Abstract Body: Purpose: Fast progressors (FP) of infarct growth in anterior circulation occlusions at 0-6 hours from onset (early window) have been associated with reduced rates of favorable outcome post endovascular thrombectomy (EVT). FP can be identified by using CT perfusion (CTP) derived infarct core volume (ICV) to calculate infarct growth rate (IGR), but has disadvantages of added time/radiation and susceptibility to motion. CTP is often performed with CT angiography (CTA) but AHA/ASA guidelines do not require perfusion imaging in the early window for EVT decision making. Our aim was to determine if machine learning (ML) can identify FP with only clinical, laboratory, CT, and CTA variables, without using CTP data.

Methods: We retrospectively included consecutive stroke patients arriving 0-6 hours from symptom onset with ICA and/or MCA occlusion on CTA and had concurrent CTP. RAPID AI generated maps of ICV and hypoperfusion volume with criteria of rCBF <30% and Tmax >6 seconds. IGR was calculated as ICV/onset-to-CT time. FP was defined as IGR ≥10 mL/hour. Eight ML algorithms for binary classification were used with 11 selected features including age, NIHSS, lab values, ASPECTS, clot burden score (CBS), and dichotomized collateral score (Tan method 0-1 vs 2-3). Data was randomly split 70:30 for training and testing. Synthetic Minority Oversampling Technique was used to balance the class distribution in the training set. 10-fold cross validation was assessed during training. The performance of the models during testing was compared using ROC AUC, precision (PPV), recall (sensitivity), and accuracy. Relative weights of features during testing were examined with Shapley Additive Explanations (SHAP).

Results: 147 patients were included with median (IQR) age 76 (67-85), onset-to-CT 2.5 hours (1.5-4), NIHSS 19 (12-24), ASPECTS 9 (8-10), CBS 7 (6-8), collateral score 1 (1-2), ICV 11 mL (0-34), hypoperfusion volume 104 mL (67-157), and IGR 3.4 mL/hour (0-12.3). 47 (32.0%) were labeled as FP with median IGR 24.0 mL/hour (13.3-40.0). 16 (34.0%) of these FP also had ICV >70 mL upon arrival. XGBoost was the best performing ML model with ROC AUC 0.90, precision 0.76, recall 0.93, and accuracy 0.89. SHAP analysis of this model ranked collateral score, time from onset, serum glucose level, systolic blood pressure, and ASPECTS as the top 5 most important features.

Conclusion: ML can accurately identify FP in the early window without using CTP.