The Effect of Predictive Analytics on Healthcare Utilization

The fast changing healthcare landscape is often characterized by a shift from fee-for-service with an emphasis on volume to a value-based reimbursement structure with an emphasis on population health management. To thrive in this new environment, providers must be proactive (anticipate problems before they arise) as opposed to reactive (be there to escalate care when an individual patient’s condition deteriorates). To support proactive clinical decision-making, providers must have and rely on effective, real time, risk stratification of their patient population. However, even with a primary care-led clinical workforce, the provider may lack access to healthcare utilization information and thus the ability to stratify their patients based on risk. This lack of information may impede processes for patient engagement and designs for community integration, critical to success in a value-based environment.

This paper evaluates the success of an insurer-driven intervention among their fully-insured Medicare Advantage member population who were diagnosed with Congestive Heart Failure. The insurer implemented a proprietary algorithm using claims data to predict patient-level likelihood of hospitalization. Patients with high risk scores were contacted and offered enrollment in a care management program. The care management program involved the assignment of health coaches to help patients manage their chronic condition by closing gaps in care. The algorithm was applied in 10 waves between July, 2013 and May, 2015. Each patient was assigned a risk score at the time the algorithm was ran and ranked from high to low. In each wave, the care management team dictated to the insurer how many patients they can approach, based on their resource availability and capacity. This created a set of arbitrary cutoff points for each wave, separating treated and untreated members with very similar predicted LOH scores.

The analysis evaluates both the success of the algorithms in predicting downstream health care utilization and the success of the intervention. We utilize the quasi-random assignment into the program to compare healthcare utilization outcomes in a regression discontinuity framework. We compare outcomes for individuals who were selected for treatment and a similar sized control group consisting of individuals whose risk score ranking fall just below the treatment cutoff (threshold).

First, we find that the risk score is a strong predictor of hospitalization and emergency room (ED) visits. Second, our estimates suggest that the outreach program reduces the number of ED visits, and some evidence of a reduction in cardiologist visits. Specifically, we find that the number of outpatient ED visits decreases by about 40% within the first year following treatment. At the same time, while directionally similar, we find little or no statistically significant impact of the outreach program on the number of PCP visits or hospitalizations. Along the extensive margin, we find strong evidence that the probability of having any cardiologist and ED visit decreases sharply. We conduct several robustness checks to test the validity of the RD design and sample selection. In all cases, our results are relatively stable and unaffected by the change of specification and bandwidth selection.