The Optum Risk Algorithm and the Quiet Denial of Care
Case Summary
§For years, one of the most widely deployed healthcare risk algorithms in the United States systematically underestimated the medical needs of Black patients.
§The algorithm, developed by Optum and used by health systems and insurers to allocate access to “high-risk care management” programs, did not explicitly consider race. Instead, it used future healthcare cost as a proxy for future health need.
§Because Black patients, on average, incur lower healthcare spending than white patients with the same disease burden, due to long-documented structural inequities, the algorithm concluded that many Black patients were healthier than they actually were.
§The result was not a marginal statistical artifact. It was a large-scale, operational outcome: millions of Black patients were less likely to be flagged for additional care, not because clinicians made a decision, but because a mathematical model quietly filtered them out.
§This case is not about malicious intent. It is about how a design choice, mathematically defensible on paper, institutionally convenient in practice, translated directly into unequal access to care.
What the Algorithm Was Designed to Do
§Healthcare systems routinely face a scarcity problem: there are more patients who could benefit from intensive care coordination than there are resources to provide it. To manage this, many organizations rely on predictive models to rank patients by “risk” and intervene only for those above a certain threshold.
§The Optum model was designed to:
§Predict future healthcare costs for individual patients
§Rank patients based on those predicted costs
§Select the highest-ranked patients for enrollment in care management programs
§From an operational standpoint, this made sense. Costs are readily available, standardized, and already tracked for billing and actuarial purposes. Predicting cost is also a well-studied machine learning task with clear objective functions and measurable accuracy.
§The problem is not that the model predicted costs poorly.
§The problem is that cost was treated as a stand-in for need.
Metric Substitution: When Convenience Replaces Meaning
§In mathematics and statistics, a proxy variable is often used when the quantity of real interest is difficult or expensive to measure directly. This is common, and sometimes necessary.
§But proxy choice is never neutral.
§In this case, the implicit assumption was:
§Patients who will cost more in the future are patients who need more care now.
§Mathematically, this assumption is attractive. It allows the model to optimize a clear numerical target. Institutionally, it is even more attractive: cost aligns with existing incentives, data pipelines, and budgeting frameworks.
§Clinically, however, the assumption is false.
§Decades of evidence show that healthcare spending is not a clean measure of disease burden. It reflects access, utilization, bias, and systemic inequity as much as, or more than, biological need.
§By choosing cost as the optimization target, the model embedded those inequities directly into its output.
A Closer Look at the Mathematics
For readers comfortable with basic modeling concepts, it is worth slowing down here.
Suppose the model estimates a function:
where represents patient features: diagnoses, past utilization, lab values, demographics, and so on.
Patients are then ranked by , and those above a threshold are selected for intervention.
The intended quantity of interest, however, is closer to:
The model implicitly assumes a monotonic relationship between and .
But if two populations have systematically different mappings from need to cost — because one receives less care, fewer referrals, delayed diagnoses, or undertreatment — then:
for the same underlying disease burden.
The model can be accurate at predicting cost and still be biased with respect to need.
This is not a paradox. It is a consequence of optimizing the wrong target.
How the Harm Manifested in Practice
§The consequences of this design choice were empirically documented by independent researchers in 2019.
§Analyzing data from millions of patients, the researchers found that Black patients assigned the same risk scores as white patients were, on average, significantly sicker—suffering from more chronic conditions and higher disease burden.
§Put plainly: the algorithm systematically ranked Black patients as lower priority for care at the same level of illness.
§Because enrollment in care management programs depended on the model’s output, this ranking translated directly into reduced access to preventive services, follow-up, and coordinated care.
§Clinicians did not need to intend discrimination for discrimination to occur. The mathematical gatekeeper had already made the decision.
Institutional Deployment and Deference
§The model was not an academic prototype. It was embedded into real clinical and administrative workflows across the United States.
§Once deployed, it carried institutional authority:
§It standardized decisions across sites
§It reduced discretionary variation
§It appeared objective and scalable
§These properties are often cited as strengths of algorithmic systems. In this case, they amplified harm.
