Maat Health Advisory

A successful Global Dialogue on AI Governance should produce outcomes that translate directly into real-world decision-making, particularly in high-stakes domains such as oncology. In oncology, AI systems increasingly inform specific clinical decisions—for example, whether to intensify treatment in patients with unfavorable intermediate-risk prostate cancer. These decisions are not abstract; they directly affect patient outcomes, toxicity, and quality of life. Success would include: 1) A shift toward decision-context governance, where AI is evaluated at the level of specific clinical decisions (e.g., treatment intensification vs. de-escalation). 2) Development of context-sensitive evidence standards, recognizing that different oncology decisions carry different levels of risk and reversibility. 3) Alignment of governance frameworks with clinical workflows, where AI informs discrete steps such as diagnosis, risk stratification, and treatment selection. 4) Clear models of distributed accountability, reflecting multidisciplinary cancer care. Ultimately, success would be defined by whether governance frameworks can be meaningfully applied at the point of care—where AI influences real oncology decisions that affect both patient outcomes and system-level resource utilization.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight
• Protection and promotion of human rights

These priorities reflect the need to align AI governance with real-world use, particularly in oncology where AI directly informs clinical decisions. In prostate cancer, AI-based risk stratification tools are increasingly used to inform whether patients with intermediate-risk disease should receive treatment intensification. These decisions carry meaningful implications for both outcomes and toxicity. "Safe, secure and trustworthy AI" and "transparency and human oversight" are therefore essential, but must be applied at the level of specific clinical decisions, where AI informs-but does not replace-clinical judgment. "Interoperability of governance approaches" is critical, as oncology care and clinical evidence are globally interconnected. Fragmented frameworks may lead to variability in how AI tools are evaluated and adopted across health systems. "Social and economic implications" are also highly relevant, as variation in treatment intensity contributes to both unnecessary toxicity and avoidable healthcare costs. Together, these priorities support governance that is grounded in real clinical use rather than abstract system-level evaluation.

A key cross-cutting issue is the absence of decision-centered governance grounded in real-world domains such as oncology. In oncology, AI systems inform discrete decision points-for example, whether to intensify treatment in prostate cancer based on patient-specific risk. These decisions occur within structured clinical pathways and are influenced by both evidence and clinical judgment. Current governance frameworks often evaluate AI at a general level, without sufficient attention to: - The specific clinical decision being influenced - The context in which that decision is made - The evidence threshold required for that decision This creates a gap between governance and implementation. Incorporating decision architecture into governance-mapping where AI enters clinical workflows and how it influences decisions-would improve alignment with real-world use and support both safety and effectiveness.

In the United States, AI governance in healthcare—particularly in oncology—is advancing, but remains uneven at the level of real-world implementation. One key gap is the disconnect between model-level evaluation and decision-level use. Regulatory pathways (e.g., FDA authorization) provide validation of safety and performance, but do not fully address how AI systems are applied within specific clinical decisions. In oncology, this is particularly evident in scenarios such as prostate cancer risk stratification, where AI tools are used to inform whether to intensify treatment in patients with intermediate-risk disease. These decisions are nuanced and occur within established clinical workflows, yet governance frameworks rarely specify how much evidence is sufficient to influence a given decision. This creates variability in: How clinicians interpret and trust AI outputs How consistently these tools are adopted across institutions How decision-making is documented and audited A second gap relates to integration into clinical workflows. Even when AI tools are clinically validated and authorized, there is limited guidance on how they should be incorporated into multidisciplinary decision-making processes. At the same time, there are important opportunities. The U.S. already has a structured pathway—clinical validation, regulatory authorization, and guideline-informed care—that can support AI integration. Emerging tools in oncology, including those used in prostate cancer risk stratification, demonstrate how AI can be aligned with real clinical decisions. This creates an opportunity to move toward decision-centered governance, where AI is evaluated not only for performance, but for its role in specific clinical contexts. Such an approach could improve consistency, support clinician trust, and ultimately enhance both patient outcomes and resource utilization.

