NAACCR Releases Statement on Responsible Use of AI

Artificial Intelligence (AI) is a fast-growing field with practical applications in cancer registries and cancer surveillance. A recent Narrative article by NAACCR President Eric Durbin, DrPH, MS, highlighted new priorities identified by the NAACCR Board of Directors, including the advancement of effective and responsible AI use. Dr. Durbin notes, “…if AI is not developed and adopted carefully and thoughtfully, cancer registry data quality, completeness, accuracy, reliability, fairness, and public trust could suffer greatly.” In a first step toward supporting the careful and thoughtful development and adoption of AI tools, the NAACCR Board of Directors has approved the following official position statement.

NAACCR Position Statement for the Responsible Adoption of Artificial Intelligence in Cancer Registry Data Collection

Cancer registries are a trusted source of high-quality, standardized cancer data that advances cancer prevention, detection, treatment, survival, and research. North American Association of Central Cancer Registries (NAACCR) leadership believes that registries can play a leading role in improving cancer data collection using artificial intelligence (AI) tools.

Computational systems, including but not limited to machine learning, natural language processing, generative AI, and agentic AI, offer significant opportunities to enhance the speed, precision, efficiency, and depth of cancer data collection. However, deploying these technologies requires a steadfast commitment to safeguarding patient privacy, maintaining data integrity, empowering cancer registrars, and preserving public confidence.

Artificial intelligence can optimize cancer registry operations across several key areas:

• Identification: Case-finding and detecting potentially reportable clinical instances.
• Extraction: Comprehending and processing electronic health records, pathology reports, laboratory results, and imaging data.
• Processing: Summarizing clinical histories and automating abstracting, coding, and staging protocols.
• Record Consolidation: Consolidating records, distinguishing cases, applying multiple primary rules, and resolving duplicates.
• Quality Assurance: validating coding and identifying systemic errors.

AI tools that interact with cancer data must be evaluated for security and their specific intended task, population, data sources, workflow, and consequences of error. AI tools, particularly generative AI and Large Language Models, must prevent clinical hallucinations, synthetic data generation, or fabrication of source documentation.

Core Principles

NAACCR encourages all entities involved in cancer data collection, from the individual cancer registrar to healthcare systems, software vendors, and national organizations, to review the following principles:

• Empowering Human Expertise: Automated systems must complement, not replace, the clinical judgment of cancer registry professionals. High-quality cancer data collection relies on expert registrar training, experience, and accountability.
• Upholding Public Trust: All of us bear ultimate accountability for data accuracy, confidentiality, and proper stewardship. Entities collecting cancer data retain responsibility for data accuracy, confidentiality, and statutory compliance.
• Rigorous, Context-Specific Validation: Prior to deployment, AI solutions must be tested and validated. Testing should consider the context of the deployment environment. Testing should consider potential bias that could result from unique characteristics of patient population, cancer types, or rare conditions.
• Auditability and Data Provenance: Registry workflows must clearly differentiate between human entries, original source documentation, and AI-generated coding and text. Automated tools must never fabricate clinical facts.
• Continuous Lifecycle Monitoring: Pre-deployment validation is insufficient. Systems require ongoing auditing to detect performance drift when clinical standards, reporting standards, or underlying data streams evolve.
• Data Security and Privacy: Protected patient data must remain within verified, secure infrastructure. Solutions must be explicit about how they handle protected data. Entities must ensure they understand how or whether their data will be used for external model training. Entities must ensure they follow institutional, provincial, state, and federal guidelines for handling protected data with the usage of AI tools.
• Regulatory Governance: Deploying AI tools for cancer data collection must strictly align with provincial, state, and federal regulations, institutional policies, and standard-setter reporting guidelines. Integrating automated technologies does not transfer legal compliance or reporting obligations away from reporting entities to software vendors or third-party developers. Entities must conduct legal and privacy reviews prior to deployment and continually evaluate systems to ensure ongoing compliance as provincial, state, and federal rules evolve.

Minimum Operational Guidelines

As this technology continues to evolve, it is important for all stakeholders to be thoughtful about minimal operating guidelines, such as:

• Accountability Constraints: Integrating AI solutions does not shift legal compliance, data security, or reporting responsibilities away from registries, cancer registry professionals, or clinical professionals to technology developers or third-party vendors.
• Procurement & Administrative Oversight: Healthcare administration and procurement teams must include cancer registrars and registry leadership in the evaluation and selection of any AI solutions impacting cancer data workflows prior to acquisition and deployment. Entities must ensure AI solution agreements explicitly prohibit the use of protected health information (PHI) or identifiable registry data for external, global, or multi-tenant model re-training without explicit, compliant Data Use Agreements.
• Pre-Implementation Assessment: Prior to deployment, entities must specify clear use cases, risk thresholds, human oversight workflows, and ownership. Appropriate legal, privacy, and technical reviews are essential, with high-consequence tasks (e.g., automated data releases or case reportability) subject to strict scrutiny. Providers of this technology must provide the entity documentation on the performance of their product, must attest to their validation methods, and must be transparent about data use and model training. Validation should encompass all targeted data sources and all applicable cancer types.
• Operational Controls: In production environments, AI suggestions must be flagged and traceable and should display decision factors alongside recommendations. Reviewers require full visibility into source documentation and a clear understanding of system boundaries. Reviewers must retain full authority to accept, modify, or reject reportability and coding recommendations. High-risk, low-confidence ambiguous outputs should automatically escalate to cancer registrar reviewers, and automated outputs should never be presumed correct by default.
• Post-Implementation Audit: Entities must routinely analyze system precision, error patterns, and drift across varied settings and versions. Metrics should track false positives/negatives, missing cases, and performance variances across all cancer types and sites. Reporting standard changes and coding updates necessitate comprehensive re-validation by both vendor and registry to ensure reliability, completeness, and accuracy. Clear procedures must exist for resolving errors, correcting downstream records, and decommissioning substandard tools.

NAACCR’s Future AI Roadmap

To support these principles, NAACCR is committing to the future development of:

1. Data Standards & Metadata Items: Establishing standard data items to identify, track, and audit AI-assisted vs. human-curated data entries.
2. Evaluation Frameworks: Exploring standard validation criteria and cost-effective frameworks for evaluating AI methods and models.

NAACCR’s Call to Action

The adoption of responsible AI practices should strengthen the security, accuracy, completeness, timeliness along with the breadth, depth, and value of cancer data.

Individual call to action: NAACCR encourages all personnel involved in cancer data collection to learn and be part of the development, evaluation, and deployment of AI solutions. As subject matter experts, cancer registry staff should develop an understanding of the benefits and risks of AI to effectively evaluate and contribute to AI implementations in their workflows.

Organizational call to action: NAACCR urges cross-sector collaboration among developers, registry professionals, and national organizations to build transparent, reliable, and iterative frameworks that reinforce trust in cancer data. Cancer registries should position themselves to confidently leverage AI solutions to remain stewards of complete, accurate, timely, and relevant standardized cancer data.

The NAACCR Board of Directors would like to acknowledge and thank Benjamin Leo, Valerie Yoder, and Kyle Ziegler for their work developing this statement.