VETNET: Assembling Benchmark Datasets for AI-Driven Veterinary Innovation

1.1 | The Opportunity

• Veterinary benchmark datasets remain deficient, while species and practice variation, privacy, inconsistent records, and sparse outcomes make direct reuse of human medical benchmarks inadequate.The current infrastructure leaves users vulnerable to shortcut learning, institutional bias, and poor generalization.Source: §1.1 · pp. 4–5

1.2 | Convening the Thought Summit

• The summit brought together 36 participants from six countries and 16 institutions, producing animal-centered, transparent, translational, and enabling working values for VETNET.Its framing kept One Health connections visible without making the first implementation target too broad.Source: §1.2 · pp. 5–6

Collaborative Infrastructure

• VETNET’s multi-institutional collaboration draws on diverse expertise, with clear governance and incentives identified as conditions for coordination.Cloud computing and federated learning are described as ways to enable secure multi-institutional collaboration.Source: Collaborative Infrastructure · p. 6

Data & Technical Potential

• Veterinary institutions hold underused multimodal data, while demand for AI and One Health concerns create an opportunity for veterinary-specific evaluation and benchmarking.The paper identifies agreed standards for dataset documentation, accessibility, and completeness as a scientific imperative.Source: Data & Technical Potential · p. 6, Market & Societal Momentum · p. 6

1.3.2 | Critical Challenges to Address

• Practice systems, local codebooks, and species diversity produce heterogeneous veterinary records, so VETNET should map local notes to interoperable forms and build on existing standards.The paper names DICOM, SNOMED-CT, OMOP CDM, and other resources as possible starting points rather than waiting for a universal ontology.Source: §1.3.2 · p. 7

Resource & Capacity Constraints

• Veterinary medicine’s scale may challenge VETNET, while smaller funding and market size constrain large-scale dataset development and narrow the available AI types, labels, metrics, and validation pathways.The paper calls for funding from industry, government, and academia to sustain a growing VETNET.Source: Breadth & Fragmentation · p. 7, Resource & Capacity Constraints · p. 7

Ground Truth and Labeling

• Sparse ground-truth documentation makes labels designed artifacts requiring explicit protocols; trustworthy use also calls for transparency, uncertainty communication, and attention to access barriers.The paper identifies under-trust, over-trust, automation bias, and deskilling as risks, and recommends lightweight baselines and resource-aware participation.Source: Ground Truth and Labeling · p. 8, Trust, Adoption, Bias, & AI Literacy · p. 8, Equity, Global Access, & Affordability · p. 8

1.3.3 | Human-Medicine Failure-Mode Lessons to Build Into VETNET

• Benchmark performance is not clinical readiness: evaluation should probe shortcuts, external and temporal validity, subgroup performance, calibration, and leakage, while deployment requires professional oversight.Veterinary records may contain client, farm, financial, and proprietary information; deidentification alone is insufficient to address reidentification and misuse concerns.Source: Legal Concerns: Privacy, Ownership,… · pp. 8–9, §1.3.3 · p. 9

2.1 | Vision

• VETNET’s vision is a trustworthy ecosystem supporting responsible AI applications and ethical innovation across animal health, well-being, and veterinary medicine.Its implementation pairs a quality-assured benchmark dataset repository with a platform for equitable, secure access and collaborative model development.Source: §2.1 · pp. 9–10
• The first year should deliver one complete benchmark rather than accumulating loosely specified datasets, with mutual trust and support underpinning data security and transparency.The guiding principle “any animal, any disease, any responsible task” is to be translated into staged, tractable benchmarks.Source: §2.1 · p. 10

2.2 | Mission

• VETNET’s mission is an accessible, well-governed platform with veterinary-specific datasets, documentation, metadata, metrics, holdout sets, and collaborative infrastructure.Standardized dataset requirements, governance, multidisciplinary collaboration, AI literacy, equitable participation, and transparent risk communication are part of that mission.Source: §2.2 · p. 10

3.1.2 | Standardization

• Standardization covers metadata, terminology, deidentification, documentation, and access, while integration spans imaging, health records, time series, and emerging data streams.The common data model is intended to harmonize uploaded data and documentation.Source: §3.1.2 · p. 11, §3.1.3 · p. 11

3.1.4 | Model Evaluation & Validation

• Evaluation plans pair industry-standard metrics with subgroup and out-of-time assessment, protected test subsets, and real-world clinical validation.Proposed evaluation subsets include breed, species, geography, and institution.Source: §3.1.4 · pp. 11–12

