ML models can simplify complex data such as neuroimaging into smaller representations called embeddings in order to expedite data processing. Many clinical variables have been previously studied to predict stroke outcomes, but here we propose a novel approach using embedding data from emergently-acquired computed tomography angiography (CTA).

Consecutive AIS patients with anterior circulation LVO from an international EVT registry were included. 256-dimensional embeddings were obtained for each patient’s pre-EVT CTA using a pre-trained deep learning algorithm. Two subsequent models were trained and tested using 5-fold cross validation to predict successful first-pass reperfusion (TICI 2b-3) and 90-day modified Rankin scale (mRS) 0-2. Our first model used the CTA embedding data (“ML-CTA”). The second model utilized the Alberta Stroke Program Early CT Score (“ML-ASPECTS”). The models were evaluated using area under the ROC curve (AUC) analysis and compared using the Mann-Whitney U test.

From 35 institutions, 285 patients were included. Mean age was 70 ±14, 55.8% were female, median NIHSS was 14 [8-19], median ASPECTS was 8 [7-9], 51.9% received tPA, 50.5% achieved TICI 2b-3 within 1 pass, and 60.4% had mRS 0-2 at 90 days. The ML-CTA model had the same performance as ML-ASPECTS for predicting first pass TICI 2b-3 (AUC 0.590 vs 0.606, p=0.045) and better performance for predicting 90-day mRS 0-2 (AUC 0.623 vs 0.595, p<0.001).

Machine learning models based on CTA embedding data perform better than models that utilize provider-assessed ASPECTS when predicting functional independence for patients with LVO AIS undergoing EVT. Further study is needed to refine the final model performance and to evaluate if fully automated imaging analysis can decrease time to treatment and improve final clinical outcomes.