MedAttention: A Self-Attentive RNN to Predict Diabetes Complications with Financial Data

Rafael T. Sousa, Lucas A. Pereira, Arlindo R. Galvão Filho, Anderson da S. Soares

Feb 12, 2018 (modified: Feb 12, 2018) ICLR 2018 Workshop Submission readers: everyone
  • Abstract: This work is aimed to introduce MedAttention, a self-attentive recurrent neural network capable of predicting diabetes complications over health plan operators financial data. We only use financial records due to the unavailability of electronic medical records and exam results in several Brazilian health care institutions. The financial records used encodes medical procedures, hospital rates, materials and medicines. Our results succeed to predict complications over 60 to 360 days after the records processing with an area under ROC curve of 0.81 to 0.76 over time gaps. Also we introduce the possibility to visualize the attention mechanism of MedAttention to understand the patterns found.
  • TL;DR: A Self-Attentive RNN model to predict diabetes complications using only financial data
  • Keywords: RNN, Self-Attention, Diabetes, EHR