A DPC code expresses a primary disease, a complication, and procedures, etc. In 2010, 1334 hospitals use DPC codes for calculations of medical fees. Since, in the hospitals, the medical fee of each case is calculated based on one DPC code, each case must be classified into one DPC code. However, the classification is difficult in some cases because patients sometimes have various conditions. Therefore, automatic DPC code selections using machine learning are being studied. Suzuki et al. evaluated automatic DPC code selections from discharge summaries using a vector space method. However, there are general machine learning methods except for the vector space method. Hence, we must evaluate other machine learning methods exhaustively for improvement of accuracy of automatic DPC code selections. Therefore, we evaluated automatic DPC code selections from discharge summaries using naïve Bayes method, SVM, concept base method, and another vector space method which is different from the vector space model used by Suzuki et al. We considered these machine learning methods as general ones. We also focus on characteristics of each machine learning methods on automatic DPC code selections and we utilize a method which combines some machine learning methods. First, the combining method estimates confidences of the machine learning methods bases on classification scores that the machine learning methods regard as classification evidence. Next, the combining method adopts the method whose confidence is highest. We compared accuracy of the methods using discharge summaries created in 2008 fiscal year in Kyoto University Hospital. As a result, SVM classified 72.2% of the cases into correct DPC codes though the vector space model utilized by Suzuki et al. classified 64.8% into correct DPC codes. Moreover the combining method classified 76.1% into correct DPC codes. In conclusion, we achieved significant improvement.