3.2.1 | Organizational Structure

• VETNET’s proposed organizational and governance work spans stakeholder representation, animal welfare, contributor protections, ethical review, and policies for consent, privacy, and liability.The organizational plan includes working groups, documentation, and interest groups; regulatory bodies and legal experts are also proposed for engagement.Source: §3.2.1 · p. 12, Governance Working Group Statement · p. 13, §3.2.2 · p. 15

Minimum Viable Product (MVP) Working Group Statement

• The MVP effort aims to deploy a small-scale benchmark, establish curation procedures, and expose knowledge and resource gaps for future efforts.Its components include task definition, governance, data retrieval, annotation, deidentification, common data formats, and secure exchange.Source: Minimum Viable Product (MVP) Working… · p. 13, §3.3.1 · p. 16

Stakeholders and Funding Working Group Statement

• The funding group proposes diversified support, institutional and industry partnerships, contributor incentives, and transparent commercial participation terms.Its stated two-year targets include 5–7 institutional data partnerships and 10–15 industry consortium members.Source: Stakeholders and Funding Working Group… · p. 14

Repositories Working Group Statement

• VETNET’s repository plan combines user-facing infrastructure with secure, consistent dataset integration and scalable data architecture.It proposes a sandbox, web interface, tiered access, living benchmarks, and contributor guidance through data ingestion and licensing.Source: §3.1.1 · p. 10, Repositories Working Group Statement · pp. 14–15

3.3.2 | Candidates for Pilot Projects

• Pilot selection should reflect community priorities and weigh stakeholder demand, data quality, risk-benefit, clinical or welfare value, equity, and expert-threshold feasibility.Other proposed candidates include scribes and client-facing chatbots, with concerns including record integrity and delayed care.Source: §3.3.2 · p. 17, Table 5 · p. 18
• Medical imaging is identified as a frontrunner pilot candidate, but its diagnostic stakes and potential marketing misuse call for expert-reviewed labels and narrow research framing.The candidate rationale includes clear input structure, DICOM precedent, and strong machine-learning interest.Source: §3.3.2 · p. 17, Table 5 · p. 18

4.1 | Phase 1: Foundation (Years 1-2)

• Phase 1 targets governance, founding partnerships and funding, up to three pilot datasets, and standardized data-collection, annotation, quality-control, and evaluation protocols.The first year’s overarching milestone is a complete benchmark loop, from task selection and data access through a runnable harness, documentation, feedback, and revision.Source: §4.1 · p. 18

4.2 | Phase 2: Expansion (Years 2-3)

• Phase 2 would scale to a small cross-species, cross-modal portfolio only after the MVP workflow is maintainable, while strengthening standards, governance, validation, and participation pathways.Planned work also includes certification, international data-sharing, exploration of hybrid or federated learning, reporting rules, and practitioner education.Source: §4.2 · p. 20

4.3 | Phase 3: Maturation (Years 4+)

• Phase 3 aims for self-sustaining operations, prospective studies of benchmark-validated tools, and VETNET influence on regulatory policy and global veterinary AI standards.The proposed operating model includes certification fees, licensing, and industry partnerships.Source: §4.3 · p. 20

4.4 | Summary of Roadmap

• Table 8 summarizes VETNET’s phases across its five-year roadmap.Source: Table 8 · p. 21

5 | Anticipating Success: Factors and Metrics

• Success may be assessed through deliverables such as deployed benchmarks and publications, alongside the number of animals whose care is influenced and less-quantifiable effects on animal health, welfare, and veterinary care.The paper also emphasizes low-friction access, clear documentation, early measurable impact, regular communication, responsive issue resolution, and stakeholder-specific value.Source: §5 · p. 21

5.1 | Community Engagement and Usability

• Community engagement depends on diverse stakeholder participation, transparent communication, workforce AI literacy, and equitable, user-friendly access and contribution processes.The value map links participation to practical benefits including credible comparisons, validation pathways, clinically meaningful thresholds, and clearer communication about AI utility and limitations.Source: §5.1 · pp. 21–22

5.2 | Quality, Rigor, & Maintenance

• Benchmark quality and maintenance priorities include expert-curated consensus labels, diverse representation, continuous quality monitoring, and refreshes to prevent dataset drift.Participation pathways should also avoid requiring high compute or elite institutional infrastructure.Source: §5.2 · p. 22

5.3 | Impact Measurement

• Impact measurement spans technical and operational indicators, validation beyond development sites, clinical outcomes, equity, environmental effects, and real-world safety and welfare.Proposed measures include external-site and time-based validation, subgroup performance, calibration, prospective validation, workflow fit, and real-world monitoring.Source: §5.3 · pp. 22–23, Table 10 · p. 23

