Retrieving cases for treatment advice in nursing using text representation and structured text retrieval

A nursing database which records patient details and treatments as fields in a standard database format is transformed into a collection, in text form, of patient case days with history. Each case is represented as text strings encoding the patient details, the current problems, treatments and their associated history. The cosine measure of similarity is used to compute a whole case similarity between a text query and the cases in text form. This standard text retrieval technique is used and compared to a simple rule base. In case-based reasoning, the similarity of cases is often computed by combining similarities of the case features involved. In this work the standard text retrieval function is modified to incorporate this case structure by combining individual matches of case components based on the cosine measure. The combination is based on a linear regression model for learning the weights assigned to the components of this retrieval function. For the 1355 records two tasks were tried: predicting the treatment for a new problem and predicting the treatment for a continuing problem when a change of treatment is required. Simple text retrieval was better than the rule base for one task and case structured retrieval was at least 18% better on both tasks. Further techniques are discussed.