Curating Naturally Adversarial Datasets for Learning-Enabled Medical Cyber-Physical Systems

Synthetic adversarial tests don't reflect real-world healthcare challenges: The paper demonstrates that existing robustness testing methods, which add imperceptible perturbations to data, fail to capture naturally occurring adversarial scenarios in medical applications—meaning AI systems that pass current tests may still fail in actual clinical settings.

Natural adversarial examples are more relevant for healthcare AI: The researchers propose evaluating medical AI models against "naturally adversarial datasets" (real data that is inherently challenging) rather than artificially manipulated examples, which better represents the messy, complex conditions AI systems will encounter in practice.

Weak labeling can help identify difficult cases: The method uses probabilistic labels from multiple noisy labeling heuristics to identify which real-world examples are most challenging for models to predict, enabling researchers to systematically build datasets that test AI robustness without requiring expensive manual annotation.

Approach validated across multiple medical domains: The evaluation covers six medical case studies, demonstrating that this dataset curation method is broadly applicable across different healthcare applications, not just a single narrow use case.

Trustworthy AI requires testing against realistic scenarios: The paper emphasizes that building trustworthy medical AI systems requires moving beyond laboratory-style adversarial testing toward evaluation methods that reflect actual clinical complexity and variability.