of these findings even after they are detected. Our team has performed the
first large-scale study of the prevalence and prognostic importance of incidentally-discovered (id)-CCD using
natural language processing (NLP) algorithms and created a real-world cohort comprising approximately
250,000 stroke- and dementia-free patients over age 50 who received either head CT or MRI, approximately
30% of whom have id-CCD. These NLP-identified id-CBI and id-WMD are both strong predictors of future
clinically-evident stroke and dementia. While this cohort has yielded important insights, there are substantial
limitations. Extractable information on WMD severity is frequently missing. Even when present in text reports,
only relatively crude information on WMD burden can be obtained. Further, other evidence of CCD—including
microbleeds, microinfarcts, prominent perivascular spaces—may also be prognostically important. Finally,
direct image analysis may discover new radiomic biomarkers predictive of future stroke and dementia that are
not captured or unknown. The ambitious goal of this project is to develop a clinically useful algorithm that can
directly read routinely obtained neuroimaging scans at scale and relate findings to stroke and dementia
outcomes through the following specific aims: Aim 1: Develop a deep learning model to identify and
characterize covert id-CCD from routinely obtained CT and MRI neuroimages and use these features, and
other deep radiomic biomarkers, to predict the development of future stroke and dementia. Aim 2: Test the
deep learning algorithms developed in Aim 1 on a large cohort of patients from an integrated health system for
the prediction of future stroke and dementia to evaluate their incremental prognostic value.