Researchers at Zhongshan Hospital in Xiamen studied 82 patients who had strokes related to non-valvular atrial fibrillation and compared them to 164 patients without strokes. The group included individuals with low-to-moderate risk scores, specifically those with CHA2DS2-VA scores of 1 or less. The study looked at demographics, medical history, lab results, and heart ultrasound images to see how well different computer models could predict who was at risk.
The team found that people who had strokes were older and more likely to have conditions like heart failure or high blood pressure. Lab tests showed higher levels of certain proteins and white blood cells in the stroke group. When the researchers tested their machine-learning models, one specific model using XGBoost algorithms performed very well, achieving high accuracy in predicting stroke risk.
Analysis of the data revealed that a specific heart marker, NT-proBNP, was the most important factor in the prediction. Higher levels of this marker and another heart measurement increased the predicted risk, while a higher left ventricular ejection fraction lowered it. This suggests that adding these specific heart details to risk calculations could help doctors make more personalized decisions about preventing strokes in this patient group.