Automatic Classification of Clinical Cases Using Deep Learning Techniques

The automatic assignation of disease codes is a complex problem that has been addressed many times throughout decades. In particular, the categorization of ICD (International Classification of Diseases) codes, which it's a compendium of symptoms, diseases, procedures and injuries. This activity is done by manually analyzing clinical cases or discharge summaries and its use has spread to areas like billing, administration or refund. Leading to associated costs close to $417 billion dollars for United States on 2012. Therefore in this investigation we propose Deep Learning models aiming to help in the task of code assignment. For this, 6 models are proposed, including architectures of Convulutional and Recurrent Neuronal Networks; both focused on NLP (Natural Language Processing) extracting features through aWord Embeddings approach. The results were obtained from the top 10, 20, 50 and 100 most frequent diseases; getting an Average Precision of 79,86% for the top 10 with an AUC of 91,37% which outperforms other methods used previously in this task.