§Few institutions conducted population-level audits to ask who was being excluded. Fewer still interrogated whether the optimization target aligned with clinical ethics.
§The model’s outputs were treated as facts, not hypotheses.
Discovery, Response, and Legal Validation
§The inequity was not uncovered by internal oversight or regulatory review. It was identified by external researchers, who published their findings in Science.
§Following publication:
§The flaw was widely acknowledged
§Model revisions were proposed and implemented
§Health systems reviewed their use of cost-based risk scores
§Class action lawsuits and regulatory scrutiny followed, alleging discriminatory denial of care. These actions did not require proving intent—only demonstrating disparate impact and foreseeable harm.
§The sequence matters. Correction came after exposure, not before deployment.
Overall Image Philosophy for the Optum Case
§Images should:
§feel procedural, administrative, and infrastructural
§suggest scale and impersonality
§emphasize abstraction replacing judgment
§look like they belong in a regulatory filing or court exhibit
§They should not:
§visualize “AI” as glowing brains
§dramatize suffering with faces
§imply insider access you don’t have
§Think: evidence of systems, not stories of individuals.
1. Healthcare Infrastructure (Abstracted, Not Human-Centered)
What works
Why
§Reinforces that decisions are being made at a distance
§Avoids patient exploitation
§Visually communicates scale and routinization
Placement
Caption style
§Administrative care coordination space in a U.S. hospital system.
§Neutral. Non-interpretive.
2. Interfaces Without Context (The Algorithm as Gatekeeper)
What works
§Generic EHR-style dashboards (non-branded)
§Risk stratification screens
§Patient lists ranked by scores
§Any interface that looks procedural
§⚠️ Important:
Use illustrative stand-ins, not leaked or proprietary screenshots.
Why
§Makes algorithmic authority concrete
§Shows how a number becomes a decision
§Helps readers visualize where judgment is displaced
Placement
Caption
§Example of a risk-stratification interface used in healthcare settings.
§Avoid naming Optum in captions unless the image is explicitly public-domain.
3. Mathematical Abstraction (Clean, Minimal, Cold)
What works
Why
§Rewards mathematically inclined readers
§Reinforces that this is a modeling failure, not a moral anecdote
§Provides visual rest without emotional manipulation
Placement
§Immediately after the section:
“A Closer Look at the Mathematics”
§This lets math-literate readers pause and engage, while others skim.
Caption
§Simplified illustration of ranking-based selection using a proxy variable.
4. Institutional Paper Trail (Regulatory, Not Accusatory)
What works
§Public-facing:
Why
§Signals validation and seriousness
§Shows this case exists in the record
§Protects you legally by grounding claims in public documents
Placement
§In the Discovery, Response, and Legal Validation section
Caption
§Public documentation following publication of research on healthcare risk models.
§Again: factual, non-accusatory.
5. Aggregate, Not Individual, Representations of Patients
What works
§Abstract silhouettes
§Crowd diagrams
§Census-style population graphics
§Waiting room wide shots (no faces)
Why
§Emphasizes population-scale harm
§Avoids exploiting patient imagery
§Keeps focus on systems, not suffering bodies
Placement
§In the “What Was Lost” section
Caption
§Care management decisions operate at population scale, not individual discretion.
6. What to Avoid (Especially Important Here)
§For the Optum case, explicitly avoid:
§❌ Faces of Black patients (exploitative, risky)
§❌ Doctors with stethoscopes looking concerned
§❌ Stock “AI brain” imagery
§❌ Flowcharts that oversimplify causality
§❌ Any image implying secret access or whistleblowing
§You are not revealing a secret—you are documenting a mechanism.
Recommended Image Count
§For a longer Optum piece:
§4–6 images total
§One per major conceptual transition
§No image should appear before the reader understands why it’s there
§If you ever ask “does this make it more engaging?”—don’t use it.
Ask instead: “does this make the system more legible?”
Reusable Image Rule for Modern Cases
§You can keep this internally:
§Modern cases should show systems, not faces.
§That single rule will keep you out of trouble—ethically and legally.