From Data to Dollars: Improving Medical Billing Accuracy Using Natural Language Processing

First Author: Julianne Gent, MPH – Analytics Developer, Emory Healthcare
Co-Author: Craig S. Jabaley, MD, FCCM – Emory University


Introduction: Maintaining accurate billing documentation in healthcare is essential to ensure timely payments and reduce claim denials. Modern electronic medical records (EMR) have limited ability to extract critical care and evaluation/management (E&M) billing time. We created a natural language processing (NLP) algorithm that processed clinical notes and accurately extracted critical care/E&M time.


Methods: Emory Healthcare (EHC) is the largest academic healthcare system in Georgia, featuring 9 acute care hospitals and over 350 ICU beds. EHC employs 400 critical care providers who write more than 250,000 billable notes per year. EHC uses the Epic Systems EMR. Using SQL, we extracted clinical notes written by critical care providers. We used the R packages ‘stringr’, ’tidyverse’ and ‘lubridate’ to extract billing data from the time a provider stated they rendered service. These data were linked to scheduling data extracted via the QGenda API, to associate clinical documentation with specific shifts worked. The primary outcome was the percentage of notes containing billing time. A secondary outcome metric was the type of billing (critical care vs. E&M time). Differences in percentages between groups were analyzed using McNemar’s Chi-squared test. All statistical analyses were performed using R (RStudio, Inc. Boston, MA).


Results: 459,519 billable notes were written by critical care providers from Jan 1 2024-July 23 2025. The NLP method extracted billing time significantly more often than the EMR SmartText method (93.6% vs. 61.2%, p < 0.0001). Critical care vs. E&M time was successfully differentiated almost 90% of the time via the NLP method, whereas the SmartText method could not differentiate critical care vs. E&M time for any of the notes. Both methods did not differ significantly in identifying notes with missing billing time documentation (20% for NLP method vs. 20% for SmartText method; p = 0.97).


Conclusions: An NLP algorithm extracted significantly more billing time data from each note compared to the EMR software’s native methods. Our algorithm differentiated whether the billing time was critical care/E&M and identified notes with missing billing documentation.