Racial Bias-The Unintended Consequences of Population Health Algorithms

The results of a 2019 research article in the journal Science uncovered significant racial bias in commonly used population health algorithms used to identify and assign care to patients with complex, active health needs. Using a large clinical data set, the researchers were able to show that at any given risk score, black patients are considerably sicker than white patients. They were able to demonstrate that remedying this disparity would more than double the percentage of black patients receiving enriched services like care management (from 17 percent to over 46 percent).

The problem at the core of these results is the fact that many of the algorithms used across the health care industry to predict risk, future utilization and cost, rely predominantly on data regarding the same patient’s prior utilization and cost. These predictions are based on the patient’s health insurance claims in the past. At the same time, the researchers found that at any given level of health, Black patients generate lower costs than White patients – on average $1,800 less per year. They were also able to show that Black patients have a very different utilization pattern than White patients with fewer inpatient surgical and outpatient specialist costs and more costs related to emergency visits and dialysis. It is thought that these differences are explained by the presence of significant barriers to access to health care, as well as reluctance and mistrust of the system that leads patients to avoid care. There is also ample evidence that physician’s biases in allocating and referring patients for care could be contributing factors. It is therefore understandable why an algorithm that is solely reliant on costs to predict risk for future costs will underestimate the true medical need of these patients.

The study is compelling in its findings and extremely well thought out and executed. It highlights some of the known pitfalls of deploying population health platforms that rely on single source data and a single predictive risk algorithm.  

Historically, most of the commercially available algorithms were deployed by health plans, and for many years utilized the only data available to these plans which was adjudicated claims data. Thankfully, health care organizations are now able to deploy contemporary population health platforms that can ingest and analyze multi-sourced data which helps create a broader and more accurate view of patients’ true burden of illness and risk. Combining electronic health record (EHR), health information exchange (HIE) and social determinants of health (SDOH) data, creates a rich tapestry of information that is less likely to introduce the kind of racial bias that is described in this study. It is also the reason why population health platforms like NextGen Population Health provide clinical teams with more than one risk algorithm they are able to apply to cohorts of patients as well as individual patients mitigating some of the risk of embedded bias in an individual algorithm.  

The Science article is timely and extremely important as our society and the health care industry struggle to identify and uproot the racial disparities in the way care is provided to all patients. At the same time, it is extremely important for clinical teams to recognize that all algorithms have limitations and unintended biases and that there is never a substitute for sound clinical judgement when it comes to clinical decisions pertaining to the complex needs of patients.