By the time a member shows up in your high-risk cohort, they have usually been high risk for months.
That is the part risk stratification doesn't advertise. The model is fine. The inputs are old. Most stratification programs run on adjudicated claims, and adjudicated claims describe what happened to a person some weeks or months ago. So the flag arrives on time by the program's own schedule, and late by the member's.
This article is about what that delay looks like from the care management side: where it comes from, what it does to the first conversation with a member, and what a team should ask of any risk model it relies on. For the version of this argument aimed at buyers comparing AI tools, see The model was never the bottleneck.
Why does risk stratification flag members so late?
Picture a composite member. Call him Walter. In March he goes to the emergency room for shortness of breath and is sent home. In April he sees a new specialist. In May the medication he has filled like clockwork for years stops being filled.
Each of those moments is a signal. None of them reaches the risk model when it happens. Claims are submitted after the visit, worked through by the plan, and land in the warehouse on their own timeline. The stratification run then happens on its own fixed schedule. The signals from March, April, and May arrive together, as a batch, in July.
In July the model does exactly what it was built to do. Walter's score jumps. He moves into the high-risk cohort. Care management gets a new name on the list.
Everyone does good work. They are doing good work on a picture of a person from months ago.
What the late flag costs the care team
Imagine Dana, a care manager, opening her new assignments on a Monday in July. Walter is on the list. The summary tells her about an emergency room visit and a lapsed medication. She plans her call around those facts.
When she reaches him, the story has moved on. The shortness of breath has become a pattern. There was a short hospital stay in June that no one on the care team was told about. He was discharged with new instructions he only half understands. The conversation Dana prepared for is about a member who no longer exists. The conversation she needs to have is about the weeks since, and she is learning them from Walter on the phone.
Nothing went wrong in a way anyone would notice. The model was accurate. The list was delivered. Dana made the call. But the window in which a timely call could have changed Walter's spring has passed, and much of her first conversation is spent catching up instead of helping.
Multiply that across a caseload and the pattern is easy to miss and expensive to live with:
- Care managers spend their first contact rebuilding the recent history the model could not see.
- Members who were rising quietly get their first outreach after the rise has already happened.
- Members whose situation has already settled stay on the list, because nothing newer has arrived to move them off.
- Leaders judge the program by outcomes, while the timing that shaped those outcomes rarely shows up in a report.
What the claims file never shows
There is a second problem underneath the first. Some of the moments that actually predicted Walter's decline were never going to be in the claims file at all.
The missed follow-up appointment produces no claim, because nothing happened. The discharge nobody was told about shows up eventually as a hospital claim, but the fact that no one followed up does not. The refills that quietly turned into none are an absence, and absences are hard to see in data built from things that were billed.
So a model trained and run on claims is not only late. It is looking in a place where several of the most telling signals rarely appear. Better math applied to the same file does not change that. This is the Insight Gap in its purest form: not missing data, but late data, and data that sits in other systems the model was never connected to.
Accuracy is a property of a date
Risk models are usually described by how accurate they are. That accuracy is real, but it is accuracy about a moment. A model that describes a member's state in March very well is not, for that reason, describing the member a care manager is talking to in July.
This is why tuning the model can disappoint. A team can invest in a better algorithm, see improvements on paper, and still find that care managers still feel they are arriving late. They are. The distance between the data and the day of the call stayed the same, and no amount of model quality closes that particular distance.
The same pattern shows up wherever a list is built from an old extract. A care gap list built on a stale extract sends teams after gaps that already closed. A late discharge signal means the transition of care is missed before anyone knows it began. Risk stratification is the same problem, applied to the people who need help most.
What good risk stratification looks like in a care team's week
Good stratification is not a better score. It is a shorter distance between something happening to a member and a care manager knowing about it. In practice, that changes the week in a few concrete ways:
- Cohorts that moveThe high-risk cohort is re-sized as new information arrives, instead of being rebuilt on a fixed schedule. Members enter it closer to the moment their risk changed, and leave it when their situation settles.
- More than claimsPharmacy fills, results, and information from the medical record sit alongside claims, so a lapsed medication or a hospital stay can register before the claim does.
- A first call about nowThe care manager's summary describes recent events, so the first conversation is about what the member needs this week rather than a reconstruction of the spring.
- Freshness you can seeLeaders can tell how recent the information behind each flag is, so a late flag is visible as late instead of looking identical to a current one.
The question to ask of any risk model
So the question worth asking of any population health program isn't "how good is the risk model." It is "how old is the youngest data the model is allowed to see."
Not the average age of the data in the warehouse. The freshest signal that actually reaches the stratification run, and then reaches the care manager. If nobody at your organization can answer that plainly, that is the finding, and it is worth the time to find out.
ElevateCare 360's Risk Module stratifies populations by clinical and utilization risk for proactive outreach and care management. Its cohort analytics segment members into living cohorts that are re-sized as new data lands, and every module reads from the same unified roster, claims, and clinical layer. You can see how that fits together on the ElevateCare 360 platform overview, or contact the Elevate team to see how ElevateCare 360 can help your organization.
Frequently asked questions
- What is risk stratification in care management?
- Risk stratification is the process of sorting a population by how likely members are to need more care, so care management can focus outreach on the people who need it most. Most programs score members with a model and assign the highest-risk group to care managers.
- Why do high-risk members get flagged late?
- Most stratification runs on adjudicated claims, which arrive weeks or months after the visit, and the model itself often runs on a fixed schedule rather than as data arrives. By the time a member's score rises, the events that drove it are already in the past.
- Won't a better risk model fix late flags?
- Usually not. A more accurate model applied to the same old inputs is more accurate about the past. The bigger gain comes from shortening the time between an event and the care team knowing about it, and from using sources beyond claims.
- What data helps risk stratification keep up with members?
- Claims remain important, but pharmacy fills, results, and information from the medical record can show changes such as a lapsed medication or a hospital stay sooner. Combining them, and updating cohorts as that data arrives, keeps the picture closer to the member's current state.