Abstract P377: Real-Word Performance of Two Automated Software Platforms for Large-Vessel Occlusion Identification in Acute Ischemic Stroke Patients: A Single Center Experience

Abstract

Background and purpose: Modern stroke treatment has been revolutionized by image-guided selection of patients for endovascular thrombectomy. Current automated platforms allow for real-time identification of large vessel occlusion and salvageable brain tissue. We had previously demonstrated that a “CTA for all” policy for stroke patients immediately upon arrival assists in the earlier identification of treatment candidates. We now sought to evaluate the performance of these platforms under this policy. Methods: All patients that presented to Henry Ford Health System hospitals over a period of 6 weeks received CTA of the head and neck upon initial presentation. The images were processed with two automated software platforms. We prospectively collected processing times, large-vessel-occlusion alerts, performance warnings, and LVO density ratios. We compared these with the interpretations of board-certified radiologists, and analyzed the performance of each platform. Results: 276 patients presented with stroke symptoms and received CT angiography upon presentation. Both platforms were able to image all stroke patients within their FDA-approved indications. Both platforms were noted to have comparable sensitivity, specificity, PPV and NPV, and excellent accuracy. The overall prevalence of LVO was extremely low (8/276). As a result, for both, NPV was much better than PPV because of the percentage of false positive results. Further ROC analysis, demonstrated an area under the ROC curve of 0.982, and overall model quality of 0.97. Optimal LVO density cutoff was <0.093 in order to maximize overall accuracy, or < 0.271, in order to maintain a sensitivity of 100% as an absolute priority (both significantly lower than the current threshold of <0.45). Conclusions: Automated software platforms are an invaluable aid in the selection of patients for endovascular thrombectomy. Different LVO detection algorithmic thresholds may be necessary (and should be part of individual stroke center validation pathways) to avoid fatigue alert, and optimize test accuracy, when LVO prevalence is low. Stroke teams should be aware of the limitations of automated analysis and need for expert review.