Program

In-person AND virtual! – We are pioneering a new conference format that will connect in-person and virtual audiences to create a unique experience. Matching the innovation that comes through our members’ work, ISPOR is pushing the boundaries of innovation to design an event that works in today’s quickly changing environment. 

In-person registration included the full virtual experience, and virtual-only attendees will be able to tune into live in-person sessions and/or watch captured in-person sessions on-demand in addition to having a variety of virtual-only sessions to attend.

Using Biased Proxy As Outcomes for Prediction Models: Are We (Re)Producing Health Inequalities?

Speaker(s)

Khor S1, Basu A1, Hahn EE2, Haupt EC2, Lyons LJ2, Henry CP1, Shankaran V3, Heagerty PJ1, Bansal A1
1University of Washington, Seattle, WA, USA, 2Southern California Permanente Medical Group, Pasadena, CA, USA, 3Fred Hutch, Seattle, WA, USA

Presentation Documents

OBJECTIVES:

It is common practice to capture health outcomes in electronic health record data using proxy based on healthcare utilization because detailed chart review is often not feasible. However, since healthcare utilization often varies systematically across racial subgroups, these proxy outcomes can be biased. We aimed to examine the racial bias in a utilization-based proxy for colorectal cancer (CRC) recurrence when the goal is to predict risk of recurrence to inform care decisions.

METHODS:

The study cohort consisted of adults with CRC who underwent resection in a large integrated healthcare system. CRC recurrence was identified using validated healthcare utilization algorithms. We compared this utilization-based recurrence against the chart-reviewed gold-standard recurrence status and assessed the positive and negative predictive values (PPV; NPV) by racial/ethnic subgroups.

RESULTS:

Among 175 patients (mean age 64;48% female), 51% were non-Hispanic White (NHW), 25% Hispanic, 11% Black/African American (AA), 11% Asian/Pacific Islander. The utilization-based proxy for recurrence had good overall performance (PPV=85%,NPV=95%). However, we observed higher false positive rates (FPR) among racial/ethnic minorities, which translated to lower PPVs of 73% in Hispanics and 82% in AAs, vs. 90% NHWs. The higher FPR also inflated the 5-year cumulative incidence of recurrence estimated using the utilization-based outcome in the Hispanic and AA subgroups compared to the gold standard (Hispanic 57%vs.41%; AA 55%vs.43%, compared to NHW 37%vs.37%). NPVs were similar across groups (NHW 95%; Hispanic 95%; AA 100%). Detailed chart review revealed that higher FPR among racial minority may be partially due to delayed primary treatment being misclassified as recurrence.

CONCLUSIONS:

The utilization-based proxy inflated the recurrence rates for the racial/ethnic minority subgroups. Using biased proxy outcome to generate risk prediction models to inform care decisions may incorrectly estimate the health needs for Hispanic/Black patients, resulting in inappropriate healthcare recommendations that further perpetuate healthcare disparities.

Code

MSR27

Topic

Health Policy & Regulatory, Methodological & Statistical Research, Study Approaches

Topic Subcategory

Artificial Intelligence, Machine Learning, Predictive Analytics, Electronic Medical & Health Records, Health Disparities & Equity

Disease

No Additional Disease & Conditions/Specialized Treatment Areas