Information Extraction of Multiple Categories from Pathology Reports

Information Extraction of Multiple Categories from Pathology Reports Yue Li NICTA VRL Melbourne, Australia y.li30@ugrad.unimelb.edu.au David Martinez NICTA VRL Melbourne, Australia davidm@csse.unimelb.edu.au Abstract Pathology reports are used to store infor- mation about cells and tissues of a patient, and they are crucial to monitor the health of individuals and population groups. In this work we present an evaluation of su- pervised text classification models for the prediction of relevant categories in pathol- ogy reports. Our aim is to integrate au- tomatic classifiers to improve the current workflow of medical experts, and we im- plement and evaluate different machine learning approaches for a large number of categories. Our results show that we are able to predict nominal categories with high average f-score (81.3%), and we can improve over the majority class baseline by relying on Naive Bayes and feature se- lection. We also find that the classification of numeric categories is harder, and deeper analysis would be required to predict these labels. 1 Introduction A pathology report is the summary of the analy- sis of cells and tissues under a microscope, and it may also contain information of the studied spec- imen as it looks to the naked eye. Pathology re- ports play an important role in cancer diagnosis and staging (describing the extent of cancer within the body, especially whether it has spread). These reports are usually written by the pathologist in natural language, and then the relevant parts are transcribed into structured form by a different per- son to be stored in a database. The use of structured information can help share the data between institutions, and can also be used to find patterns in the data. For this reason, some recent initiatives are exploring better ways to man- age pathology reports. For instance, the Depart- ment of Health and Ageing of Australia is fund- ing the project Structured Pathology Reporting of Cancer since 2008 to develop standard reporting protocols for cancer reports1. Another way to promote the creation of structured data is to use standard terminologies, such as SNOMED CT2, which is a large collection of medical terminology covering most areas of clinical information such as diseases, findings, procedures, microorganisms, pharmaceuticals etc. The National E-Health Tran- sition Authority (NEHTA) has recently launched an adapted terminology (SNOMED CT-AU) to be used by the Australian health sector3. These initiatives will help to increase the reposi- tories of structured data, but they will not be a sub- stitute to the flexibility of natural language. The relevant fields in structured reports change over time as different clinical tests are made available, and it is difficult to design a specific form to cover all the possible cases that will be observed in the pathology analysis. Clinicians need time to learn the different standards, and they prefer the flexibil- ity of free text to record their analyses and conclu- sions. Ideally their natural language input would be used to automatically extract the structured data that different protocols demand. This scenario is promising for text …

Content truncated to 500 words