Collaborative Research Project 1

Principal investigators: Tessa Cook (University of Pennsylvania) and Curtis Langlotz (Stanford)

Natural language processing of radiology reports for COVID-19.

Updated January 20, 2023

Collaboration between University of Pennsylvania and Stanford researchers building document level classification algorithms to determine whether narrative radiology reports indicate the presence of COVID-19, other viral pneumonia or other lung disease.  This research team has also been developing an information extraction tool to identify specific associated imaging findings and other clinical conditions, as well as a high performing open-source de-identification algorithm for clinical radiology reports.

Current plans include

  • Build a RadGraph model for modalities other than XR and body regions other than chest.  Assess the performance and generalizability for extraction of entities and relations.

  • Conduct a head-to-head comparison of the radiology report de-identifier with open-source algorithms and algorithms commercially available from cloud vendors.

Effect of fine-tuning strategies on validation loss in deep learning natural language processing as a function of the learning rate

Effect of fine-tuning strategies on validation loss in deep learning natural language processing as a function of the learning rate

Next

CRP 2: Algorithmic approaches for improving health equity