SUDO: a framework for evaluating clinical artificial intelligence systems without ground-truth annotations

SUDO addresses a critical real-world problem: Clinical AI systems often fail when deployed on new data from different hospitals or electronic health record systems, yet evaluating performance is difficult when ground-truth labels don't exist—a common situation in actual clinical practice.

The framework uses a novel temporary labeling approach: SUDO assigns provisional labels to unlabeled real-world data and trains multiple models; the best-performing model indicates the most likely correct label, allowing researchers to assess AI reliability without manual annotation by experts.

Validated across multiple clinical domains: The researchers demonstrated SUDO's effectiveness on three different types of medical AI systems (dermatology images, histopathology patches, and clinical reports), suggesting the approach has broad applicability in healthcare settings.

Enables bias detection in real-world conditions: SUDO allows researchers to identify algorithmic bias in AI systems when applied to actual clinical data without ground-truth annotations—a capability previously unavailable that is essential for ensuring fair and equitable AI deployment.

Improves clinical AI safety and trustworthiness: By identifying unreliable predictions that warrant human review and detecting potential biases before deployment, SUDO helps ensure that clinical AI systems can be used responsibly and that clinicians know when to trust or question AI recommendations.