Model Predicts Risk of Stroke

A group of researchers at Harvard and Massachusetts General Hospital have published findings on a new, more accurate method to predict stroke risk through a statistical model that screens for thousands of potentially stroke-causing genes.

The model—called a Bayesian network—differs from existing methods that examine only the effect of a single factor, instead allowing researchers to study the interaction of multiple genes.

The study, released last month, found the formula could predict a person’s risk for the most common stroke in the U.S. with 86 percent accuracy—which the researchers said was a significant improvement over previous models, which only predicted risk with 50 percent accuracy.

Originally used for artificial intelligence, the Bayesian network was used to screen 1,313 genes in 569 individuals and turned up 37 that worked collectively to forecast a cardioembolic stroke.

Lead author and Medical School Associate Professor Marco F. Ramoni said he first explored predicting stroke risk after observing a high rate of strokes in a 2005 sickle-cell anemia study he conducted.

That began a collaboration with Karen L. Furie, Mass. General’s director of stroke service, who was instrumental in providing the raw data needed to construct the Bayesian model.

“We’re fortunate Karen had been collecting this data,” said Rachel L. Ramoni, an instructor at the Harvard School of Dental Medicine who worked on the study. “She had an ongoing study to collect blood and other information from people affected with a stroke and those not affected by a stroke.”

Both the Ramonis (the two researchers are married) said using the Bayesian model was particularly important for predicting strokes because the disease is so complex and often results from interactions of many different factors.

Marco Ramoni said his next step will be to analyze even more genes, hopefully one million by next year.

“The more you see the more accurate the predictions will be,” Ramoni said, adding that he hoped to get enough data to build models for different ethnic groups.

Deputy Director of the National Institute of Neurological Disorder and Stroke Walter J. Koroshetz said the findings were interesting, but more studies would need to be conducted to test whether they were applicable to the general population, which would indicate whether the interactions identified are universal or only pertain to the 569 individuals in the study.

“If this pattern is predictive and you see the same thing then that’s very robust,” Koroshetz said. “The model works no matter where you go.”