Practical Problems and Solutions in Hospital Information System Data Mining

Abstract

Information mining from textual data becomes a very challenging task when the structure of the text record is very loose without any rules. Doctors often use natural language in medical records. Therefore it contains many ambiguities due to non-standard abbreviations and synonyms. The medical environment itself is also very specific: the natural language used in textual description varies with the personality creating the record (there are many personalized approaches), however it is restricted by terminology (i.e. medical terms, medical standards, etc.). Moreover, the typical patient record is filled with typographical errors, duplicates, ambiguities, syntax errors and many nonstandard abbreviations.

This paper describes the process of mining information from loosely structured medical textual records with no apriori knowledge. The paper concerns mining a large dataset of ~50,000–140,000 records × 20 attributes in relational database tables, originating from the hospital information system (thanks go to the University Hospital in Brno, Czech Republic) recording over 11 years. This paper concerns only textual attributes with free text input, that means 650,000 text fields in 16 attributes. Each attribute item contains approximately 800–1,500 characters (diagnoses, medications, anamneses, etc.). The output of this task is a set of ordered/nominal attributes suitable for automated processing that can help in asphyxia prediction during delivery.

The proposed technique has an important impact on reduction of the processing time of loosely structured textual records for experts.

Note that this project is an ongoing process (and research) and new data are still received from the medical facility, justifying the need for robust and fool-proof algorithms.

In the preliminary analysis of the data, classical approaches such as basic statistic measures, word (and word sequence) frequency analysis, etc., have been used to simplify the textual data and provide a preliminary overview of the data. Finally, an ant-inspired self-organizing approach has been used to automatically provide a simplified dominant structure, presenting structure of the records in the human readable form that can be further utilized in the mining process as it describes the vast majority of the records.