After obtaining approval from our institutional review board we used data from our institutional clinical data warehouse to create a dataset of 5203 patients who had at-least one admission to the hospital or were treated in the ER for a primary diagnosis of CHF (ICD9 diagnosis of 402.01, 402.11, 402.91, 404.01, 404.11, 404.91, 428.0, 428.1, 428.20, 428.21, 428.22, 428.23, 428.30, 428.31, 428.32, 428.33, 428.40, 428.41, 428.42, 428.43, or 428.9 tied to an encounter with an encounter date of 1/1/2011 or later in which the patient was >= 18 years old at the time of the encounter and in which the ICD9 diagnosis code is a primary diagnosis) at one of the 5 hospitals in our hospital system over a period of 3 years. All patients in the dataset had at least one Transthoracic Echocardiogram (TTE) with an assessment of the ejection fraction (EF) along with at least one NT-proBNP measurement throughout this period (either as an inpatient or outpatient). Patients with end stage renal disease on renal replacement therapy (585.5, 585.6,403.11,403.91,404.02,404.03, 404.12,404.13,404.92,404.93,V45.11, V45.12, V42.0) and patients who had follow-up for less than 6 months were excluded from the dataset.(Figure 1) The patient’s date of entry into the study was taken as the date of the first available NT-proBNP measurement during the 3-year study period. Patients received NT-pro-BNP levels at variable periods and frequencies across the study period according to the discretion of the treating physicians. The median interval between two successive values NT-proBNP measurements was of 14 (IQR 3-44days).

Using text-mining techniques we extracted time-stamped ejection fraction (EF) values from all the available echocardiogram reports in the EMR and verified the extraction by manual review. The final dataset contained time stamped NT-proBNP measurements along with the nearest available albumin, creatinine, systolic BP (SBP), diastolic BP(DBP) measurements. If there were multiple values for these data points on a particular calendar day, then the maximum value for the day was used for the patient. If the EF was reported as a range in the Echo report, then the lower end of the range was used (for e.g. if the EF was reported as 45-50%, 45% was used). EF was measured by the biplane method (the method was specified only in 26.7% of the reports). A patient was classified has having HFpEF if their minimum EF across all their echocardiograms was greater than or equal to 50% and HFrEF if their maximum EF across all echocardiograms was below 40%⁶. Patients whose EF fluctuated and who moved between the two categories were not classified within either category; instead they were classified as a separate mixed category. Patients who had more than a 25% net increase or decrease in serum NT-proBNP from their initial NT-proBNP in the study period were classified as having an up-trending or downtrending NT-proBNP levels respectively. Patients who had less than a 25% change in their pro-BNP levels were noted to have stable NT-pro-BNP values. To calculate the clinical Charlson comorbidity scores for the patients, all available ICD-CM codes, across all encounters during the study period were used. The Deyo-Quan modification of the Charlson index for was used for the calculation of the co-morbidity scores and classifying co-morbidities.⁷All baseline labs, demographic and clinical variables for each patient were used from the first available measurement during the study period (baseline variables). We had multiple NT-proBNP measurements for each patient and these were treated as a time dependent co-variate in the regression. NTproBNP assays were performed using the proBNP reagent pack (Roche Diagnostics, Meylan, France) with an Elecsys immunoanalyzer (Roche).