Introduction: Safe surgical resection of brainstem cavernous malformations (BSCMs) can be challenging. Furthermore, the balance of operative risk versus benefit remains elusive. In the present analysis, unsupervised and supervised machine learning (ML) models are utilized to understand clusters of patient outcomes.
Methods: A retrospective analysis of BSCMs at a high-volume center was performed from 1987-2022. 32 features were extracted with minimal missing data, with 3 outcome features in the postoperative period. Uniform Manifold Approximations and Projections (UMAP) and K-means clustering were used to identify unique clusters of outcomes. Recursive Feature Extraction (RFE) was used to identify highly predictive preoperative features mapped to each cluster. A logistic regression model predicted membership into the worst-performing cluster, from which a score was constructed and validated to predict cluster membership.
Results: 680 patients were identified to have BSCMs, of which 413 had complete outcome data. Four clusters were identified, of which cluster 3 was found to have the poorest post-operative Glasgow outcome scale (GOS) and modified Rankin scale (mRS) (p < 0.001). RFE, narrowing predictive variables from 32 to 20, demonstrated the size of the lesion (VI = 5.7, OR = 1.59, p = 0.04, and pre-operative CN10 deficit (VI = 3.72, OR = 7.68, p = 0.01) were the two most predictive. Our model with these variables was 85.23% accurate, predicting cluster 3 membership. A score was constructed from the beta-coefficients of the 6 most predictive variables. This score, thresholded at 5.7, was 83.22% accurate, with an AUC of 0.72 in predicting cluster 3 membership.
Conclusion : Herein, unsupervised and supervised ML was utilized to identify patients that perform poorly after surgery for BSCM. We also internally validated a score constructed from our ML model with strong performance. Further studies ought to be done to make ML approaches clinically utilizable.
