Personalized Predictions for the Value of Medical Device Utilization

Precision medicine holds the potential to improve quality and control costs by tailoring treatments to individual patients. Clinical trials, however, are rarely scaled to capture the heterogeneity present in real-world populations. Recent methodological developments combine econometric and machine learning methods to estimate individual-level treatment effects. We use these methods to study the value of personalized medicine for medical device utilization. In particular, we measure the value of left Ventricular Assist Devices (VAD) in Advanced Heart Failure (AHF) patients. VAD’s are expensive, with an implantation cost of nearly $200,000, and existing evidence suggests substantial variation in treatment effects. We measure patient-level treatment effects for both survival and costs.

We employ detailed data on each patient’s diagnoses and treatment history. Treatment selection, however, is almost certainly based on information not observed in our data. We take advantage of two sources of exogenous variation to address this likely selection bias. First, VAD insertion procedures are only performed at a handful of centers. Opening a new center is expensive, but entry is distributed across our study period. VAD center entry should be uncorrelated with changes in the distribution of patient severity, but could be correlated with the average level of severity and treatment effects could differ across centers. Second, we observe distances between patients and providers. The differential distance between the closest VAD hospital and the next closest AHF hospital should be uncorrelated with the treatment outcome. We will restrict the analysis to markets with patients that are similar to those with VAD centers to avoid incidental correlation between differential distance the errors of the outcome equation.

Models will be estimated using causal forest techniques (Wager and Athey, 2016; Athey, Tibshirani, and Wager, 2017). This approach combines traditional econometric identification strategies such as instrumental variables (IV) with random forests to estimate heterogeneous and non-parametric treatment effects. Random forest algorithms are, in effect, used to test for treatment effect heterogeneity while conventional IV methods are applied locally. This approach can generate patient- or provider-specific treatment effects. Models will be estimated separately for both treatment costs and survival.

This study employs Medicare claims data from 2008 – 2015. These data span all services covered by fee-for-service Medicare, including: inpatient, outpatient, and pharmacy claims (e.g., Parts A, B, and D). These data document patients’ demographics, comorbidities, and utilization. Our sample comprises more than 500,000 patients hospitalized with an advanced heart failure diagnosis during our study period. These data also provide a high-dimensional space for treatment heterogeneity.

Parameter estimates will be used to generate counterfactual predictions for survival and treatment costs for each patient. These counterfactuals will provide guidance on the value VAD treatment in the real-world settings faced by patients and providers. Estimates will also provide insight into the potential survival gains and cost savings derived from a shift to personalized medical decision making.