Racial and Sex Inequities in the Use of and Outcomes After Left Ventricular Assist Device Implantation Among Medicare Beneficiaries

This cohort study assesses inequities in access and outcomes in the receipt of left ventricular assist device (LVAD) therapy for Black and female patients in the US.

Data challenges for sample selection. While heart failure is common, ventricular assist devices are only relevant for a small fraction of advanced heart failure (AHF) patients. Claims data generally, and ICD-9 codes, in particular, do not capture AHF and extant codes are measured with error. Diagnoses may be documented imprecisely (e.g., with three, as opposed to five digits) or not at all. These data do, however, capture many variables that are correlated with advanced heart failure such as repeated heart failure hospitalization as well as contraindications for LVAD implantation.
We thus employ a data-driven approach to selecting our analytic sample. We constructed a large dataset of patientlevel variables that might be correlated with LVAD treatment. These variables are further described below. These data are drawn from the current hospitalization (demographics and diagnoses present on admission) as well as from inpatient and outpatient encounters in the preceding six months. The variables exclude provider-level factors and variables that could be a consequence of treatment decisions in index hospitalization as these could be endogenous. The variables also exclude patient race and sex as we do not wish to "control" for these factors when we measure disparities.
Empirical challenges. Our approach is similar to propensity score estimation, but we face two important challenges. First, we have a high-dimensional set of patient variables and an unknown functional form. We will address this issue using machine learning methods and cross-validation. Second, our data have a large class imbalance. Heart failure hospitalizations are prevalent (486,017 in our sample), but LVAD treatment is rare (7,135 in our sample). This imbalance exists because LVAD would be inappropriate for the vast majority of patients -who aren't relevant to our analysis -and because access may be limited by proximity and access to LVAD treatment centers. This large class imbalance illustrates the importance of selecting a relevant sample of non-LVAD patients, but it also poses a practical challenge for estimating our LVAD propensities -a model that predicted LVAD=0 for all observations would be 98.6% accurate (and 100% useless) in our sample. Furthermore, accurate prediction among HF patients with a low LVAD propensity is all but useless, we need to train a model that's accurate for patients with a nontrivial probability of receiving LVAD treatment.
Estimation strategy. We address the class imbalance using the synthetic minority oversampling technique (SMOTE). The SMOTE technique (Chawla et al., 2002) generates a synthetic sample that oversamples the rare outcome and undersamples the more prevalent outcome, to generate more accurate predictions in imbalanced data. We employed a 1:3 ratio for oversampling and undersampling. 1 We used eXtreme Gradient Boosting Training (XGBoost) algorithm (Chen and Guestrin 2016) to generate propensity scores with our synthetic sample. Cross-validation (10fold) was used to select appropriate hyperparameters that maximized out-of-sample area under the ROC curve. This measure was chosen because it provides better discriminative power between the two values of the variable (between LVAD treatment and no LVAD treatment). Using the hyper-parameters chosen using the cross-validation, we then build a final XGBoost model for estimating the propensity scores. 2 The final model had an AUC of 0.9045.
Patients with a predicted probability of LVAD treatment ( � ) of less than 0.05 were eliminated from our sample. This restricts our sample size to 15,076 observations with a common support for the LVAD propensity.
We further eliminated 19 observations with missing geographic data a further 13 observations resided in zip codes for which social deprivation data did not exist. Finally, we eliminated the 311 Hispanic patients in our sample.
While this is an important and interesting population, the sample sizes were too small to calculate meaningful parameters.
Our analytic sample comprises 6,825 LVAD patients and 7,908 non-LVAD patients, a total of 14,733 observations.
Subsamples. Our analyses also examine subsamples using patient's low-income-subsidy (LIS) status and survival conditional upon LVAD receipt. LIS status is a beneficiary-specific measure of income and is only available for Medicare Part D beneficiaries. These data are only available for about 64% of our sample. In Table 2, Model 5, we lose 5372 observations. Two additional observations are lost in Table 2, Model 6. These observations are "present" in the SDI data, but the index value is missing.
The models reported in Table 3 are based on the 6,825 observations that received LVAD therapy. The sample size falls to 6,739 observations in Model 4 as we are limiting the model to within hospital variation. This produces collinearity problems for low-volume VAD centers. 3 As in Table 2, we lose 2,512 observations when the sample is restricted to Medicare Part D beneficiaries.

A. Severity Adjustment for LVAD and survival.
We predict one-year survival for each LVAD patient to measure severity. These predictions use the same independent variables used to estimate the LVAD propensity. The variables include diagnoses present on or in the six months preceding the index admission, prior utilization, and demographics excepting race and sex. The model is estimated using an XGBoost algorithm. SMOTE was not used as one-year survival among LVAD patients is reasonably balanced across outcome classes (i.e., N = 304, 796 with ≥ one-year survival and N=188,056 with < oneyear survival). 10-fold cross-validation was employed to select hyperparameters of the XGBoost algorithm. The final model was then used to predict one-year survival for each LVAD patient. This model had an AUC of 0.94513.