Information extraction from clinical and radiology notes for liver cancer staging

Thesis (Ph.D.)--University of Washington, 2016-06

Medical practice involves an astonishing amount of variation across individual clinicians, departments, and institutions. Adding to this condition, with the exponential pace of new discoveries in biomedical research, medical professionals, often understaffed and overworked, have little time and resources to analyze or incorporate the latest research into clinical practice. The accelerated adoption of electronic medical records (EMRs) brings about great opportunities to mitigate these issues. In computable form, large volumes of medical information can now be stored and queried, so that optimization of treatments based on patient characteristics, institutional resources, and patient preferences may be data driven. Thus, instead of relying on the skillsets of patients' support network and medical teams, patient outcomes can at least have some statistical guarantees. In this dissertation, we focused specifically on the task of hepatocellular carcinoma (HCC) liver cancer staging using natural language processing (NLP) techniques. Staging, or categorizing cancer patients by extent of diseases, is important for normalizing over patient characteristics. Normalized stages, can then be used to facilitate research in evidence-based medicine to optimize for treatments and outcomes. NLP is necessary, as with other clinical tasks, a majority of staging information is trapped in free text clinical data. This thesis proposes an approach to liver cancer stage phenotype classification using a mixture of rule-based and machine learning techniques for text extraction. Included in this approach is a careful, layered design for annotation and classification. Each constituent part of our system was characterized by detailed quantitative and qualitative analysis. Two important modules in this thesis are a framework for normalizing text evidence related to specific conditions and an algorithm for tumor reference resolution. The overall results of our system revealed an F1 performance of 0.55, 0.50, 0.43 for AJCC, BCLC, and CLIP liver cancer stages, respectively. Although outperforming baseline classifications, these accuracies are not viable for clinical use. Analysis of error suggests that performance for some constituent stage parameters would improve through additional annotation. However, one identified crippling bottleneck was the requirement of reference resolution and discourse-level reasoning to determine the number of tumors in a patient, a crucial part of cancer staging. Still our work provides a methodology to classify a complex phenotype, whose strength includes its interpretability and modularity while maintaining ability to scale and improve with greater amounts of data. Furthermore, submodules of our system, for which perform at higher accuracies, may be used as tools to decrease annotation costs.