Key result
XGBoost outperforms CHA₂DS₂-VASc by ~0.16 AUROC points for predicting stroke in AF.
Why the study?
Risk stratification for ischaemic stroke and systemic embolism in AF typically relies on clinical scores with predefined predictors, prompting comparison of conventional CHA2DS2-VASc scoring against logistic regression and machine learning models using routine clinical data.
Does machine learning-based risk stratification using routine clinical data improve the prediction of ischaemic stroke or systemic embolism within 24 months compared to conventional CHA₂DS₂-VASc scoring in adults with atrial fibrillation?
Cohort (n=248,500)
Yes
Does machine learning-based risk stratification using routine clinical data improve the prediction of ischaemic stroke or systemic embolism within 24 months compared to conventional CHA₂DS₂-VASc scoring in adults with atrial fibrillation?
Machine learning models, particularly XGBoost incorporating routine multidimensional clinical data, significantly outperformed the conventional CHA₂DS₂-VASc score in predicting 24-month thromboembolic events in patients with atrial fibrillation.
Answers are grounded in this paper and cite their sources.
Machine learning may improve AF stroke risk prediction; leaves open need for prospective validation before clinical adoption.
Background: Risk stratification for ischaemic stroke and systemic embolism in atrial fibrillation (AF) typically relies on clinical scores with predefined predictors. This study compared conventional CHA₂DS₂-VASc scoring with logistic regression and machine learning models using routine clinical data.Methods: A retrospective cohort of 248,500 adults with AF was identified in the TriNetX Global Collaborative Network from 2015 to 2025. Development and validation sets comprised 198,800 and 49,700 individuals, respectively. The primary outcome was ischaemic stroke or systemic embolism within 24 months. Discrimination, calibration, net benefit, and subgroup performance were evaluated.Results: A total of 12,425 events were recorded (5.0% of the baseline cohort); 221,680 participants had ascertainable 24-month outcome status (event proportion, 5.6%). In the validation analysis, XGBoost achieved the highest AUROC (0.842; 95% CI, 0.832–0.852), followed by random forest (0.826), support vector machine (0.811), and logistic regression (0.794), compared with 0.681 for CHA₂DS₂-VASc. The XGBoost–CHA₂DS₂-VASc difference in AUROC was 0.161 (95% CI, 0.145–0.177; p < .001). XGBoost had a Brier score of 0.041 and higher estimated net benefit across the evaluated 2%–15% risk thresholds.Conclusion: In this internal validation study, models incorporating multidimensional clinical data, particularly XGBoost, discriminated thromboembolic events more effectively than the conventional CHA₂DS₂-VASc score. The comparison highlights the potential contribution of clinical and laboratory variables to AF risk prediction; external validation remains necessary before clinical implementation.
No takes yet. Share an insight, caveat, or question.
Nebonta et al. (2026) conducted a cohort in Atrial fibrillation (n=248,500). Machine learning models (XGBoost, random forest, support vector machine) and logistic regression vs. Conventional CHA₂DS₂-VASc score was evaluated on Ischaemic stroke or systemic embolism within 24 months (AUROC difference 0.161, 95% CI 0.145-0.177, p=<0.001). XGBoost outperformed CHA₂DS₂-VASc in predicting 24-month stroke or systemic embolism in AF patients.
Synapse has enriched 10 related papers. Consider them for comparative context: