A FrameNet for Cancer Information in Clinical Narratives: Schema and Annotation

A FrameNet for Cancer Information in Clinical Narratives: Schema and Annotation Kirk Roberts1, Yuqi Si1, Anshul Gandhi1, Elmer V. Bernstam1,2 1 School of Biomedical Informatics 2 Division of General Internal Medicine, Department of Internal Medicine, McGovern Medical School The University of Texas Health Science Center at Houston Houston, TX USA kirk.roberts@uth.tmc.edu Abstract This paper presents a pilot project named Cancer FrameNet. The project’s goal is a general-purpose natural language processing (NLP) resource for cancer-related information in clinical notes (i.e., patient records in an electronic health record system). While previous cancer NLP annotation projects have largely been ad hoc resources to address a specific and immediate information need, the frame semantic method employed here emphasizes the information presented in the notes themselves and its linguistic structure. To this end, three semantic frames (targeting the high-level tasks of cancer diagnoses, cancer therapeutic procedures, and tumor descriptions) are created and annotated on a clinical text corpus. Prior to annotation, candidate sentences are extracted from a clinical data warehouse and de-identified to remove any private information. The frames are then annotated with the three frames totaling over thirty frame elements. This paper describes these steps in the pilot project and discusses issues encountered to evaluate the feasibility of general- purpose linguistic resources for extracting cancer-related information. Keywords: clinical information extraction, cancer, frame semantics 1 Introduction Important medical information about cancer patients is of- ten only available in free text (natural language) notes in electronic health records (EHRs). This information is fre- quently needed for research, quality improvement, surveil- lance, and other important functions. However, the man- ual abstraction of this information can be incredibly time- consuming and expensive, often making it infeasible for both early-stage research and clinical quality improvement projects. Thus, natural language processing (NLP) ap- proaches can offer a tremendous service in oncology. Existing NLP systems for cancer-related information gen- erally fall under a type of biomedical NLP task known as phenotyping, which is the task of identifying patients that meet a certain set of criteria. Phenotyping specifications are often highly task-specific, which frequently yields one-off NLP datasets and algorithms that do not generalize to simi- lar phenotyping tasks. For example, one phenotype method may identify lung cancer patients with a tumor of at least 2cm in diameter, while another method may identify lung cancer patients with at least two tumors that are 1cm in di- ameter. Such methods are quite similar, yet often produce incompatible annotations and algorithms. The key insight is that phenotyping methods often conflate extraction (identifying relevant portions of text) and rea- soning (determining if the extracted text fulfills the task’s needs). By separating these steps and focusing on general- purpose extraction, it will be possible to easily and rapidly develop phenotyping methods. This is because the bulk of the effort is typically spent in the extraction step (annotating data and developing NLP algorithms), while the reasoning step is often a straightforward set of rules. Continuing the example above, an NLP system capable of extracting all tu- mor …

Content truncated to 500 words