Explainable Classification of Medical Documents Through a Text-to-Text Transformer

Death certificates are important medical records which are collected for the purpose of public healthcare and statistics by multiple organizations around the globe. Due to their importance, those certificates are compiled by experienced medical practitioner according to a standard defined by the World Health Organization including rules to select an underlying cause of death (UCOD). For this reason, the coding of death certificates is a slow and costly process. To overcome these issues, the scientific community proposed deep learning approaches to perform such a task. Despite those systems achieve high accuracy scores (close to 1), their complexity makes the obscure to the final user, making it unfeasible the adoption as a decision support system. In this paper, we propose a model based on text-to-text transformers which is able to provide a UCOD as well as to generate a human-readable explanation for its classification. We compare the proposed approach to state-of-the-art interpretable rule-based systems.