• VETNET seeks institutional resources, collaborative industry participation, regulatory dialogue, and philanthropic investment to build benchmark and validation infrastructure.Source: §6 · p. 23
• Potential contributors can nominate pilot tasks, build evaluation tools, provide data or infrastructure under transparent rules, clarify legal and appropriate-use expectations, fund stewardship, and identify caregiver workflow and welfare needs.Source: §6 · pp. 23–24

• Participants are asked to comment on VETNET’s mission, vision, values, launch definition, and explicit year-one non-goals.The discussion grounding identified responsible data and AI innovation for animals, animal health and well-being, veterinary science, and One Health interconnections.Source: §7 · p. 24
• The first pilot should be feasible, useful, and trust-building; participants are asked to review a ranked shortlist using feasibility-risk-value criteria and identify an owner.Source: §7 · p. 24
• The pilot’s access tier and data-use rules remain open because open, registered, controlled, and federated access carry different implications for participation, privacy, and contributor trust.Source: §7 · p. 24
• Company participation requires clear boundaries for commercial use, benchmark claims, product evaluation, and derivative products; the requested output is a pilot policy.Source: §7 · p. 24
• Domain experts must define success thresholds before release because machine-learning metrics alone cannot establish clinical or welfare meaning.The proposed output is a completed threshold worksheet.Source: §7 · p. 24

8 | Conclusion

• A complete, useful, well-governed first benchmark loop should demonstrate that veterinary relevance and technical reproducibility can work together.The near-term task is to build the package, show stakeholders can use it, govern it responsibly, and learn from the process.Source: §8 · p. 25
• VETNET’s long-term aspiration is to support AI tools for “any animal, any disease, any responsible task.”The conclusion distinguishes this broad aspiration from the narrower proof-of-concept task.Source: §8 · p. 25
• If VETNET builds a credible foundation, it stands to become a trusted reference point for responsible veterinary AI innovation.Source: §8 · p. 25

Funding

• The AI Thought Summit was funded by Cornell AI Initiative, and author KK was funded by NIH T32ODO011000.Source: Funding · p. 25

Conflicts of Interest

• The authors declare no conflicts of interest.Source: Conflicts of Interest · p. 25

Disclosure

• The content is solely the authors’ responsibility and does not necessarily represent the official views of government agencies or private entities, including NIH.Source: Disclosure · p. 25

Appendix B | Minimum Dataset Documentation Template

• Record dataset identity, version, contributors, contacts, represented populations and settings, collection timeframe, and inclusion criteria.Source: Appendix B · p. 29
• Document raw and processed data, cleaning and deidentification steps, transformations, known data losses, label protocols, expertise, disagreement handling, and uncertainty.Source: Appendix B · p. 29
• Specify metadata and missingness, access and license terms, permitted and prohibited uses, and withdrawal or correction procedures.Source: Appendix B · p. 29
• Record limitations, anticipated biases, subgroup gaps, clinical-risk warnings, equity and access limitations, and interpretation guidance.Source: Appendix B · p. 29
• Plan external and temporal validation, subgroup reporting, artifact and shortcut probes, and drift monitoring where feasible.Source: Appendix B · p. 29

Appendix C | Tiered Access Model

• A benchmark is a defined evaluation package covering task, data or access, labels, metrics, code, documentation, governance, and interpretation rules.An evaluation harness is runnable code and file formats for submissions and standardized scoring.Source: Appendix D · p. 30
• A holdout set protects test labels from model developers to reduce overfitting; external validation uses data from different sites, practices, equipment, geographies, or institutions.Source: Appendix D · p. 30
• Temporal validation tests robustness to change over time by evaluating on data from a later period than the training data.Source: Appendix D · p. 30
• Shortcut learning means using an easy but misleading correlate instead of the intended clinical signal; distribution shift means deployment data differ from development data.The glossary notes that distribution shift often reduces performance.Source: Appendix D · p. 30
• Calibration describes how well predicted confidence matches observed correctness; automation bias and deskilling describe over-reliance on outputs and erosion of human judgment.Source: Appendix D · p. 30
• Federated evaluation keeps data with the holder while submissions are evaluated without broadly releasing raw records; data cards and Croissant document data and support dataset discoverability and interoperability.Source: Appendix D · p. 30
• Inner-loop evaluation is offline benchmarking under controlled conditions; outer-loop validation covers prospective, workflow, safety, welfare, ethical, legal, and deployment evidence before real-world use claims.Source: Appendix D · p. 30