Introduction Mechanical Thrombectomy (MT) is the standard-of-care in the interventional management of Acute Ischemic Stroke (AIS). The NVQI-QOD AIS Thrombectomy Registry documents detailed patient characteristics, pre-operative imaging, procedural metrics, and post-operative outcomes. Although these data are highly informative, there is inherent uncertainty in all medical interventions, so patient outcomes remain variable after intervention.

Methods We identified three groups of feature variables from the NVQI-QOD registry, including data available prior to MT (Group Preop), post MT (Group Postop), and at discharge (Group DC). We introduced a Probabilistic Neural Network (PNN) that predicts the expected distribution of NIH Stroke Scale (NIHSS) changes from pre-intervention to discharge and the binary severity categories derived from a 90-day follow-up modified Rankin Scale (mRS), using the three groups of feature variables as inputs. Numerous machine learning studies and competitions have demonstrated that XGBoost is a high-performance prediction model, so the predictions of the PNN were compared to those predicted by XGBoost. Furthermore, both PNN and XGBoost were trained using bagging ensemble learning, a technique for training an ensemble of multiple member models based on bootstrapping to improve prediction robustness.

Results For both regression and classification, there were almost no differences in the prediction performance between the PNN and XGBoost ensembles. Prediction accuracy was improved when more correlated feature variables were available (from Group Preop to Group DC). For example, the best performance was achieved using Group Postop in regression of NIHSS changes (RMSE: 4.34 for PNN vs. 4.30 for XGBoost) and using Group DC in classification of mRS severity (accuracy: 0.78 for PNN vs. 0.77 for XGBoost; the same trend using different feature groups can be found in other measures, such as precision, recall, F1-scoure, and AUC-ROC). Furthermore, PNN accurately described the distributions of NIHSS changes represented by predicted means and SDs. Notably, in Group Preop, even patients with the worst predicted outcomes had an approximately 50% chance of improvement. Feature importance analysis showed that both the predictions of the NIHSS changes and mRS severity primarily relied on earlier NIHSS, Pre mRS, and patient age.

Conclusions This study demonstrates the utility of probabilistic ensemble learning in clinical decision-making and prognosis. It can provide robust predictions as well as quantify data uncertainty. Our results regarding NIHSS changes reinforce the substantial benefits of MT, that can improve outcomes in nearly half of patients. The degree of disability relevant to the 90-day follow-up mRS can be determined by probabilistic learning available as early as discharge.

Disclosures C. Zhou: None. S. Faruqui: None. A. Patel: None. R. Abdalla: None. A. Shaibani: None. M. Potts: None. B. Jahromi: None. S. Ansari: None. D. Cantrell: None.