Jae Hyun Lee

Medical AI is often evaluated using aggregate measures of discrimination, calibration, and accuracy. However, these measures can obscure clinically important variation across patient groups, institutions, devices, and workflows. This viewpoint defines refined exclusion as a governance condition in which an AI system appears successful in aggregate, while uncertainty, error, or reduced clinical reliability is concentrated in populations that are insufficiently represented, measured, validated, or monitored. The concept does not replace algorithmic fairness, hidden stratification, dataset shift, or subgroup performance analysis. It connects these mechanisms to a distinct consequence: an unequal distribution of safety that remains inadequately detected or corrected. Drawing on purposively selected, illustrative evidence from population health management, chest radiography, dermatology, computational pathology, medical foundation models, and clinical measurement, we distinguish model-level disparity, patient safety signals, and documented patient harm. We then frame data justice as a complementary governance approach with distributional, procedural, and substantive dimensions. The proposed lifecycle decision gates address intended use, subgroup learnability, data provenance, validation, procurement, local deployment, monitoring, updates, and patient feedback. Each gate links minimum evidence to decision authority and 1 of 4 actions: proceed, enrich or validate, restrict use, or pause or retire. Governance intensity should be proportionate to clinical risk and evidentiary uncertainty. By linking subgroup evidence gaps to institutional decisions and corrective action, the framework shifts attention from whether a model performs well on average to whether its safety is demonstrable for the populations and settings in which it will be used.

UNSTRUCTURED Medical artificial intelligence (AI) systems are often evaluated through aggregate performance metrics and output-level fairness measures. However, clinically meaningful harms may remain hidden when systems perform well on average while underperforming for data-poor, underrepresented, or structurally marginalized populations. This Viewpoint uses the concept of refined exclusion to synthesize a recurring pattern in medical AI: systems may appear technically successful at the population level while transferring uncertainty, misclassification, delayed recognition, or reduced clinical reliability to groups that are less visible within training data, validation cohorts, proxy definitions, and deployment workflows. Drawing on representative cases from population health management, chest radiograph AI, dermatology, computational pathology, and foundation model applications, we argue that refined exclusion should not be treated merely as algorithmic bias or a defect of model outputs. Rather, it reflects a data governance failure with direct implications for patient safety. Moving beyond output-centered algorithmic fairness, we propose data justice as a governance foundation for medical AI, organized across distributional, procedural, and substantive dimensions. We further outline operational checkpoints across the medical AI lifecycle, including subgroup learnability assessment, data provenance documentation, local validation, procurement-stage accountability, explainability-based proxy audits, post-deployment subgroup monitoring, and patient participation. Reframing refined exclusion as a patient safety problem shifts the central governance question from “Is this model accurate on average?” to “For whom is this system safe, reliable, and clinically accountable?”

Artificial intelligence (AI) in health care is increasingly defined not by static algorithms but by adaptive intelligence—systems that evolve over time through interactions with data, clinicians, and clinical environments. This adaptive capacity creates a structural mismatch with regulatory frameworks built for technologies whose behavior remains static. As AI models drift, recalibrate, or degrade in real-world contexts, they dissolve the linear boundaries between design, deployment, and clinical interpretation. These temporal, epistemic, and organizational frictions expose responsibility gaps that cannot be resolved through incremental modifications to legacy oversight structures. Regulators across major jurisdictions are beginning to respond to these challenges, though with differing orientations. The United States advances mechanisms for predictable adaptation, including Predetermined Change Control Plans, real-world evidence frameworks, and life cycle–oriented quality management reforms. The European Union emphasizes precautionary, rights-based governance through the European Union Artificial Intelligence Act (AI Act) and modernized liability rules. South Korea, operating within a hyperconnected digital health ecosystem, has introduced the Digital Medical Products Act (DMPA), one of the world’s first comprehensive statutory frameworks for learning medical AI. Despite philosophical differences, these regulatory trajectories converge on a shared insight: learning AI systems cannot be governed by static rules or episodic evaluation. This viewpoint proposes Co-Lifecycle Governance as a conceptual framework to synchronize regulatory oversight with adaptive intelligence. Rather than treating oversight as a discrete event, Co-Lifecycle Governance frames regulation as a continuous, synchronized process grounded in 4 pillars: continuous validation, agile change management, proactive performance surveillance, and distributed accountability. Each pillar functions as a structural antidote to the responsibility frictions that arise when AI systems evolve faster than expectations surrounding them. Together, these pillars provide a governance grammar capable of supporting safe, iterative model improvement while maintaining system-level trust. Drawing from the strengths of US predictability, European Union accountability, and Korean scalability, this paper outlines a hybrid convergence pathway that synthesizes predictability, accountability, and operational feasibility. Learning AI will not wait for governance to catch up; oversight must evolve in lockstep with adaptive intelligence. Co-Lifecycle Governance offers a foundation for regulatory systems that not only regulate learning AI but also learn with it—at the speed at which adaptive intelligence actually changes.

UNSTRUCTURED Artificial intelligence (AI) in health care is increasingly defined not by static algorithms but by adaptive intelligence—systems that evolve over time through interactions with data, clinicians, and clinical environments. This adaptive capacity creates a structural mismatch with regulatory frameworks built for technologies whose behavior remains static. As AI models drift, recalibrate, or degrade in real-world contexts, they dissolve the linear boundaries between design, deployment, and clinical interpretation. These temporal, epistemic, and organizational frictions expose responsibility gaps that cannot be resolved through incremental modifications to legacy oversight structures. Regulators across major jurisdictions are beginning to respond to these challenges, though with differing orientations. The United States advances mechanisms for predictable adaptation, including Predetermined Change Control Plans (PCCPs), real-world evidence frameworks, and lifecycle-oriented quality management reforms. The European Union emphasizes precautionary, rights-based governance through the AI Act and modernized liability rules. South Korea, operating within a hyper-connected digital health ecosystem, has introduced the Digital Medical Products Act (DMPA), one of the world’s first comprehensive statutory frameworks for learning medical AI. Despite philosophical differences, these regulatory trajectories converge on a shared insight: learning AI systems cannot be governed by static rules or episodic evaluation. This Viewpoint proposes Co-Lifecycle Governance as a conceptual framework to synchronize regulatory oversight with adaptive intelligence. Rather than treating oversight as a discrete event, Co-Lifecycle Governance frames regulation as a continuous, synchronized process grounded in four pillars: continuous validation, agile change management, proactive performance surveillance, and distributed accountability. Each pillar functions as a structural antidote to the responsibility frictions that arise when AI systems evolve faster than expectations surrounding them. Together, these pillars provide a governance grammar capable of supporting safe, iterative model improvement while maintaining system-level trust. Drawing from the strengths of U.S. predictability, EU accountability, and Korean scalability, this paper outlines a global hybrid pathway that synthesizes predictability, accountability, and operational feasibility. Learning AI will not wait for governance to catch up; oversight must evolve in lockstep with adaptive intelligence. Co-Lifecycle Governance offers a foundation for regulatory systems that not only regulate learning AI, but also learn with it—at the speed at which adaptive intelligence actually changes.