Most risk stratification in this country is built to answer one question — who will cost the most next year? It sounds reasonable. It is also the wrong question, and there is a landmark study showing exactly how wrong, and exactly who gets missed when you ask it.

If your care-management targeting runs on prior cost and utilization, this is not an abstract concern. It is probably shaping your target lists right now.

Cost is a proxy for need — and a biased one

A model that predicts cost is not predicting illness. It is predicting spending, and spending follows access, not just sickness.

In 2019, researchers publishing in Science examined a widely used population-health algorithm — the kind applied to millions of patients to flag who needs extra care. It worked the way most do: it predicted future cost and used that as a stand-in for future need.

The problem is that cost and need are not the same. Less money gets spent on patients with worse access, so the model learned to call them healthier than they actually were.

The size of the miss

The finding was not subtle. At any given risk score, Black patients were considerably sicker than White patients with the same score, as measured by uncontrolled chronic conditions.

Because the algorithm equated lower spending with lower need, it under-flagged them. Correcting the model — having it predict illness instead of cost — would have raised the share of Black patients identified for extra care from 17.7% to 46.5%. Nearly a threefold change, from a single decision about what the model was told to predict.

This isn’t only an equity problem — it’s an accuracy problem

Even if fairness weren’t your concern, a model that confuses cost with need is simply wrong about who is sick.

It is tempting to file this under “health equity” and move on. That undersells it. The model was not just unfair; it was inaccurate about the clinical reality it was supposed to measure.

Any organization using cost-trained risk scores to target care management is, to some degree, aiming its resources with a distorted map — missing patients who are genuinely sick but historically under-treated, and over-weighting patients who are simply expensive. Under value-based contracts, that misallocation costs outcomes and dollars.

Where this bites in a real program

Most commercial and homegrown stratification tools still lean heavily on prior cost and utilization. So the bias is not a historical footnote — it likely lives in your current target lists.

The patients most affected are exactly the ones value-based care is meant to reach: under-resourced, under-treated, high genuine need, low historical spend. They look “low risk” precisely because the system has failed them before.

The fix is more tractable than people assume

The same study showed the remedy. When the researchers retrained the model to predict active chronic conditions rather than cost, bias dropped by roughly 84%.

You do not have to abandon analytics — you have to interrogate the label. Ask what your model is actually trained to predict. If the answer is “cost” or “utilization,” you have found a problem worth fixing before you build your next high-needs or LEAD strategy on top of it.

Final Thoughts

The uncomfortable truth is that a risk model inherits the inequities of the data it learns from, and most are trained on the most biased target available: dollars.

As someone trained as a physician who later worked in coding and audit, I have seen how easily a number on a dashboard gets mistaken for clinical truth. The question to bring to your analytics team this quarter is simple and consequential: are we predicting who will be expensive, or who is sick? They are not the same patients — and the gap between them is where both your missed outcomes and your missed savings are hiding.

If your risk stratification is trained on prior cost, it is likely under-identifying the very patients value-based care exists to reach. At HealtheNomics I help organizations audit what their models actually predict — and rebuild identification around need, not spend.

Explore the services: https://healthenomics.com/services-2/

Request a strategy conversation: https://healthenomics.com/contact-us/

Connect on LinkedIn: https://www.linkedin.com/in/muhammad-ayoub-ashraf/

Website: https://www.drayoubashraf.com

Watch on YouTube: https://www.youtube.com/@HealtheNomics