Identification and Classification of Heart Failure

The first inpatient diagnosis of heart failure from Medicare claims using ICD-9 codes (see Appendix Table S1) was identified during follow-up. We classified these heart failure cases as rEF, pEF or uncertain EF by extracting LVEF values that were most proximal to the date of heart failure diagnosis from electronically recorded cardiology reports (echocardiograms or cardiac catheterization) using natural language processing. Specifically, the Medical Text Extraction, Reasoning and Extraction System21 was used to extract expressions identifying LVEF along with important contextual information such as negations in reports and collect the nearest valid value to the expression. A sample of the notes (n=200) was manually reviewed to ensure the basic rules of LVEF extraction were working. Rules were further refined by manually reviewing all false positives for these 200 notes and removing sources of obvious errors. We then validated a random sample 100 of notes by comparing NLP extracted EF values against manual extraction by a chart reviewer. In 99 of these notes, the NLP algorithm extracted numerical EF values accurately. At this point, the NLP algorithm was deemed reliable and was used for extraction of EF values from notes of all included patients. Based on EF values recorded around the time of HF diagnosis, within 1 month prior to the first episode of heart failure up to 1 year after the diagnosis, the cases were classified into heart failure with reduced EF defined as LVEF <45% or with preserved EF defined as LVEF ≥ 45%.22 Patients with heart failure diagnosis without any evidence of LVEF recordings around the time of the diagnosis in our EHR were still included in the analysis and classified to be heart failure with uncertain EF.