Risk and Benefit Informed MTM Pharmacist Intervention in Heart Failure

Heart Failure

Heart Failure, Machine Learning, Medication Therapy Management, Supervised Machine Learning, Population Health Management

Out-of-hospital care of complex diseases, such as heart failure, is transitioning from an individual patient-doctor relationship to population health management strategies. As an example, at our institution, medication therapy management (MTM) pharmacists are being deployed to patients with heart failure with the intent of improving patient outcomes (through proper medication management and adherence) while reducing cost (e.g., keeping these patients out of the hospital). The success of such strategies will be dependent on the ability to effectively direct scarce resources to deliver appropriate/needed care to patients. In this prospective, pragmatic randomized and matched controlled study, the investigators hypothesize that the combination of accurate, data-driven benefit models and MTM pharmacist intervention in patients with heart failure will result in reduced 1-year mortality and hospital admissions. Using our extensive historical electronic health record data, the investigators have developed a machine learning model that, for individual patients with heart failure, predicts risk and benefit (that is, reduction in risk) associated with closing specific care gaps. These care gaps represent standard evidence-based treatments that may be missing for an individual patient, such as beta blockers or flu shots. The investigators will use this model to define three cohorts to be studied: 1) a high risk/high benefit group to be referred for MTM pharmacist intervention, 2) a high risk/high benefit group to continue with existing standard of care (not necessarily involving MTM pharmacy), and 3) a high risk/low benefit group to be referred for MTM pharmacist intervention. Comparison of groups 1 and 2 (for which assignment is randomized) will evaluate the effectiveness of the MTM pharmacy intervention, while comparison of groups 1 and 3 will evaluate the accuracy of the benefit model prediction and importance of appropriate patient selection for treatment. The primary study outcomes will be mortality and number of hospital admissions during 1-year follow-up following study enrollment.

Heart failure is a highly prevalent, complex disease associated with significant morbidity and cost. For example, Geisinger manages over 900 heart failure admissions per year, with each admission costing an estimated $10,000-$12,000. As payment models continue to shift from fee-for-service to value-based, significant investments are occurring in care team resources to help manage populations of patients with heart failure. These care team resources have demonstrated effectiveness. For example, internal Geisinger metrics indicate that interventions led by clinical pharmacists aimed at poorly controlled type II diabetics have resulted in a sustained median 1% (absolute) drop in hemoglobin hemoglobin a1C (glycated hemoglobin). In this new environment, intelligent deployment of limited resources is critical to drive quality and contain costs. In heart failure, current risk prediction have demonstrated poor prognostic abilities and present a barrier to precision delivery of care team resources. Currently approaches are limited due to not fully utilizing rich, highly granular objective data such as imaging, laboratory values, and vital signs, and therefore are not optimized to accurately predict outcomes. The investigators have generated a machine learning model to predict both 1-year survival and heart failure hospitalization within 6 months of echocardiography. This model utilized 169 input variables including clinical data, imaging measures, and 18 care gap variables. Our results showed not only that the machine learning model had far superior accuracy to predict the morbidity endpoints compared to current approaches utilizing billing code data, but also that care gap variables were important for predicting 1-year survival. Moreover, the investigators showed that closing four of the care gap variables (flu vaccination, evidence-based beta blocker treatment, ACE (angiotensin-converting-enzyme) inhibitor/ARB (angiotensin receptor blockers) treatment, and control of diabetic a1C (i.e., values in goal)) resulted in a predicted improvement in 1-year survival of \ 1200 (out of \ 11,000) patients. This study therefore aims to apply this machine learning approach to direct care team resources in a clinical setting to evaluate its impact on patient survival and healthcare utilization.

No related papers found yet.

United States