
How research at Kaiser Permanente found that NLP can assist in identifying hospitalizations for worsening heart failure
Heart health is nothing to be taken lightly. A healthy functioning heart is a key factor in providing every part of your body the blood it needs to provide nutrients and oxygen and remove the waste necessary for life. If there is presence of disease there are times when the heart decompensates, meaning that the heart is no longer able to maintain an efficient circulation allowing for this imperative exchange amongst tissues. According to the Center for Disease Control, in 2017 approximately 6.5 million adults in the United States had heart failure, and it was a major factor of mortality in 1 of 8 deaths. In 2017 it was reported that average wait times for a replacement heart in the U.S. is 191 days, and the cost can approach 1.4 million dollars. So, it is of no surprise that it is better to take care of the heart you have and imperative to capture decompensation as early as possible. Unfortunately, sometimes decompensation can be missed.
NLP workflows in contrast to manual abstraction
Imperative information is often hidden within free text and reports that are attachments within the Electronic Health Record (EHR). A myriad of key cardiology measures can be abstracted from EHRs, be it text based notes to test reports that are PDF attachments, for example: ejection fraction measurement; symptoms such as shortness of breath, fatigue and palpitations; New York Heart Association classifications, and B-type natriuretic peptide (BNP).
There was a recent article published in the Journal of the American College of Cardiology by Rishi Parikh, Senior Consulting Data Analyst at Kaiser Permanente, discussing how NLP can assist in identifying hospitalizations for worsening heart failure (WHF).
In this study, the Kaiser Permanente team randomly chose:
- 50 hospitalizations with a primary discharge code for heart failure
- and 75 hospitalizations with an equal mix of primary and secondary discharge codes for heart failure
Worsening Heart Failure (WHF) was determined based on “the presence of ≥1 symptom, ≥2 objective findings including ≥1 sign, AND ≥1 change in HF-related therapy.” Provider notes and radiology reports were examined using Linguamatics NLP. Researchers established that the 75 patients within the validation cohort had a mean age of 76±13 years, consisting of 44% women. Of them, 31% had an ejection fraction <40%. Ten (26%) hospitalizations had a secondary diagnostic code for HF which met the definition of WHF based on manual review and were correctly detected by the NLP algorithm with only one additional false positive. They found that the NLP algorithm had exceptional performance in contrast to manual abstraction with a sensitivity of 100.0% and a specificity of 96.4%.
Early detection for decompensating heart failure allows for more aggressive treatment, therefore allowing for delays in mortality. By adding NLP workflows in either real-time or in overnight batches, many lives could be positively extended. Although not a cure, implementing these NLP workflows would offer a much better outcome in improved clinical care.
This has proven to be another instance where artificial intelligence like natural language processing (AI-NLP) has the capacity to augment & outperform manual abstraction and allow you to find that critical information quickly and effectively.
Find out how other hospitals and leading academic medical centers are using NLP to improve patient outcomes.
