Grant Number:	1R01CA323939-01 Interpret this number
Primary Investigator:	Zhou, Xuanyu
Organization:	Arizona State University-Tempe Campus
Project Title:	SCH: Proof-Certified Cancer Registry Extraction
Fiscal Year:	2026

PROJECT SUMMARY (See instructions): Cancer registries are essential infrastructure for cancer surveillance, oncology research, and clinical quality improvement. Certified Tumor Registrars synthesize pathology reports, imaging summaries, operative notes, oncology clinic notes, and other clinical narratives into standardized fields that support programs such as SEER. The growing volume and complexity of oncology documentation make manual abstraction increasingly difficult, while conventional AI systems still provide predictions rather than registry-grade evidence that can be checked, audited, and governed. This project develops proof-certified cancer registry extraction: a neuro-symbolic architecture that treats each extracted registry field as a formal claim that must be verified before it can be advanced for automation. The project will build a typed domain-specific language and clinical proof kernel for registry claims; construct lemma libraries from authoritative registry guidelines and high-precision natural language processing primitives; train proof-carrying information extraction models that propose values, evidence, and proof sketches from oncology EHR text; and convert verified proofs into risk certificates, evidence cards, and registrar-in-the-loop workflows. Claims whose proofs verify under explicit risk budgets can be advanced, while uncertain or conflicting cases are escalated with localized explanations and evidence. The expected outcome is a validated framework for selectively automating cancer registry abstraction while preserving transparency, reproducibility, and human oversight. The work directly supports the NCI mission by improving the timeliness, reliability, and auditability of cancer surveillance data that underlie population level cancer research and quality improvement.





None