Abstract:
Optimal solutions for abstractive summarization of Electronic Health Record (EHR) content have yet to be discovered. An automated clinical note summary would assist with issues of physician burnout and the challenges of “note bloat” in EHRs . For our project, we will explore transfer learning techniques applied to deep learning based models for abstractive text summarization, namely BERT and BART. We will construct a text summarization model through the MIMIC-III and MIMIC-CXR data sets with discharge summaries and radiology reports. We anticipate fine tuning the model will be challenging given the complex nature of summarizing clinical content. Our evaluation of the model will be performed by both a ROUGE score and a convenience sample of 20 notes for two clinicians to compare the differences of the human-generated reports to the computer generated ones. Ultimately, this automated tool for summarizing EHR content has the potential of supplementing the physician’s discharge and radiology workflows and improving the overall patient care experience.
Vince Hartman - vch6@cornell.edu
Sai Patnala - sgp78@cornell.edu
Claus-Philipp Feuring - cf483@cornell.edu
Yichen Shao - ys2233@cornell.edu
Colin Gladue - cpg39@cornell.edu
Cyrus Tam - lt469@cornell.edu