Health plans have always known which members are high-cost. The claims system produces a clear record of who has generated the most medical spend in the prior year. What most health plans cannot do reliably is identify which members will be high-cost next year — specifically, which currently lower-cost members are on a trajectory toward a high-cost event that proactive care management could prevent.
The distinction matters enormously. A care management programme that targets the members who were most expensive last year is managing an existing high-cost population. Some of those members have stabilised chronic conditions that are under good clinical management and will remain expensive but predictably so. Others are on a deteriorating trajectory. And the members who are not yet expensive but will be in three to six months — the rising-risk population — receive no attention because the claims record does not yet flag them.
That population is where the most preventable cost sits. A preventable hospital admission costs between $12,000 and $20,000. The clinical deterioration that leads to that admission is typically visible in pharmacy adherence patterns, laboratory result trends, emergency department utilisation frequency, and primary care visit gaps — weeks or months before the admission itself. A model that reads those signals can identify the rising-risk member in time for a care manager to make effective contact, address the clinical gap, and prevent the event.
The lag problem in claims-based identification
Claims data is a retrospective record. A member who fills their diabetes medications regularly will have pharmacy claims that support good adherence. A member who stops filling those medications — perhaps because of a coverage change, a cost sensitivity issue, or an early sign of declining engagement with their care — will not appear as an adherence problem in the claims record until enough time has passed for the pattern to be visible. By the time the claims-based risk score reflects the change, the clinical deterioration may already be advanced.
The same lag applies across every chronic condition signal in the claims record. Laboratory result trends are not captured in claims data at all — only the fact of the lab visit. Emergency department visits that do not result in admission appear as individual claims rather than as the utilisation pattern they represent. Primary care visit gaps are visible only in their absence, which requires comparing against a baseline that most claims-based models do not construct at the individual member level.
The three to six month lag between a clinical change and its full representation in the claims risk score is the window that separates a preventable high-cost event from one that has already occurred. Closing that window requires signals that are more current than claims.
The economic case
On a population of 100,000 members, preventing 500 avoidable hospital admissions annually — 0.5 percent of the population — is worth between $6 million and $10 million in avoided medical cost. That estimate uses the industry benchmark of $12,000 to $20,000 per preventable admission. The actual number achievable depends on how many of those admissions are genuinely preventable through care management intervention, which varies by condition type and member engagement responsiveness.
The care management resource cost of achieving that prevention rate depends on how efficiently the right members are identified. A model that generates a list of 10,000 high-risk members for 5,000 available care management contacts will miss half the highest-priority interventions and spend resources on members who would not have had a preventable event regardless of outreach. A model that generates a prioritised list of 5,000 rising-risk members responsive to intervention can direct the same care management capacity toward the cases where it will have the greatest impact.
The efficiency of identification is therefore as important as the accuracy of identification. A highly accurate model applied to the wrong population — the historically high-cost rather than the currently rising-risk — produces lower returns per care management dollar than a less precise model applied to the right population.
What effective risk stratification looks like
The feature set that identifies rising-risk members earlier than claims alone uses pharmacy adherence patterns as the primary signal: a member who has been filling chronic medications regularly and then shows a gap is demonstrating a change that, depending on the condition, is predictive of clinical deterioration 30 to 90 days ahead. Laboratory data, where available through data sharing or integrated delivery network relationships, provides direct clinical insight that pharmacy data can only approximate. Emergency department visit frequency and pattern — is this member using the ER for primary-care-sensitive conditions — provides a utilisation signal that is faster-moving than hospitalisation claims.
The model should produce a prioritised list of members with inferred clinical risk drivers, not just a risk score. A care manager who receives an alert that a member has a rising risk score and an inferred driver of medication non-adherence in a specific condition can make a targeted, informed outreach call. A care manager who receives a risk score without clinical context makes a generic outreach call that is less likely to engage the member and less likely to address the underlying clinical issue.
The technology dimension
Risk stratification models require integrating claims, pharmacy, laboratory, and behavioral data across the member population. For health plans running their core operational data on IBM Z, the claims and pharmacy data that provides the foundation of the risk model is resident on the same platform. Deploying the risk stratification model on IBM Z enables daily or weekly scoring of the full member population against the most current available data, without the data extraction and movement overhead that off-platform analytics environments require. The scored output — members ranked by rising-risk probability with inferred clinical drivers — feeds the care management workflow system that manages outreach assignment.
What success looks like
The primary metrics are avoidable ER visits, 30-day readmission rate, preventable hospital admission rate, and cost per managed member. These should be tracked against a baseline established before the programme deploys and measured against a comparable control population where possible. The quality of the identification model is measured by the degree to which the highest-ranked members actually have the clinical events the model predicted, compared against members of similar risk score who did not receive the intervention. That comparison is the evidence base for the programme’s continued investment.
