Contributors

Andrew P. Ambrosy, MD, Cardiologist, TPMG SFO. Associate Program Director for Research (Fellowship), Assistant Medical Director, Clinical Trials Program, Kaiser Permanente Northern California Division of Research
Alan S. Go, MD, Regional Medical Director, Clinical Trials Program, Kaiser Permanente Northern California. Associate Director, Division of Research, The Permanente Medical Group. Professor, Departments of Epidemiology and Biostatistics and Department of Medicine, University of California San Francisco School of Medicine

Despite the increasing burden of hospitalizations for HF, there are significant limitations of the existing research on the epidemiology, clinical characteristics, care management, and outcomes of patients admitted for worsening HF (WHF). “Our current understanding of WHF comes from observational studies using administrative or claims data, national databases, or quality improvement registries,” explains Andrew P. Ambrosy, MD. “However, these data sources are inherently limited by the accuracy and completeness of diagnostic coding and voluntary reporting.”

Randomized clinical trials assessing drugs and devices apply standardized definitions to ensure the enrollment of well-defined target populations and to validate outcomes of interest, but this approach is time-consuming and resource intensive. “The promise of Big Data is it can potentially improve efficiency,” says Alan S. Go, MD. “EHRs and advances in data science are making it possible to more quickly automate approaches on a healthcare system-wide level that leverage the full range of structured and unstructured data.”

Assessing Temporal Trends in WHF Hospitalization Rates
In JAMA Open Network, Dr. Ambrosy, Dr. Go, and colleagues reported on rule-based natural language processing (NLP) algorithms that were applied to state-of-the-art EHR data to describe the epidemiology and temporal trends in hospitalization rates for WHF overall and by the degree of left ventricular systolic dysfunction from 2010 to 2019. Using a consensus definition in line with FDA guidance, hospitalizations for WHF were defined as patients having one or more symptoms, two or more objective findings (including one or more signs), and a change in HF-related therapy (eg, new administration of intravenous loop diuretics and/or initiation of hemodialysis or continuous kidney replacement therapy). Signs and symptoms were identified using NLP algorithms that were applied to EHR data.

The study evaluated data on nearly 288,000 HF hospitalizations, with 65,357 patients having a principal discharge diagnosis of HF and 222,635 having a secondary discharge diagnosis of HF. Of note, 49.5% of patients had HF with a preserved ejection fraction (HFpEF), 11.6% had HF with a midrange ejection fraction (HFmrEF), 20.8% had HF with a reduced ejection fraction (HFrEF), and 18.2% had HF with unknown ejection fraction.

 

NLP-Based Algorithms Better Characterize Hospitalizations for WHF
The study team found that the overall rate of hospitalizations for WHF identified using NLP-based algorithms increased from 5.2 to 7.6 per 100 hospitalizations annually during the 10-year study period (Figure). “A gradual increase was seen in hospitalization rates for WHF, and a more prominent increase was seen in patients with HFpEF, particularly in those with systolic function below normal,” says Dr. Ambrosy. Subgroup analyses demonstrated the following based on NLP-based algorithms:

  • HFpEF hospitalization rates increased from 2.6 to 3.9 per 100 hospitalizations
  • HFrEF hospitalization rates increased from 1.5 to 1.9 per 100 hospitalizations
  • HFmrEF hospitalization rates increased from 0.6 to 1.0 per 100 hospitalizations

Applying NLP-based algorithms to structured and unstructured EHR data was technically feasible and highly accurate, according to the study group. “We observed a more than two-fold increase in the burden of WHF hospitalizations when compared with estimates based entirely on principal discharge diagnoses,” says Dr. Ambrosy.


Broader Application May Help Personalize Medicine in WHF
Using NLP algorithms and the full range of EHR data to identify hospitalizations for WHF is an important first step to reducing reliance on claims data and patient registries, according to Dr. Ambrosy. “The next step is to apply this model to identify WHF cases presenting to emergency departments and clinics,” he says. “With a better understanding of the epidemiology, healthcare systems can triage resources appropriately when managing patients with WHF. Hospitalizations for HF are the tip of the iceberg—these are sentinel events that should trigger clinicians to review medications and treatment plans.”

Dr. Go says the NLP algorithms used in the study are adaptable for other healthcare systems. “The key for each institution is to test, refine, and customize algorithms based on the needs of the healthcare system,” he notes. “Importantly, these algorithms are not intended to replace clinical judgment. Instead, they should be used to improve efficiency and accuracy when identifying WHF on a large scale. This technology can help us in our efforts to personalize medicine and improve individual patient- and population-level outcomes.”

Author