Advancing stroke prevention in atrial fibrillation: a systematic review of machine learning-based risk prediction models.
Md Mohaimenul Islam, Arinze Nkemdirim Okere
✦ AI-curated · Sources linked
In 30 seconds
This systematic review evaluated machine learning (ML) models for predicting ischemic stroke in patients with atrial fibrillation (AF) using electronic health record data. Eight studies involving 809,523 patients were analyzed. ML models showed improved predictive performance over the CHA2DS2-VASc score, with AUROCs ranging from 0.66 to 0.91 compared to 0.54 to 0.68 for the traditional score, though methodological limitations were significant.
Key findings
- ML models achieved AUROCs from 0.66 to 0.91, outperforming CHA2DS2-VASc's 0.54 to 0.68.
- Only one study performed external validation, highlighting a critical methodological gap.
- 88% of studies had a high risk of bias in the analysis domain under PROBAST.
- Several ML models showed only marginal gains, with AUROCs between 0.63 and 0.69.
Why it matters
Understanding the predictive capabilities of ML models for stroke risk in AF patients is crucial, as current tools may underestimate or overestimate risk, impacting treatment decisions and patient safety.
What to keep in mind
The review identified high analytic risk of bias, lack of external validation, and insufficient model interpretability as significant limitations.