Predictors of Outcome Clusters in Unruptured Aneurysms: An Unsupervised Machine Learning Analysis

Introduction: Identifying patients who are prone to poor outcomes is crucial to adapt treatment and ensure optimal outcomes. Our aim was to use unsupervised machine learning to reveal patient outcome subgroups and predictors of attribution to these groups.

Methods: 400 patients with unruptured aneurysms, microsurgically treated at a large quaternary center from 1/1/2014 to 12/31/2020, were retrospectively reviewed. Outcomes evaluated included new neurological deficit at discharge and poor neurological outcome at follow-up, as defined by mRS > 2. Outcomes were clustered using a k-means model after the number of optimal clusters were obtained using a total within sum of squares (TWSS) algorithm. Uniform maniform projections (UMAP) were used to project the outcomes into a 2-D space. Upon cluster determination, a multivariate regression with p-inclusion of
Results: Clustering analysis yielded 3 unique clusters. Cluster 1 represented patients who had no poor neurological outcome (0%) and no neurological deficits at discharge (0%). Cluster 2 represented patients with a high burden of new neurological deficit at discharge (76/400 (53%)) but no poor neurological outcome (0%). Cluster 3 represented patients with a new neurological deficit at discharge (52/400 (45%), p < 0.001) and a poor neurological outcome (47 (41%), p < 0.001). PHASES scoring was highest in Cluster 3 compared to Cluster 2 and 1 (5.03 (sD 2.83) vs 4.71 (sD 2.56) vs 3.91 (sD 2.66), p < 0.001). With multivariate analysis, diabetes and hyperlipidemia were factors that significantly predicted non-membership Cluster 1. Hyperlipidemia and admission mRS were the only variables that significantly showed increase in stepwise membership among the clusters.

Conclusion : Herein, unsupervised machine learning adequately grouped outcomes into three distinct, clinically meaningful clusters. Vascular comorbidity rather of aneurysm characteristics significantly predicted attribution to these patient outcome clusters.