Grant Number:	1R01CA316685-01 Interpret this number
Primary Investigator:	Ellingson, Sally
Organization:	University Of Kentucky
Project Title:	Ai-Enhanced Methods for Cancer Surveillance: Improving Automated Information Extraction with Dac for Clinically Actionable Cancer Types
Fiscal Year:	2026

Project Summary/Abstract A model developed from an extensive research and development (R&D) collaboration between the US National Cancer Institute (NCI) and United States Department of Energy (DOE) is currently deployed by all the Surveillance, Epidemiology, and End Results (SEER) cancer registries for automated information extraction and classification of cancer pathology reports. This model is able to achieve sufficient (97% or higher) accuracy to replace human annotators, producing considerable time savings. However, in order to achieve that accuracy, it abstains on (or declines to classify) a large fraction of reports (∼80%), due to reasons including missing information at report level, ambiguous information, the hierarchical nature of histology classifications, confounding factors as a result of the constantly evolving understanding of cancer and terminologies used in cancer research, or combination of these reasons. While adopting more sophisticated machine learning (ML) models can improve information extraction, they typically do not address issues related to evolving problem definitions or missing data. This project aims to improve the model abstention by improving the training protocols, emphasizing the problem formulation and custom loss functions that can alleviate these problems. This project seeks to better-inform the registries of the potential causes and mitigations for “errors” or mismatches between human and model annotations, allowing informed improvements in the registry operations, ultimately leading to increased efficiency and reduced workload for human annotators.





None