NR 588AI Week 4 Bias and Equity Review Example

Bias and Equity Review: How the Readmission Risk Model Performs for Seven Patient Groups, Commissioned by the Governance Committee With the Consequence Agreed First

Student Name

College of Nursing, Chamberlain University

NR 588AI: Advanced Nursing Leadership in Artificial Intelligence-Integrated Complex Systems

Instructor Name

Month Day, Year

Bias and Equity Review: How the Readmission Risk Model Performs for Seven Patient Groups, Commissioned by the Governance Committee With the Consequence Agreed First

The Prior Question

Before asking whether the model treats groups differently, the review asked whether a model should select patients for transitional care at all. The governance committee's answer was yes, provisionally, because our silent run showed it selected more of the patients later readmitted than our current criteria. This review asks whether that gain is shared across groups.

Commissioned by Whom, and the Consequence Agreed First

The System Artificial Intelligence Governance Committee commissioned this review on 18 November 2026, before go-live. It agreed in advance what would follow: if the model's capture rate for any group fell more than 8 percentage points below its overall rate, use of the model for that group would be changed before go-live. The committee holds the go-live decision open until it acts on this review.

Choosing Groups for a Reason

Seven groups were chosen because there is a plausible reason to expect the model to perform worse for them.

Black patients and Hispanic patients: because models trained on records of use and spending can under-rank groups who receive less care for the same illness (Obermeyer et al., 2019).

Patients whose preferred language is Spanish: because fewer contacts and interpreted visits may leave thinner records.

Patients living in rural ZIP codes: because they may use hospitals outside our system, so their prior visits are missing.

Patients with no primary care provider on record: because the model uses prior outpatient visits as inputs.

Patients covered by Medicaid or uninsured: because coverage affects access to prior care.

Patients aged 85 and over: because the model may weight age heavily.

The Data

All 36,400 adult discharges from our four hospitals in 2025, scored silently. A patient was "captured" if they were in the top 20% of scores on the day of discharge and were readmitted within 30 days. Capture rate is the share of a group's readmitted patients whom the model selected.

Disaggregated Results

Interpreting the Results

Four groups fell well below the overall capture rate of 41%: Black patients (34%, 7 points below, just inside the threshold), Hispanic patients (33%, 8 points), Spanish-preferring patients (29%, 12 points) and patients with no primary care provider (27%, 14 points). Strictly, the threshold was crossed by Spanish-preferring patients and patients with no primary care provider; Hispanic patients sit at the line and Black patients just inside it.

The pattern is consistent: these groups have higher readmission rates than average but are less likely to be placed on the list. Patients with no primary care provider are readmitted most often of any group except the oldest, yet only 11% are listed. The likeliest explanation is the model's reliance on prior outpatient visits, which these patients have fewer of. The model is reading thin records as low risk. That is the same mechanism Obermeyer et al. (2019) described, working through visits rather than spending.

Patients aged 85 and over are listed more often than any group and captured best, which suggests age is weighted heavily. That benefits them but uses list places.

Calibration by Group

Capture rate measures how many readmitted patients the model selects. Calibration asks whether a given score means the same risk in each group. For patients with no primary care provider, the observed readmission rate in each score band was about 3 to 5 points higher than the model predicted; for White patients the two matched closely. The model is not only listing these patients less often; it is underestimating their risk at every level. This supports the explanation that thin records read as low risk.

How This Compares With Readmission Models Generally

Readmission models have generally shown modest ability to separate patients who will and will not be readmitted (Kansagara et al., 2011), and many rely on prior use of care. The pattern here is therefore a known risk of the model type, not a peculiarity of this vendor.

Groups the Analysis Could Not See

Patients without a fixed address are not reliably recorded. Race and ethnicity were missing for 8% of discharges, who are excluded from those rows. Patients who were readmitted to hospitals outside our system are counted only if the regional exchange reported them, which may undercount rural readmissions. Undocumented patients may avoid giving information. Rajkomar et al. (2018) recommend that fairness reviews state which groups could not be assessed, because the absence of data is itself a finding.

Applying the Agreed Consequence

For the two groups beyond the threshold, the committee's consequence applies. I recommend:

A floor: any patient with no primary care provider on record, or with Spanish as preferred language, whose score is in the top 35% is placed on the list. In a rerun, this raised capture for those groups to 39% and 37% while adding about 4% to the list.

Because the floor adds patients, the list grows from 20% to about 24% of discharges, which needs about two more transitional care nurses across the system.

For Black and Hispanic patients, who sit at or near the threshold, quarterly monitoring with the same 8-point rule.

What the Committee Is Asked to Decide

Approve the floor for patients with no primary care provider and for Spanish-preferring patients, approve the request for two more transitional care nurses to absorb the larger list and approve quarterly monitoring of Black and Hispanic patients against the 8-point rule. The committee agreed in November that go-live waits on these decisions.

What the Review Does Not Settle

It does not settle whether 8 points is the right threshold; the committee chose it in advance, and it can be revisited at the annual review.

It does not show whether the transitional care itself works equally well for each group once patients are on the list. A Spanish-preferring patient called by an English-speaking nurse through an interpreter may benefit less. That is a question for the care management program, not the model.

References

Kansagara, D., Englander, H., Salanitro, A., Kagen, D., Theobald, C., Freeman, M., & Kripalani, S. (2011). Risk prediction models for hospital readmission: A systematic review. JAMA, 306(15), 1688-1698. https://doi.org/10.1001/jama.2011.1515

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447-453. https://doi.org/10.1126/science.aax2342

Rajkomar, A., Hardt, M., Howell, M. D., Corrado, G., & Chin, M. H. (2018). Ensuring fairness in machine learning to advance health equity. Annals of Internal Medicine, 169(12), 866-872. https://doi.org/10.7326/M18-1990