Dilated Convolution Networks for Classification of ICD-9 based Clinical Summaries

Deployment of Artificial Intelligence for understanding clinical notes in the healthcare sector is a crucial step to extract meaningful phrases based on diseases. Electronic Health Records (EHR) are stored in the health care system in an unstructured and event associated way. Public clinical records can be used for billing, monitoring and insurance purpose. These clinical notes contain abbreviations, acronyms, and a non-uniform dictionary. Various Machine learning models are used with different approaches to understand these notes, these models are evaluated in various criteria based on datasets. These techniques differed mainly in pre-processing and code assignments as well as architecture for reading long medical documents. In this thesis work, we propose a layered model of Convolution Neural Networks with pre-trained embeddings. This architecture uses multiple dilation layers with a label-specific dot-based attention mechanism. We have extracted the embeddings from Common Crawl Glove (Global Vector). The architecture of the model is designed to calculate attention to words and their context. The model is trained on MIMIC-III data set labelled with the ICD-9-CM hierarchy. This research is helpful in highlighting and perceiving useful information from medical reports to a physician, this step will apparently increase treatment quality and support administrative tasks. We conclude with optimal results compared to state-of-the-art models, proposing certain limitations and possible developments in the near future. It will address a significant role in health security for nations using their public healthcare systems.