Casematch goes Deep Learning
With effect from 2017 and the implementation of SwissDRG 6.0 and retroactive for the versions 4.0 and 5.0, many innovations will occur for our coding control software Casematch. The emphasis of the development is the complete replacement of the statistical backend from a Java based solution to a Python based. With the used frameworks Keras and Tensorflow, that are developed and used by Google, we rely consequently on Deep Learning. Deep Learning is a new approach of machine learning that set new standards in multiple fields over the last five years. This includes among others object recognition in images or videos, bioinformatics, speech recognition or text analysis.
We analyze the medical coding with similar tools as we use for text analysis of natural language, even though the medical coding is more structured than natural language and consists of a standard vocabulary (ICD and CHOP catalogs). For instance we use recurrent neural networks that can handle different lengths of sequences (different number of secondary diagnosis and procedures) and can display the order of the coding. Further, we use with word embeddings a method to map ICD codes in a vector space so that similar codes lie close together. To create this proximity between similar codes, we use the following information: Codes were often coded together, codes are assigned to the same chapter, group or non-terminal in the ICD catalog or codes that have linguistically similar description texts.
Thus, more semantic can be displayed in the statistical model. More precise proposals and selections of noticeable cases are the result. Additionally, this is a first step towards the analyses of free text based primary documentation for the support of the coding. However, this is not only a lot of work ahead, at the moment we lack particularly the data for the validation.
Another development is the integration of our search API (https://search.eonum.ch) that now allows the search of inclusiva and synonyms.
The actualizations happen automatically and in the background. Users do not need to make any adjustments.