Can a single output score effectively distinguish urgent from non-urgent cases on non-contrast head CT studies?

Post-processing of AI model (Annalise Enterprise CTB) output into a single score, with 51/130 findings included as urgent. 2,807 non-contrast head CT cases were included, ground-truthed by consultant radiologists. Bootstrapping method on 2,000 cases with 12.5% prevalence of urgent findings was performed at different thresholds for non-urgent predictions (25%, 40%, 50%).

The post-processed output achieved an AUC of 0.92 on the overall dataset (42% urgent cases). On sampled datasets for 50, 40, and 25% non-urgent thresholds, sensitivity was 0.97 – 0.99, specificity was 0.40 – 0.18, with 7 – 1 false negatives.

Single score predictions from the AI model showed effective performance for distinguishing urgent from non-urgent cases, with potential clinical efficacy and financial benefits in settings, such as emergency departments.