AI-based individual risk prediction of oral anticoagulation in patients with cirrhosis and atrial fibrillation
I S Kim, H W Lee, J S Lee, M N Kim, S U Kim, J Y Park, D Y Kim, S H Ahn, B K KimAbstract
Background
Oral anticoagulation (OAC) should be carefully administered especially among atrial fibrillation (AF) patients with liver cirrhosis (LC). We developed OAC-dose selection calculator considering risk-to-benefit ratio for such population.
Methods
A total of 600 AF patients with LC (420 derivation/180 validation sets) were recruited. Machine learning (ML)-based risk prediction model for major bleeding (MB) and ischemic stroke (including transient ischemic attack) were constructed, using components of CHA2DS2-VASc and Child-Pugh scores.
Results
In the derivation set, our ML-based model showed significantly higher c-statistics to predict MB (0.768) and ischemic stroke (0.743), compared to each existing model, i.e. ORBIT-AF for MB (0.642) and CHA2DS2-VASc for ischemic stroke (0.663) (both p<0.05). Both net reclassification improvement and integrated discrimination improvement indices were significantly positive (p<0.05). Proportions with 3-year absolute risk increases (ARI)>30% for MB, compared to non-OAC, increased across low-dose direct oral anticoagulants (DOACs) (12%), standard-dose DOACs (18%), and warfarin (21%). Proportions with 3-year absolute risk reductions (ARR)>15% for ischemic stroke, compared to non-OAC, also increased across low-dose DOACs (17%), standard-dose DOACs (22%), and warfarin (32%). The optimal cutoffs of ARI for MB and ARR for ischemic stroke when using OAC vs. non-OAC were >2.4%/year and >1.2%/year, respectively, with c-statistics (95% confidence intervals) of 0.81 (0.78-0.83) and 0.75 (0.72-0.79).
Conclusions
By ML-based model, individual risks for MB and ischemic stroke can be effectively predicted through readily available clinical variables among AF patients with LC, also guiding appropriate OAC dosing.