Learning from Disagreement: Clinician Overrides as Implicit Preference Signals for Clinical AI in Value-Based Care

Researchers propose a framework that treats clinician overrides of AI recommendations as preference signals for training clinical decision-support systems in value-based care settings. The approach combines preference learning with capability modeling to improve AI alignment with patient outcomes rather than encounter economics, addressing a failure mode called suppression bias.

This research addresses a critical gap in clinical AI deployment: how to systematically learn from the moments when experienced clinicians disagree with algorithmic recommendations. Rather than treating overrides as failures, the authors recognize them as rich behavioral data that reveals what domain experts truly value when stakes are high and outcomes measurable.

The framework's innovation lies in decomposing clinician capability into execution and alignment components, then training dual models that jointly optimize reward signals and capability assessment. This prevents suppression bias—where models learn to suppress valid but difficult recommendations when they exceed clinician execution capability—a particularly dangerous failure mode in healthcare where missing a correct recommendation can harm patients.

The approach gains leverage from value-based care environments where financial incentives align with patient outcomes rather than service volume. These settings naturally produce the longitudinal data, outcome labels, and capability variation necessary for robust preference learning. Unlike consumer applications where RLHF signals come from annotators with limited domain expertise, clinical overrides carry irreplaceable expert judgment backed by observable patient trajectories.

For healthcare organizations, this framework offers a path to systematically improve clinician-AI teams rather than replacing clinicians or blindly following algorithms. The operational implementation demonstrates maturity beyond theory—insights emerged from live deployment in value-based care programs. As healthcare systems increasingly adopt outcome-based payment models, this methodology becomes increasingly applicable, potentially accelerating the transition from algorithmic compliance to genuine shared decision-making where technology amplifies clinical expertise rather than circumventing it.

• →Clinician overrides of AI recommendations contain rich preference signals analogous to RLHF but with domain-expert annotators and measurable real-world consequences.
• →The framework prevents suppression bias by separately modeling clinician capability, preventing AI systems from suppressing correct-but-difficult recommendations.
• →Value-based care environments with outcome-based contracts provide uniquely favorable training conditions: longitudinal density, concentrated decision spaces, outcome labels, and natural capability variation.
• →Joint optimization of reward and capability models enables better alignment with patient trajectories rather than encounter economics or per-interaction metrics.
• →The methodology demonstrates practical applicability through implementation in live clinical deployments, not theoretical research alone.