Using radiological brain health to predict recurrence after ischemic stroke and transient ischemic attack: A population-based study.
Authors
Affiliations (7)
Affiliations (7)
- Department of Neurology, University of Cincinnati School of Medicine, Cincinnati, OH.
- Department of Neurology, Yale University, New Haven, CT.
- Department of Emergency Medicine, University of Cincinnati School of Medicine, Cincinnati, OH.
- Department of Radiology, University of Cincinnati School of Medicine, Cincinnati, OH.
- I-MED Radiology Network, Melbourne, VIC, Australia.
- Cincinnati Childrens Hospital Medical Center, Cincinnati, OH.
- Department of Neurology, University of Michigan, Ann Arbor, MI.
Abstract
Approximately one-fourth of all strokes are recurrent, which is associated with a higher risk of morbidity/death. Identifying stroke survivors at greatest risk of recurrence could help direct secondary prevention efforts, but prior models based on clinical data have limited predictive ability. Neuroimaging markers of brain health may be uniquely able to capture the duration and severity of traditional and nontraditional risk factors for recurrence, and we examined whether they could improve the prediction of stroke recurrence beyond clinical factors alone using advanced machine-learning techniques. Using a well-validated approach, all ischemic strokes and TIAs during 2015 were ascertained in a representative population of the United States. Standard-of-care neuroimaging was collected for all patients across 18 different hospitals and centrally read/analyzed for more than 20 different measures of brain health. Only patients who underwent standard of care MRI were included. Patients were then followed for 3 years to assess for stroke recurrence. Using both clinical and neuroimaging variables, prediction models of recurrence at 90-days and 3-years were built using random survival forests. A total of 1999 patients (46.6% male and 21.2% Black) had sufficient clinical and imaging data to be included in the final study population, including 275 (13.8%) patients who experienced a recurrent stroke event. Models that included clinical and imaging variables outperformed those with clinical data alone at 90 days (C-statistics increased from 0.63 with 95% CI 0.520-0.740 , to 0.667 with 95% CI 0.567-0.767; <i>P</i>=0.0498) and at 3 years (C-statistics increased from 0.618 with 95% CI 0.572-0.663 to 0.683 with 95% CI 0.637-0.729; <i>P<</i>0.001).Among imaging variables, global cortical atrophy scale, microbleed count, deep white matter hyperintensities (WMH), and baseline CT ASPECTS were highly important variables at both time points. Neuroimaging data can improve the prediction of stroke recurrence but only moderately improves discriminative ability. Other strategies will be needed to better identify stroke patients at highest risk of recurrence, including dynamic prediction models or other non-imaging biomarkers.