3C-08 - Are heart failure diagnoses sufficient for heart failure classification? Evaluating NYHA class and ejection fraction extracted from unstructured EHR notes in improving heart failure classification using real-world data

Background: When studying heart failure (HF) using real-world data, HF identification often relies on billing diagnosis codes. Unstructured electronic health record (EHR) data, such as free-text notes and echocardiograms, can supplement and enrich the way that we study HF, especially when extracted at scale using large language models.

Objectives: To examine how echocardiogram measures, focusing on left ventricular ejection fraction (LVEF), and New York Heart Association (NYHA) class information extracted from unstructured EHR records can enrich and support HF information captured through diagnosis codes alone.

Methods: Using de-identified EHR data from Truveta US member health systems, we identified patients with a diagnosis code for a HF type (HFrEF, HFpEF) or NYHA class (I-IV). LVEF and NYHA classification data were extracted from unstructured clinical notes using machine learning models in Truveta . We used this data to examine the additional number of patients who could be classified into one of the HF categories. For HFrEF and HFpEF, we quantified the proportion of patients with a diagnosis code who also had an accompanying LVEF value within two weeks of their diagnosis that supported the classification of HFrEF or HFpEF.

Results: We identified 238,425 total patients with NYHA class data using diagnosis codes alone (Class I: 72,133 patients; Class II: 77,984 patients; Class III: 108,810 patients; Class IV: 31,120 patients). Supplementing with extracted NYHA class data resulted in an increase in the number of patients identified across all classes (increases of 86.8%, 173.1%, 81.7%, and 93.2% for Classes I–IV, respectively).

We identified 2,416,407 patients with HF (1,102,347 HFrEF patients and 1,314,060 HFpEF patients). We found an additional 290,356 patients that did not have HF diagnosis codes, but had extracted LVEF data that could classify them as HFrEF or HFpEF (197,752 patients as HFrEF and 92,604 as HFpEF). For patients who had both HF diagnosis codes and LVEF values available within two weeks of diagnosis, we found that most patients had LVEF values that supported their HF diagnosis codes, though the extent of the agreement differed between the HF types (79% and 51% agreement for HFpEF and HFrEF, respectively).

Conclusions: Extracted LVEF and NYHA data from unstructured EHR can enhance and enrich HF classification. Most LVEF values recorded within two weeks of diagnosis show parity with HFrEF and HFpEF diagnosis codes, though discrepancies may arise from timing and data selection. Our study highlights the need for researchers to carefully leverage unstructured EHR data to build robust HF classifications and populations.