Determination of Disease from Discharge Summaries

Abstract

Determining whether correct disease codes are included in discharge summaries is important for hospital management because submission of medical receipts with incorrect disease codes can result in loss of insurance reimbursement. Because medical information managers in large hospitals must evaluate more than 1000 summaries per month, an automated determination of discharge summaries will reduce their workload, allowing information managers to focus on complicated cases. This paper proposes a method of constructing classifiers of discharge summaries. In the first step, morphological analysis generated a term matrix from text data extracted from the hospital information system. Subsequently, important keywords were selected from an analysis of correspondence, training examples were generated, and machine learning methods were applied to the training examples. Several machine learning methods were compared using discharge summaries stored in the information system of Shimane University Hospital. A random forest method was found to be the best classifier when compared with deep learning, SVM and decision tree methods. Furthermore, the random forest method had a classification accuracy greater than 90%.