Sovereign Large Language Models for Structured Data Extraction from Pathology Reports: A Perspective for the Clinical Laboratory

The surgical pathology report remains one of the richest yet least computable artefacts in the clinical record. Diagnostic, prognostic, and treatment-relevant information is recorded predominantly as a free-text narrative that resists aggregation for research, quality monitoring, and cancer registration, while manual abstraction is slow, costly, and difficult to scale. Large language models (LLMs) have rapidly emerged as a means of converting unstructured pathology narrative into structured, analysis-ready data. This perspective examines the current state of the evidence, with particular reference to breast pathology, and foregrounds the distinction between proprietary cloud-hosted models and locally deployed open-weight models. Recent comparative studies indicate that open-weight models can approach the accuracy of proprietary systems for structured extraction, offering a privacy-preserving and cost-controlled alternative that keeps protected health information inside the institutional firewall—a decisive advantage under data-protection regimes such as Singapore’s Personal Data Protection Act (PDPA) and Human Biomedical Research Act (HBRA). We argue that hybrid architectures—pairing deterministic rule-based extraction for unambiguous fields with local LLMs for narrative reasoning—currently offer the most defensible route to laboratory deployment. We also highlight the “reality gap” between synthetic benchmark performance and real-world clinical accuracy, and the need to align studies with emerging reporting and appraisal frameworks (TRIPOD-LLM, PROBAST + AI). Structured extraction is compatible with the quality and traceability expectations of accredited laboratories only when it is verified before use, monitored over time, and kept under human oversight.