Emilia Apostolova1, Joe Morales2, Ioannis Koutroulis3 and Tom Velez2, 1Language.ai, USA, 2Computer Technology Associates, USA, 3Children's National Health System, USA
While there has been considerable progress in building deep learning models based on clinical time series data, overall machine learning (ML) performance remains modest. Typical ML applications struggle to combine various heterogenous sources of Electronic Medical Record (EMR) data, often recorded as a combination of free-text clinical notes and structured EMR data. The goal of this work is to develop an approach for combining such heterogenous EMR sources for time-series based patient outcome predictions. We developed a deep learning framework capable of representing free-text clinical notes in a low dimensional vector space, semantically representing the overall patient medical condition. The free-text based time-series vectors were combined with time-series of vital signs and lab results and used to predict patients at risk of developing a complex and deadly condition: acute respiratory distress syndrome. Results utilizing early data show significant performance improvement and validate the utility of the approach.
Natural Language Processing, Clinical NLP, Time-series data, Machine Learning, Deep Learning, Free-text and structured data, Clinical Decision Support, ARDS, COVID-19