APEHR: Automated prognosis in electronic health records using multi-head self-attention

Automated prognosis has been a topic of intense research. Many works have sought to learn from Electronic Health Records using Recurrent Neural Networks that, despite promising results, have been overcome by novel techniques. We introduce APEHR, a Transformer approach that leverages medical prognosis using the latest technology Neural Network Transformer, which has demonstrated superior results in problems whose data is organized in sequential fashion. We contribute with an innovative problem modeling along with a detailed discussion of how Transformers can be used in the medical domain. Our results demonstrate a prognostic performance that surpasses previous works by at least 6% for metric Recall@k in the public dataset MIMIC-III.