Bridging the unstructured and structured worlds: An adaptive self-learning medical form generating system

The prevalence of medical report standards and structured reporting systems reflects the increasing demand for representing and preserving medical and clinical data with controlled vocabularies in well structured format. Strictly formatted medical reports offer high human readability and facilitate further data processing, such as querying, statistical analysis, and reasoning to support decision making. However, many medical reports, such as pathology reports, nursing notes and physician's notes, are written in free-text narration. Manually extracting free text reports by filling predefined data fields is cumbersome and error-prone. Meanwhile, information extraction tools try to automate such process, for example, through machine learning based methods. Such methods often require large volumes of training datasets annotated manually by humans, which is expensive to obtain. Furthermore, they are also limited by their accuracy (both precision and recall). To facilitate the process of extracting information from narrative medical reports and transforming extracted data into standardized structured forms, we present in this paper a semiautomatic system, ASLForm, that interacts with users, analyzes free text input and generates normalized answers to populate forms in real-time. This system learns from users' feedback transparently and establishes decision models incrementally. It requires no additional configurations and training datasets. ASLForm is not constrained to any domain, and is adaptable to free text input in any format. These features of the system offer high usability and portability. Its design also enables easy integration with existing reporting systems.