OBJECTIVES: Our objective was to determine if severity measures commonly present in unstructured clinical data of congestive heart failure (CHF) patients could be extracted using a text-to-text transfer transformer (T5) model.

METHODS: Comprehensive clinical notes collected from a large United States medical institution were included if they contained strings suggesting a patient history of CHF. These notes were randomly sampled and analyzed using a pretrained T5 model available in a licensed software package using free-text sentences as the questions (e.g., “What is the ejection fraction?”) and complete clinical notes as the context. For New York Heart Association (NYHA) class, returned answers were marked correct if they were integers or roman numerals between 1 and 4. For ejection fraction (EF), returned answers were marked correct if they were in percentage form, integers between 0 and 100, or decimals between 0 and 1. Ranges for these data elements were also marked as correct.

RESULTS: Of approximately 9.6 million notes collected, 440 thousand contained CHF-specific strings, and 200 of these notes were randomly sampled for analysis. Processing time was approximately 0.7 hours on a 128-core, 512 GB RAM CPU processor. Four questions (2 for EF and 2 for NYHA class) were presented to the 200 notes for a total of 800 possible answers. Valid EF and NYHA class values or ranges were detected in 31/400 (7.7%) and 16/400 (4.0%) answers, respectively.

CONCLUSIONS: These results suggest that T5 models can extract disease-specific knowledge related to CHF severity from unstructured clinical data. Potential applications include characterizing patients for incorporation into health economic models and improving predictive accuracy for CHF readmission and mortality. Further work is required to determine sensitivity and specificity of valid data elements in analyzed notes and to improve performance, yield, and process automation of extracting these disease severity concepts.