A fast unsupervised assignment of ICD codes with clinical notes through explanations.

In healthcare industry a set of ICD codes are assigned to a clinical note (which can be patient visit, a discharge summary and the like) as part of medical coding process which is mandatory for medical care and patient billing. Most of this assignment task use supervised framework for which a collection of the clinical notes are a-priori labeled with ICD codes. But in lot of cases there may not be enough labeled texts, and transfer learning is not a viable option. These necessitate an unsupervised assignment of codes. In this work we propose an unsupervised method of assignment of codes to clinical notes which takes diagnosis description (DD) of a patient and assigns the most appropriate ICD codes to it. This framework employs a novel approach of finding similarity between a clinical note and a list of ICD codes, called Word Mover's Similarity (WMS), which is based on the concept of Word Mover's Distance (WMD) used to compute the distance between text documents. We also speed up our proposed algorithm by developing an approach to derive similarities with the provision of a faster computation. Moreover, using this framework one is able to explain the rationale behind the assignment codes to a DD. Lastly, we show the efficacy of our unsupervised approach on clinical notes containing 59k patient visits.