Introduction The first pass effect (FPE) in mechanical thrombectomy, characterized by successful recanalization of an occluded vessel in the initial thrombectomy attempt, has been demonstrated to predict positive patient outcomes. Unsupervised machine learning (ML) techniques offer a promising approach to analyzing large-scale data without the need for explicit labels or training data, enabling the identification of patterns and relationships that are not readily apparent. In this study, we sought to investigate the predictive value of FPE using unsupervised clustering of outcomes.

Methods Consecutive mechanical thrombectomies were retrospectively tabulated at a large volume stroke center. This study was approved by the local institutional review board (IRB). Clustered outcomes to formulate clusters included the number of passes, the final reperfusion obtained, the discharge NIHSS, and complications. FPE was defined as TICI 2C or 3 reperfusions on the first pass. Uniform manifold approximations and projections (UMAP) and K-means unsupervised clustering were done to identify outcome clusters. Logistic regression with and without interaction coefficients was used to determine associated factors with cluster membership or FPE. P-values less than 0.05 were considered significant.

Results 287 consecutive patients were identified, of which 187 were used in the final analysis due to missing outcomes data. The average age was 70 years (sD = 15.14), with 81.8% of patients with MCA occlusion. Three outcome clusters were identified in the analysis. Cluster 2 had the greatest congruency with FPE (p < 0.001) and was also associated with poor outcomes. Patients in Cluster 2 were more likely to have ICA occlusions and were older (p = 0.043, 0.001, respectively). Predictors of FPE were analyzed using cluster membership and adjusted for aneurysm properties. Adjusted for other factors, cluster 2 membership was significantly associated with achieving FPE in our cohort (OR = 5.419 (95% CI: 2.349 - 13.464), p < 0.001). Next, the predictive value of FPE and cluster membership was compared against Discharge NIHSS. FPE was inversely associated with discharge NIHSS (B (SE) = -4.497 (1.39), p = 0.0014). Cluster 2 membership was also inversely associated with discharge NIHSS (B (SE) = -11.27 (0.77), p < 0.001). Adjusting for interaction, cluster 2 membership was independently more predictive (p < 0.001) while FPE predictive value was no longer significant.

Discussion In this study, we identified novel outcome clusters using unsupervised machine learning. We examined the congruency of the predictive value of FPE against cluster membership. We also demonstrate that cluster membership was able to exceed the predictive value of FPE in predicting discharge NIHSS. This study demonstrates that unsupervised machine-learning methods can augment established metrics to better predict outcomes. Further validation in large multi-institutional cohorts Is required.

Disclosures J. Catapano: None. A. Naik: None. S. Koester: None. R. Singh: None. H. O’stinnington: None. I. Rangel: None. S. Desai: None. E. Winkler: None. A. Ducruet: None. F. Albuquerque: None. A. Jadhav: None.