OC 24.5 - Modest Performance of Available Risk Prediction Models when Used to Predict Hemorrhage in Patients with Chronic Liver Disease

Background: Patients with chronic liver disease (CLD) have a unique hemostatic profile with simultaneous reduction in factors promoting and opposing thrombosis, and multiple abnormalities on conventional coagulation tests. Available risk prediction models for anticoagulant-related major bleeding (MB) were not developed and validated in the CLD population, and hence, may not predict MB among these patients to assess the risk-versus-benefit of anticoagulation.

Aims: We aimed to evaluate the performance of existing (anticoagulant-related) bleeding risk prediction models in a cohort of CLD patients.

Methods: Using the Veterans Health Administrative data, we identified patients with CLD (using previously validated methods) who were started on anticoagulant therapy between 2001 and 2018. We identified hemorrhage within 12 months of anticoagulant therapy through previously validated ICD-9/10 codes present in primary or secondary position of inpatient diagnoses codes. We evaluated the predictive performance of three risk prediction models in the CLD cohort, and assigned points as recommended (Table 1). The association between the score assigned by each model and MB was measured using competing-risk analysis by Fine and Grey. We evaluated each model’s discrimination using Harrell’s c-statistic.

Results: Among 19,871 CLD patients, 761 experienced hemorrhage (variceal or non-variceal) within twelve months of initiation of anticoagulation. The median time to hemorrhage from initiation of anticoagulation was 60 days. The increase in risk of MB per point increase in risk prediction score is presented in Table 2. The c-statistic for HEMORR2HAGES and HAS-BLED was 0.56, and it was 0.51 for VTE-BLEED.

Conclusion(s): In this cohort of 19,871 veterans with CLD, available risk prediction models predicted anticoagulant-related MB only slightly better than random chance. There is hence, a need to develop and validate a risk prediction model in patients with CLD to accurately identify those at a higher risk of anticoagulant-related bleeding.