Identifying patients with atrial fibrillation and truly low
Identifying patients with atrial fibrillation and "truly low" thromboembolic risk who are poorly characterized by CHA 2 DS 2 VASc: Superior performance of a novel machine learning tool in GARFIELD AF Keith A. A. Fox, Joseph E. Lucas, Karen S. Pieper, Jean Pierre Bassand, A. John Camm, David A. Fitzmaurice, Werner Hacke, Gloria Kayani, Ali Oto, Ajay K. Kakkar for the GARFIELD AF Investigators The GARFIELD Registry is funded by an unrestricted research grant from Bayer Pharma AG www. tri london. ac. uk
Background and Context • • The role of anticoagulation for patients with AF and ≥ 1 risk factor for stroke/systemic embolism is defined by trial evidence and guidelines Between 2010 2011 and 2014 2015, anticoagulation usage rose from 57% to 71% of patients with AF However, the balance of risk and benefit is poorly defined for “low risk’ AF 100 Proportion of patients, % • 90 80 70 60 50 40 30 20 10 0 Cohort 1 2010– 11 (n=5311) VKA+AP Cohort 2 2011– 13 (n=11, 562) Cohort 3 2013– 14 (n=11, 343) FXa. I DTI FXa. I+AP DTI+AP Camm AJ et al. Heart 2016 (in press) www. tri london. ac. uk Cohort 4 2014– 15 (n=10, 923) AP None
How are low and high risk AF patients managed in practice? Contrary to international guideline recommendations, – 28% high risk patients (CHA 2 DS 2 VASc ≥ 2) are not anticoagulated – 51% of very low risk patients (CHA 2 DS 2 VASc 0) are anticoagulated CHA 2 DS 2 -VASc 0 1 ≥ 2 100 90 Proportion of patients, % • 80 None 70 AP 60 Factors beyond those in current risk scores appear to DTI+AP influence DTI 50 prescribing decisions on anticoagulation, including risk of bleed FXa. I+AP 40 FXa. I 30 VKA+AP 20 VKA 10 0 (n=352) (n=1336) Camm AJ et al. Heart 2016 (in press) www. tri london. ac. uk (n=9027)
Purpose: To provide accurate estimates of risk as the basis of decisions on prescribing or withholding anticoagulation Aim: To derive and validate a more accurate and user friendly method of stratifying patients according to risks of death, stroke and bleeding www. tri london. ac. uk
Statistical Methods: The GARFIELD Score A “machine learning” approach to risk modelling • Coalescent regression avoids the need to specify levels of relatedness in the statistical model, it allows joint modeling of all outcomes. • Models were based on 38984 patients in GARFIELD 2010 to 2015 for: – all-cause mortality, – ischaemic stroke/thromboembolism, and – haemorrhagic stroke/major bleed that occurred within 1 year of enrolment into GARFIELD AF. • Also, a simplified model was also derived to facilitate web applications • The performance of both models were compared with CHA 2 DS 2 VASc in all patients and those with a low risk of stroke • External validation was undertaken using an independent contemporary registry ORBIT AF www. tri london. ac. uk
Number of events in low and higher risk patients at 1 year Number of events determined using one year Kaplan-Meier rates Event Low risk* (n=7 861; 20. 2%) Higher risk All cause mortality 94 (1. 4%) 1387 (4. 9%) Ischaemic stroke/ Systemic embolism 35 (0. 5%) 396 (1. 4%) Haemorrhagic stroke/ Major bleed 26 (0. 4%) 295 (1. 1%) • • (n=31 123) Low risk patients (defined as CHA 2 DS 2 VASc 0 or 1 for men and 1 or 2 for women) represent 20. 2% of overall cohort Total number of patients: 38, 984 enrolled between March 2010 and July 2015 www. tri london. ac. uk
GARFIELD Score performance characteristics in all patients All-cause mortality C statistic: 0. 78 Ischaemic stroke / Systemic embolism C statistic: 0. 63 www. tri london. ac. uk Haemorrhagic stroke / Major bleed C statistic: 0. 67
Comparison of GARFIELD Score with CHA 2 DS 2 VASc in all patients Performance Event GARFIELD Score CHA 2 DS 2 -VASc measure C statistic All cause mortality 0. 78 0. 66 Ischaemic stroke / systemic embolism 0. 63 Haemorrhagic stroke / major bleed 0. 67 0. 61 www. tri london. ac. uk
Comparison of GARFIELD Score with CHA 2 DS 2 VASc in low risk patients CHA 2 DS 2 VASc 0 or 1 for men and 1 or 2 for women Performance Events GARFIELD measure CHA 2 DS 2 -VASc Score C statistic All cause mortality 0. 72 0. 56 Ischaemic stroke / Systemic embolism 0. 62 0. 56 Haemorrhagic stroke / Major bleed 0. 72 0. 57 www. tri london. ac. uk
Performance of the new simplified GARFIELD Score in patients enrolled in GARFIELD AF and ORBIT I Population Endpoint C Statistic Events (n / N) (95% CI) GARFIELD AF Ischaemic stroke/SE 0. 70 (0. 68, 0. 73) 438 / 38, 607 ORBIT I Any stroke/SE 0. 69 (0. 64, 0. 75) 91 GARFIELD AF Haemorrhagic stroke/major bleed 0. 68* (0. 64, 0. 72) 187 / 12, 249 1 ORBIT I Major bleed 0. 61 (0. 58 0. 64) 625 / 7, 442 / 9, 743 *C statistic for HAS-BLED is 0. 64 (95% CI 0. 59, 0. 68) 1. Evaluation of a subset of patients who were prescribed oral anticoagulants in countries where at least 1% bleeding rate was recorded www. tri london. ac. uk
Conclusions • Performance of GARFIELD Score was superior to CHA 2 DS 2 VASc in predicting ischaemic stroke or major bleed in all patients, and those with a low risk of stroke • This integrated risk tool has the potential for incorporation in routine electronic systems www. tri london. ac. uk
Next steps • A simplified GARFIELD Score, validated using data from ORBIT AF, is being developed, with web based and mobile device applications* • The GARFIELD Score may help physicians assess the appropriateness of anticoagulation in low risk patients *http: //colab sbx 322. oit. duke. edu: 3338/ www. tri london. ac. uk
BLEEDING SCORE 20% Risk of major bleed in 1 year
- Slides: 13