Background:
BAVM is an important cause of intracranial hemorrhage (ICH) in younger persons. Accurate and reliable prediction models for determining ICH risk in the natural history course of BAVM patients are needed to help guide management. The purpose of this study was to develop a prediction model of ICH risk, and validate the performance independently using the Multicenter AVM Research Study (MARS).
Methods:
We used 3 BAVM cohorts from MARS: the UCSF Brain AVM Study Project (n=726), Columbia AVM Study (COL, n=640), and Scottish Intracranial Vascular Malformation Study (SIVMS, n=218). Cox proportional hazards analysis of time-to-ICH in the natural course after diagnosis was performed, censoring patients at first treatment, death, or last visit, up to 10 years. UCSF served as the model development cohort. We chose a simple model, including known risk factors that are reliably measured across cohorts (age at diagnosis, gender, initial hemorrhagic presentation, and deep venous drainage); variables were included without regard to statistical significance. Tertiles of predicted probabilities corresponding to low, medium, and high risk were obtained from UCSF and risk thresholds were validated in COL and SIVMS using Kaplan-Meier survival curves and log-rank tests (to assess whether the model discriminated between risk categories).
Results:
Overall, 82 ICH events occurred during the natural course: 28 in UCSF, 41 in COL, and 13 in SIVMS. Effects in the prediction model (estimated from UCSF data) were: age in decades (HR=1.1, 95% CI=0.9-1.4, P=0.41), initial hemorrhagic presentation (HR=3.6, 95% CI=1.5-8.6, P=0.01), male gender (HR=1.1, 95% CI=0.48-2.6; P=0.81), and deep venous drainage (HR=0.8, 95% CI=0.2-2.8 P=0.72). Tertiles of ICH risk are shown in the
Figure
, demonstrating good separation of curves into low, medium and high risk after 3 years in UCSF (left, log-rank P=0.05). The model validated well in the COL referral cohort with better discrimination of curves (middle, P<0.001). In SIMVS, a population-based study, the model separated curves in the earlier years but a consistent pattern was not observed (right, P=0.51), possibly due to the small number of ICH events.
Conclusion:
Our current prediction model for predicting ICH risk in the natural history course validates well in another referral population, but not as well in a population cohort. Inclusion of additional cohorts and risk factors after data harmonization may improve overall prediction and discrimination of ICH risk, and provide a generalizable model for clinical application.
CT Diagnosis of Complications Arising from the Natural History and Treatment of COVID-19
