(AS2) Novel Outcome Clusters of Patients with Aneurysmal Subarachnoid Hemorrhage Using Unsupervised Machine Learning: A Post-hoc Analysis of Pooled PBRAT and BRAT Data
Joshua Catapano: No financial relationships to disclose
Introduction: Patients presenting with ruptured aneurysms can often have poor outcomes despite aggressive early intervention. Multiple outcome metrics presently exist to quantify different aspects of their status and improvement during their hospitalization. We sought to identify morphologies of patient outcomes in ruptured aneurysms by using unsupervised machine learning (UML) approaches to the Barrow Ruptured Aneurysm Trial (BRAT) and patients in the post-BRAT (PBRAT) era.
Methods: The BRAT and PBRAT registries were queried, and common variables were pooled. Uniform Manifold Approximations and Projections (UMAP) were used to dimensionally reduce the length of stay, discharge, and one year follow-up modified Rankin Scale (mRS), retreatment, observation of recurrence or residual aneurysm, and the occurrence of delayed cerebral ischemia (DCI). K-means was used to cluster the dimensionally reduced data. Firth’s logistic regression was used to identify predictors of cluster membership.
Results: A total of 1444 patients were identified. Cluster 2 varied significantly from cluster 1, including length of stay(SMD=2.391,p < 0.001), DCI(SMD=0.303,p < 0.001), and final mRS(SMD=2.388,p < 0.001). Important predictors of cluster 2 included age(p=0.028), peripheral vascular disease(p=0.008), aneurysm size(p=0.026), Hunt-Hess 5(p < 0.001), and Fisher Grade 4(p=0.013). Cluster 2(N=1262) was further sub-clustered, into 2 sub-clusters. LOS was longer in sub-cluster 1(SMD=2.247,p < 0.001), and had higher rates of DCI(50.9% vs. 21%,SMD=0.655,p < 0.001). Membership of sub-clusters was driven primarily by higher Hunt-Hess Grades 3 through 5(p < 0.01).
Conclusion : We demonstrate unique morphologies of recovery and progression after ruptured aneurysms by applying UML clustering to BRAT and PBRAT. Predictors of improvement and poor outcomes in our analysis remain consistent with previously published analyses in the literature. Furthermore, the present study demonstrates the utility of machine learning for large databases in cerebrovascular diseases.