Systematic Review and Meta-analysis of the Predictive Performance of Stroke and Bleeding Prediction Models in Atrial Fibrillation Patients With Kidney Disease.

RATIONALE &

OBJECTIVE

Patients with atrial fibrillation (AF) and chronic kidney disease (CKD) are at high risk for ischemic stroke (IS) and bleeding. The applicability of prediction models in this population remains debated. This study aimed to (1) identify external validations of CHA2DS2-VASc, CHADS2, HAS-BLED, and HEMORR2HAGES model scores in patients with AF undergoing dialysis or with CKD, (2) provide pooled estimates, and (3) assess their risk of bias (ROB).

STUDY

DESIGN

Systematic review and meta-analysis.

SETTING &

PARTICIPANTS

We searched Web of Science, PubMed, MEDLINE, Embase, Emcare, PMC, Cochrane Library, and Academic Search Premier for studies externally validating IS and bleeding prediction models in patients with AF undergoing dialysis or with CKD.

EXPOSURES: AF and CKD or dialysis.

OUTCOMES: IS and bleeding.

ANALYTICAL APPROACH: Eligible studies were reviewed, discrimination was pooled using random-effects meta-analysis, calibration was calculated and plotted, and the ROB score was assessed using the prediction model ROB assessment tool.

RESULTS

The CHA2DS2-VASc score was validated in 35 studies, the CHADS2 and HAS-BLED scores in 19 each, and the HEMORR2HAGES score in 1. Among 627,199 patients, 28,493 (4.5%) experienced IS and 25,695 (4.1%) bleeding. Only 12 studies presented c-statistic scores. In patients with AF and CKD, the CHADS2 model score showed nominally better discrimination predicting IS (pooled c-statistic score of 0.70) than the CHA2DS2-VASc model score (0.64). In patients with AF undergoing dialysis, the CHA2DS2-VASc and CHADS2 model scores showed similar discrimination predicting IS (both 0.70), and the HAS-BLED and HEMORR2HAGES model scores showed similar c-statistic scores predicting bleeding (0.55 and 0.56, respectively). Calibration was good in the most relevant high-risk group.

LIMITATIONS: All studies were at high ROB scores, contained within- and between-study heterogeneity, and often merged scoring categories or populations, limiting comparability.

CONCLUSIONS

Although modest, the discrimination of prediction models in patients with AF undergoing dialysis or with CKD is similar to patients with AF without CKD. Despite the described limitations, these models can be used in clinical practice for patients with CKD and patients undergoing dialysis.