Text Mining in Radiological Data Records: An Unsupervised Neural Network Approach

Abstract

The rapid growth in digitalized medical records presents new opportunities for coalescing terra bytes of data into information that could provide us with new knowledge. The knowledge discovered as such could assist medical practitioners in a myriad of ways, for example in selecting the optimal diagnostic tool from among many possible choices. We analyzed the radiology department records of children who had undergone a CT scanning procedure at Nagasaki University Hospital in the year 2004. We employed self organizing maps (SOM), an unsupervised neural network based text-mining technique for the analysis. This approach led to the identification of keywords within the narratives accompanying the medical records that could contribute to reduction of unnecessary CT requests by clinicians. This is important because overuse of medical radiation poses significant health risks to children in spite of the invaluable diagnostic capacity of such procedures