Probabilistic and Explainable Machine Learning Methods for the Prediction of Technical Success and Stroke Outcomes Following Mechanical Thrombectomy

About This Grant

PROJECT SUMMARY Acute ischemic stroke (AIS) resulting from large vessel occlusion (LVO) is a primary cause of global morbidity and mortality. Mechanical thrombectomy (MT) is the current standard of care for LVO, and it is one of the most effective interventions in modern day medicine. This procedure involves catheter-based endovascular access to the intracranial vasculature for clot retrieval. The number of retrieval attempts required for successful recanalization is variable, influenced by clot characteristics, patient anatomy, and procedural technique. Clinical trials have established the efficacy of MT in AIS, with recent studies demonstrating a strong correlation between successful first-pass recanalization and favorable clinical outcomes. Despite the established population-level efficacy of mechanical thrombectomy (MT) in acute ischemic stroke (AIS), individual patient outcomes exhibit significant variability and prognostic uncertainty. The ability to more accurately predict MT outcomes for individual patients from clinical data would offer highly valuable guidance for clinical decision-making. Prior published works have utilized classic deterministic machine learning methods to generate singular expected outcome estimates, which do not adequately describe the broad distribution of potential outcomes. In this work, we propose to develop new probabilistic machine learning models, which directly infer outcome distributions, providing a more informative assessment of the anticipated prognosis in the setting of real-world clinical uncertainty. For this purpose, we leverage the NeuroVascular Quality Initiative – Quality Outcomes Database, which contains highly granular clinical, procedural, and outcomes data for over 10,000 MT procedures. In our first aim, we will develop multiple state-of-the-art probabilistic models to predict MT procedural success, MT first-pass success, and functional and neurological outcomes from the clinical data available in the NVQI-QOD. As part of this first aim, we will evaluate a new and extremely powerful probabilistic foundational model for Bayesian inference, and compare it to state-of-the-art tree-based probabilistic methods. In our second aim, we will utilize methods from Explainable AI (XAI) to identify the most important predictors of procedural success and clinical outcomes and then to also characterize their interactions. We anticipate that this work will provide valuable insight into the complex determinants of large vessel stroke outcomes and identify high-priority targets for future research.