Structured Clinicopathologic Data as a Prerequisite for Generalizable AI in Ocular Oncology

Artificial intelligence (AI) has shown promise in ocular oncology in several narrow but clinically relevant tasks. Existing studies suggest that machine-learning models can assist in distinguishing small choroidal melanoma from choroidal nevus and that deep-learning approaches can detect ocular surface squamous neoplasia from slit-lamp images, highlighting potential roles in diagnosis, triage, and teleconsultation [1, 2]. Reviews of this literature are encouraging, but they also converge on the same limitations: narrowly defined image-based tasks, small retrospective datasets, single-center development, heterogeneous reference standards and label definitions, limited external validation, and uncertainty about real-world deployment [3]. Much less attention has been paid to the data infrastructure required to generate consistent diagnostic, prognostic, and outcome labels from pathology, molecular testing, and longitudinal follow-up. This perspective argues that structured, auditable, and interoperable clinicopathologic variables are a necessary precondition for generalizable, evaluable, and safely deployable AI in ocular oncology. This argument concerns data infrastructure, not a claim that structured data alone resolves broader questions of validity, safety, transparency, privacy, or fairness in clinical use [4].Ocular oncology decisions are inherently multimodal. A single case may include imaging, cytology/histopathology, immunohistochemistry, molecular testing, and longitudinal outcomes. The reference standard for that case is often established through multidisciplinary discussion and documented across narrative reports and free-text clinic notes, which are frequently dispersed across systems and over time.This is why ocular oncology is both promising and challenging for AI. While imaging-based models can be trained, the target variables they depend on (e.g., diagnosis, stage, histologic risk factors, molecular class, treatment intent, and outcome definitions) are often not captured as discrete fields, are defined differently across institutions, and are difficult to audit or harmonize. Models trained in one workflow may not generalize to another, even when the images look similar, because they may learn local documentation patterns rather than durable clinicopathologic signals.If the immediate bottleneck is the portability and auditability of clinicopathologic data, pathology is a logical place to start. By computable ocular oncology, we mean the routine capture of key clinicopathologic variables in structured, queryable, versioned, and interoperable forms that can be audited within and across institutions.Ocular oncology already has a strong foundation for standardization in pathology content. The College of American Pathologists (CAP) uveal melanoma protocol and the Royal College of Pathologists (RCPath) dataset specify discrete, clinically meaningful elements for staging and prognostication and promote structured reporting [5, 6]. Recent pathology informatics work has shown that published synoptic cancer datasets can be rendered semantically and syntactically interoperable using Systematized Nomenclature of Medicine – Clinical Terms (SNOMED CT) and Fast Healthcare Interoperability Resources (FHIR), underscoring that “computable” reporting is a practical informatics goal rather than an abstract ideal [7].The opportunity now is to treat these synoptic elements as more than documentation standards. They can become the shared data language that makes multicenter AI, registries, and real-world evidence feasible across institutions.A key distinction matters: “synoptic on paper” versus “synoptic as data.” A report that lists required elements but is stored as formatted text remains human-readable, but it is not readily reusable as structured data. If we want AI that can be trained, validated, and compared across sites and over time, core clinicopathologic variables of ocular tumors should be captured as discrete, versioned fields (e.g., protocol version; AJCC edition) so that these data elements and their definitions are temporally and semantically stable.This is timely because Cancer PathCHART, the US National Cancer Institute’s initiative updating site-morphology validation standards for cancer surveillance, lists Eye and Orbit among the pending sites for 2027 implementation-year review [8]. That creates an opportunity for ocular oncology to influence terminology and coding that will affect registry abstraction, interoperability, and the harmonization of multicenter AI datasets.If data structure is the immediate bottleneck, the first AI applications worth deploying are not autonomous classifiers but documentation-first tools that improve the fidelity of routine care. Three practical applications stand out:1.Automated completeness and consistency checks for pathology reporting. Tools can flag missing required CAP/RCPath elements prior to sign-out (e.g., absent staging descriptors, incomplete margin/extension documentation, missing required special studies fields), reducing omissions in prognostic documentation and improving downstream data quality.2.Structured extraction from legacy narrative reports (under human sign-off). Many centers have years of high-quality ocular oncology reports that are not readily queryable. Governed natural language processing, potentially including large language models used only within secure environments, could convert free text into structured variables aligned to CAP/RCPath standards, with pathologist verification, unlocking existing cases without changing clinical practice.3.Treat the tumor board as a computable event. Multidisciplinary discussion often converts heterogeneous inputs into a single plan (e.g., surveillance intensity, metastatic risk group, biopsy recommendation). Capturing tumor board outputs in a small structured set would create a practical bridge among pathology, imaging, and outcomes, and would support audit and prospective evaluation of decision support.The rarity of ocular malignancies makes multicenter collaboration essential, but data-sharing friction is real. Federated learning and related privacy-preserving approaches can help, but they are only as good as the labels and endpoint definitions they learn from. If each site encodes key endpoints differently, federated training simply scales inconsistency. The first step is not federation but standardized, auditable clinicopathologic variables and endpoint definitions.An implementation-ready agenda could look like this:Define a minimum computable dataset for priority conditions (starting with uveal melanoma and ocular surface squamous neoplasia): staging, key pathology risk factors, key ancillary studies, treatment category, and a small number of outcomes with explicit definitions.Make synoptic reporting truly computable: capture core elements as discrete laboratory information system/electronic health record fields (not only formatted text) and include explicit versioning (CAP protocol version; AJCC edition) so model inputs remain stable over time [5].Start with documentation-first tools (report completeness checks, legacy extraction, tumor board structured outputs) before autonomous diagnosis. These are lower risk, immediately useful and generate higher quality data as the byproduct.Evaluate success at two levels: data quality (completeness, concordance, portability) and AI performance/governance (external validation, calibration, workflow safety, and fairness), rather than assume that better labels alone are sufficient for robust evaluation or clinical deployment.Reward data rigor in AI publications: encourage explicit label definitions, missingness handling, and a brief variable dictionary mapping study variables to CAP/RCPath elements when applicable. Contemporary reporting guidance (e.g., TRIPOD+AI for prediction models) can support this without overburdening authors [9].In conclusion, the claim of this perspective is not that structured data alone solve the broader challenges of clinical AI evaluation and deployment. Rather, structured, auditable, and interoperable clinicopathologic variables are a prerequisite infrastructure for external validation, multicenter comparison, error analysis, and governance. Ocular oncology is unusually well positioned to lead this work because it already integrates clinical, imaging, pathology, and outcomes, and it already possesses mature reporting standards. If the field invests first in computable reporting and structured multidisciplinary outputs, it will be better positioned to develop AI systems that are not merely accurate in one dataset but generalizable and safely deployable across rare-disease settings.The authors have no conflicts of interest to declare.This work received no specific funding. No external funders had any role in the conception, planning, writing, or decision to publish this manuscript.Conceptualization and writing – original draft: H.B. Writing – review and editing: H.B. and M.D.