iSPAN: Explainable prediction of outcomes post thrombectomy with Machine Learning
Brendan S Kelly1, Prateek Mathur2, Silvia D Vaca3
1St Vincent's University Hospital, Dublin, Ireland; Insight Centre for Data Analytics, UCD, Dublin, Ireland; Wellcome Trust - HRB, Irish Clinical Academic Training, Dublin, Ireland; School of Medicine, University College Dublin, Dublin, Ireland; Lucille Packard Children's Hospital at Stanford, Stanford, CA, USA.
European Journal of Radiology
|February 24, 2024
Summary
A new clinical score, iSPAN, improves prediction of outcomes for stroke patients undergoing endovascular thrombectomy. It outperforms existing scores and is externally validated.
Area of Science:
- Neurology
- Medical Informatics
- Clinical Prediction Models
Background:
- Endovascular thrombectomy is a critical treatment for acute ischemic stroke.
- Accurate prediction of patient outcomes is essential for treatment planning and resource allocation.
- Existing clinical scores have limitations in predicting outcomes post-thrombectomy.
Purpose of the Study:
- To develop and evaluate a machine learning (ML) model for predicting outcomes in stroke patients.
- To create a novel clinical score (iSPAN) integrating ML-derived features.
- To compare the predictive performance of iSPAN against existing scores (SPAN, PRE) and ML models.
Main Methods:
- Retrospective analysis of 812 patients with anterior circulation stroke treated between 2010-2020.
- Development of ML models (XGB) and derivation of the iSPAN score by optimizing SPAN with ML features.
- External validation of iSPAN and SPAN on a separate cohort of 63 patients.
Main Results:
- The XGB model showed higher accuracy (0.738) than the SPAN score (0.628).
- Key predictive features identified by ML were Age, mTICI, and number of passes.
- The novel iSPAN score demonstrated comparable accuracy to the XGB model and significantly outperformed SPAN in external validation (accuracy 0.79 vs 0.67).
Conclusions:
- The iSPAN score effectively integrates machine-derived insights for improved outcome prediction in endovascular thrombectomy patients.
- iSPAN is not inferior to advanced ML models and demonstrates strong external generalizability.
- This novel score offers a more accurate and reliable tool for clinical decision-making in stroke management.
