Abstract WP180: Novel Machine Learning Model for Prediction of Futile Recanalization in Acute Ischemic Stroke Patients With Anterior Circulation Large Vessel Occlusion
Conor Cunningham, Youssef Zohdy, Sameh Samir Elawady, Hidetoshi Matsukawa, Mohammad-Mahdi Sowlat, Kazutaka Uchida, Sara Zandpazandi, Atakan Orscelik, Ilko Maier, Sami Al Kasab, Pascal M Jabbour, Joon-Tae Kim, Stacey C Quintero, ansaar rai, Robert Starke, Marios Psychogios, Edgar A Samaniego, Adam S Arthur, Shinichi Yoshimura, Hugo Cuellar, Brian Howard, Ali M Alawieh, Daniele G. Romano, Omar Tanweer, Justin Mascitelli, Isabel Fragata, Adam Polifka, Joshua Osbun, Roberto Crosa, Charles C Matouk, Min S Park, Michael Levitt, Waleed Brinjikji, Mark Moss, Travis Dumont, Richard Williamson, Pedro Navia, Peter Kan, Reade A De Leacy, Shakeel A Chowdhry, Mohamad Ezzeldin, Alejandro M Spiotta- Advanced and Specialized Nursing
- Cardiology and Cardiovascular Medicine
- Neurology (clinical)
Introduction: Up to 50% of acute ischemic stroke (AIS) patients who undergo successful mechanical thrombectomy (MT) fail to achieve favorable outcomes (futile recanalization). In this study we aim to develop a machine learning (ML) model to predict futile recanalization (FR) in AIS patients who undergo MT.
Methods: We used data from an ongoing large, multicenter database from 2013 to 2023. We included AIS patients treated with MT for ICA, M1, or M2 occlusion with successful recanalization (modified Thrombolysis in Cerebral Infarction [mTICI] score ≥ 2C) and procedure durations under 60 minutes. FR was defined as successful recanalization with 90-day modified Rankin Scale (mRS) 3-6. The dataset was divided into 75% for training and 25% for external validation. Using the Caret Package in R, multiple models were tested, and their performances were evaluated by the area under the curve (AUC) of receiver operating. Both baseline and pre-interventional characteristics were incorporated into the model. The selected model was then externally validated on a 25% validation dataset.
Results: Among 2,546 qualified patients, FR occurred in 1,342 (52.7%). In univariate analysis, baseline characteristics were significantly different between FR and non-FR groups. The M5P model demonstrated the highest performance (AUC: 0.833; 95% CI: 0.7989-0.852; PPV: 0.8101) in comparison to other tested models such as logistic regression (AUC: 0.74), RF (AUC: 0.78), J48 (AUC: 0.78), SVM (AUC: 0.79), and GB (AUC: 0.79). The external validation of the model showed satisfactory results (AUC: 75.25; 95% CI: 70-80; PPV: 76.87).
Conclusion: Utilizing clinical, pre-procedural, and imaging parameters, the M5P model can efficiently predict F) in AIS patients before attempting MT. This tool can assist neurointerventionalists in adequately choosing their MT candidates.