`Probabilistic ensemble learning for prediction of stroke thrombectomy outcomes from the NeuroVascular Quality Initiative-Quality Outcomes Database (NVQI-QOD) Acute Ischemic Stroke Registry

Abstract

Introduction

Mechanical Thrombectomy (MT) is the standard of care in the interventional management of Acute Ischemic Stroke (AIS). The NVQI-QOD registry records detailed patient characteristics, pre-operative imaging, procedure metrics, and post-operative outcomes of neurointerventional surgical procedures. Although these data are highly informative, there is substantial uncertainty in all medical interventions, so patient outcomes remain variable after intervention. In this work, we leverage a probabilistic machine learning paradigm to predict MT outcomes in the context of this inherent uncertainty.

Methods

Using data from the NVQI-QOD AIS registry, we identified three groups of feature variables: those available prior to MT (Group Preop), post MT (Group Postop), and at discharge (Group DC). Using Probabilistic Neural Networks (PNNs) and XGBoost, we predicted 1) the change in NIH Stroke Scale from presentation to discharge (∆NIHSS), and 2) a binary measure of functional outcome, which was aggregated from the 90-day follow-up Modified Rankin Scale (mRS).

Results

Both XGBoost and the PNN are capable of binary probabilistic classification of mRS scores, with accuracies ranging from 0.69 using preoperative feature variables to 0.80 when utilizing input features that are available at the time of discharge. XGBoost and the PNN had similar mean squared error performance for the ∆NIHSS regression task as well, however the PNN can also perform probabilistic regression, predicting distributions of ∆NIHSS with means and standard deviations (SDs). Feature importance analysis showed that predictions of both ∆NIHSS and mRS severity primarily depended upon the presenting NIHSS, Pre mRS, and patient age.

Conclusions

The probabilistic machine learning paradigm allows for quantification of predictive uncertainty through outcome probability distributions and may offer clinicians critical insights beyond traditional deterministic methods. With the clinical information available prior to MT, patients with the worst predicted outcomes will have a nearly ∼50% chance of neurological improvement, while those with the best anticipated outcomes have a > 98% probability for improvement, reinforcing both the safety and profound benefits of MT.