Healthcare utilization in clinically significant tricuspid regurgitation patients with and without heart failure

This study used data from the Optum^® Integrated File, which contains data from both claims and EHRs for patients across the USA [20]. The combination of claims and clinical data offers a more comprehensive view of the patient’s journey and interaction within the healthcare system. Optum’s Integrated Claims – Clinical dataset combines adjudicated claims data with Optum’s EHR data. Optum’s longitudinal clinical and claims repository from Optum Analytics is derived from more than 50 healthcare provider organizations in the USA that include more than 700 hospitals and 7000 clinics, treating more than 103 million patients receiving care in the USA. The Integrated dataset, which leverages a proprietary Optum algorithm that uses both salting and cryptographic hashing, links both claims and clinical data for approximately 24M matched individuals. The Integrated dataset includes historical, linked administrative claim data from pharmacy claims, physician claims and facility claims, with EHR information, inclusive of medications prescribed and administered, lab results, vital signs, body measurements, diagnoses, procedures and information derived from clinical notes using Natural Language Processing [20]. Optum data provide a continuum of treatment and cost including medications by therapeutic area, provider notes with treatment rationale and cost by procedure and condition. All data used to perform this analysis were de-identified and accessed in compliance with the Health Insurance Portability and Accountability Act. As a retrospective analysis of a de-identified database, the research was exempt from Institutional Review Board review under 45 Code of Federal Regulations 46.101(b)(4).

Adults 18 years or older in the Optum^® Integrated File were included for analysis if they had a minimum of one diagnosis code (ICD10: I36.1 nonrheumatic tricuspid valve insufficiency) for tricuspid insufficiency between October 2015 and June 2019 without a record of end-stage renal disease, valve prolapse or primary organ disease. Patients were also required to have 12 months of continuous health plan enrollment before their first TR diagnosis and 1 month of continuous enrollment post-TR diagnosis. The 1-month post-TR diagnosis was used as a landmark period to capture TR severity. No continuous enrollment was applied to the outcome period and because patients had varying follow-up times, all outcomes were annualized per patient year.

The impact of severe TR on mortality among patients with HF is known [21]. However, the impact of TR associated with other conditions (RSHD and OVD) remains unknown. Conventional wisdom assumes that HF should account for the majority of a patient’s clinical burden, so we constructed a hierarchical cohort classification of TR patients as follows: All patients with TR were categorized based on whether or not there was a record of HF (diastolic, systolic, combined or unspecified) before or 1 month (landmark period) following TR diagnosis. For those patients with no record of HF, they were further subdivided first by presence of OVD and next by presence of RSHD at baseline or landmark if OVD was absent.

RSHD was defined as one or more of the following conditions: RVD, COPD, IPAH or AF. Based on this hierarchy four cohorts emerged: TR-only (reference group); TR with HF; TR with OVD (no HF); and TR with RSHD only (no OVD or HF).

Patients were categorized as having severe TR disease with pulmonary hypertension plus presence of either ascites, edema or hepatic insufficiency before or 1 month following (landmark period) TR diagnosis. Patients were also considered to have severe TR if they had a record of TR surgery during the landmark period, or if TR was quantified as moderate or severe in EHR physician notes of the Humedica database.

Outcomes included all-cause hospitalizations, cardiovascular (CV)-related hospitalizations, total hospital days and the percentage change in total expenditures. All outcomes were annualized to reflect the total outcome in the post period divided by the number of years of follow-up. All count outcomes (hospitalizations and total hospital days) were rounded to the nearest whole number to employ Poisson and negative binomial count-regression methodologies.

Zero-inflated methodology was used due to excess number of zeros (>50%) for each count outcome. Both zero-inflated Poisson and negative binomial regression models were created to model all-cause hospitalizations, CV-related hospitalizations and total hospital days. Final models were determined by using diagnostics for overdispersion. Least square means were calculated to show the estimated outcomes given the model results.

Due to the inherent right-skewness of cost data, total expenditures were modeled using gamma log link regression models. For the total expenditure models, relative counts were reported comparing each cohort to the reference group (TR-only) because of restrictions on reporting dollar amounts using Optum data. All models adjusted for the following covariates: age, sex, race, Elixhauser Comorbidity Index (ECI) [22] score and TR severity. Region, payor and income were not included in modeling because each was found to be nonsignificant, or were related to other variables already included (e.g., payor and age). Confidence intervals were reported for each model, with any confidence interval including or crossing one considered statistically not significant. All statistical analyses were performed using SAS software version 9.2-9.4 (SAS Institute, Inc., NC, USA).

These results should be interpreted with caution for several reasons. We recognize that TR was ascertained through sources of automated data that rely on coding. This could have been biased in terms of over- or undercoding. We could only control for known confounders and lacked the ability to control for unknown confounders, ideally by randomization, stratification and matching. Statistical modeling was used to control for the potential confounding effect of known variables with between-group differences. While statistical models controlled for several factors, models could not control for some variables that are not included in an administrative database, such as echocardiography results. Proxies in automated data were relied upon that suggested significant TR without having the quantitation of TR from imaging data. The use of proxies could have significantly reduced the proportion of patients with severe TR, and our data analysis should not be interpreted as a prevalence study. A strength of the present study, however, is that the data reflect real-world patient characteristics and outcomes across the country from different hospitals and physicians as compared with evidence from controlled clinical trials.