Learning to Defer with Guidance: 3-way triage on real-world medical data

Extending medical triage beyond binary autonomous vs defer decisions: a framework that routes cases to autonomous prediction, human clinician, or human with AI guidance.

Primary source: arXiv:2609.26384: https://arxiv.org/abs/2609.26384

What’s new: Traditional Learning to Defer (L2D) systems make an all-or-nothing choice: either the model predicts autonomously or it passes the case to a physician. In Learning to Defer with Guidance on Real World Medical Data (arXiv:2609.26384, MICCAI 2026 HAIC), Sun, Strong, and Noble expand the decision space to three options: predict autonomously, defer to a clinician, or defer to a clinician accompanied by AI guidance.

Tested on multi-label chest radiography datasets with real clinician annotations (Collab-CXR, VinDr-CXR, and CheXpert) in a predictor-rejector setup, the guided triage policy outperformed standard two-way L2D, standalone human reading, and unguided deferral baselines. Crucially, the authors demonstrate this performance using simpler loss objectives than typical complex surrogate losses in prior L2D literature.

Why it matters: Real clinical workflows are collaborative, not adversarial handoffs. When an AI model defers a borderline or subtle finding, providing saliency or candidate differentials alongside the deferral reduces clinician cognitive load. Testing on datasets with recorded clinician disagreements makes the evaluation realistic rather than synthetic.

Caveats: The study evaluates fixed predictor-rejector pipelines; joint training of guidance generation and rejection policies in live clinical settings remains to be demonstrated. Performance gains depend on clinician receptivity to AI prompts without introducing automation bias.