The AI Dialogue can advance international cooperation by aligning governance frameworks around shared approaches to real-world use cases, particularly in globally relevant domains such as oncology. Cancer care is inherently international, with shared evidence bases, clinical trials, and treatment guidelines. AI tools used in oncology—such as those informing prostate cancer risk stratification—are therefore evaluated and adopted across multiple health systems. The Dialogue can support convergence on: - Context-sensitive evidence standards for clinical AI - Decision-level evaluation frameworks - Shared approaches to accountability and oversight By focusing on concrete use cases like oncology decision-making, the Dialogue can move beyond abstract alignment and toward practical interoperability.

In healthcare, and particularly in oncology, there is already a well-established pathway for translating innovation into clinical practice. A clear example can be seen in AI-based risk stratification tools in prostate cancer. Systems such as those developed by Artera AI have followed a structured progression: - Clinical validation through peer-reviewed studies - Regulatory authorization through the U.S. Food and Drug Administration (e.g., De Novo pathway) - Integration into clinical practice, including incorporation into guideline-based care (e.g., National Comprehensive Cancer Network) These tools are used to inform specific clinical decisions, such as whether to intensify treatment in patients with intermediate-risk prostate cancer—an area where decision-making is inherently nuanced and patient-specific. This progression reflects a governance model that is: - Anchored in real clinical use cases - Aligned with existing decision pathways - Structured around evidence, regulation, and implementation Rather than developing entirely new frameworks, there is an opportunity to connect and extend these domain-specific approaches into a more globally coherent governance structure.

In oncology, effective AI governance requires coordination across stakeholders involved in clinical decision-making: - Clinicians provide insight into how AI influences real treatment decisions - Patients experience the impact in terms of outcomes and quality of life - Developers define system capabilities and limitations - Regulators establish safety and validation standards - Health systems and payers influence adoption and implementation For example, in prostate cancer, decisions about treatment intensification involve trade-offs between benefit and toxicity, making stakeholder alignment essential. The AI Dialogue should incorporate: - Case-based discussions grounded in oncology scenarios - Domain-specific working groups - Iterative engagement across stakeholders

Domain-level practitioners, particularly clinicians involved in oncology decision-making, are underrepresented in global AI governance discussions. In practice, clinicians must determine how AI outputs influence specific decisions—for example, whether to intensify treatment in prostate cancer based on AI-derived risk. These decisions involve: - Clinical judgment - Patient preferences - Interpretation of evidence Additional underrepresented groups include: - Patients navigating treatment decisions - Clinicians in resource-limited settings - Health system operators responsible for implementation Incorporating these perspectives will improve both relevance and impact.

Domain-level practitioners, particularly clinicians involved in oncology decision-making, are underrepresented in global AI governance discussions. In practice, clinicians must determine how AI outputs influence specific decisions—for example, whether to intensify treatment in prostate cancer based on AI-derived risk. These decisions involve: - Clinical judgment - Patient preferences - Interpretation of evidence Additional underrepresented groups include: - Patients navigating treatment decisions - Clinicians in resource-limited settings - Health system operators responsible for implementation Incorporating these perspectives will improve both relevance and impact.

A concrete example of effective AI governance can be seen in oncology, particularly in prostate cancer decision-making. AI-based tools, such as those developed by Artera AI, are designed to inform risk stratification and guide treatment decisions in patients with intermediate-risk disease. A key clinical question is whether to intensify treatment by adding hormone therapy. This represents a well-recognized gray zone: Some patients benefit from intensification Others experience unnecessary toxicity without added benefit This decision also has broader implications: Overtreatment may lead to unnecessary toxicity and increased healthcare utilization Undertreatment may result in disease progression and more intensive interventions These tools do not replace clinical judgment. Instead, they inform a specific decision point within a structured clinical pathway. This model highlights key governance principles: Evaluation tied to specific decisions, not general performance Alignment with clinical workflows Clear role as decision support within physician-led care This provides a concrete example of how AI governance can be aligned with real-world implementation.