High Risk: Intensive Management
Complex, high-utilization patients requiring active care coordination. This tier is usually small, expensive, and the hardest to move, which is why it should not consume the entire program.
Custom Software
Risk Stratification Is a Capacity Allocation Problem, Not a Prediction Contest
The purpose of stratifying a panel is to decide where finite care management capacity goes. A health system with eight care managers and forty thousand attributed lives is not asking who is sick. It is asking which two hundred people should receive outreach this month. That reframing changes what a good model looks like. A model that predicts next year’s costliest patients with impressive accuracy may identify a cohort nobody can help, while a less accurate model that surfaces patients whose trajectory can still be changed produces measurably better outcomes for the same staffing.
Care management capacity is the binding constraint in almost every program. Stratification exists to ration that capacity intelligently, which means the output must be a ranked, actionable list.
A model can be right about who will be expensive and useless for deciding who to call. Optimize for actionability, then measure accuracy as a constraint rather than a goal.
A risk score in a dashboard nobody opens changes nothing. Stratification output belongs in the care management worklist, refreshed on a cadence the team works to.
Care teams act on categories with defined protocols. A continuous score between zero and one requires every user to invent their own threshold, which produces inconsistent practice.
Stratification is one component of a broader program. Our population health entry covers cohort definition and the surrounding program structure.
Tell Us Your Requirements
Our experts are ready to understand your business goals.
Trusted Partners
Trusted by Industry Leaders Worldwide
Recognition
Awards & Recognitions
The Four Tiers and What Each One Is Actually For
Most programs land on four tiers, and the value is not in the labels but in having a distinct intervention attached to each. High-risk patients receive intensive care management. Rising-risk patients receive preventive intervention before they escalate. Stable patients receive lower-intensity monitoring. Low-risk patients receive wellness and screening outreach. If two tiers share the same intervention, you have three tiers pretending to be four, and the stratification is doing less work than the dashboard suggests.
Rising Risk: The Highest-Value Tier
Patients trending toward high risk but not yet there. Intervention here is cheaper, more effective, and routinely underinvested because the patients do not yet look alarming in the data.
Stable: Monitoring and Adherence
Managed chronic conditions needing continuity rather than intervention. Automated monitoring and adherence support handle most of this tier without care manager time.
Low Risk: Prevention and Screening
Largely healthy population where the objective is closing screening gaps. Population-level outreach rather than individual case management is the appropriate mechanism.
Every Tier Needs an Owner
Assign the intervention, the owner, and the review cadence per tier before building the model. Tiers without defined actions become reporting artifacts.
Monitoring Feeds the Tiering
Device and remote monitoring data sharpens tier assignment considerably. Our remote patient monitoring app development guide covers that data pipeline.
Impactability: The Question Most Risk Models Do Not Answer
This is the gap that separates a useful stratification program from an expensive one. Conventional risk models identify patients at risk of elevated cost, utilization, or chronic disease burden, and published work on population risk stratification is direct about the limitation: focusing only on high-cost high-need patients misses those at lower but rising risk who may be considerably more impactable. Some groups have built explicit impactability and propensity-to-succeed models to target patients likely to respond to intervention. Integrating those with conventional risk scoring is the harder engineering problem and the one worth solving.
01
Risk and Impactability Are Different Predictions
A patient may be certain to incur high cost and impossible to help. Another may be moderate risk with a trajectory that outreach genuinely changes. These require separate models.
02
Propensity to Succeed as a Second Score
Model the likelihood a patient responds favorably to the specific intervention available, then rank on the combination rather than on risk alone.
03
Rank on Expected Benefit
The operationally correct ordering is expected change in outcome from contacting this patient, not predicted severity. That is what maximizes return on fixed care management capacity.
04
Integrate Rather Than Choose
Accretive approaches combine complementary models rather than selecting one. Each captures a different subgroup, and integration retains the value each contributes.
05
Engagement Likelihood Is Part of It
A patient who will not answer the phone is not impactable through telephonic outreach regardless of clinical opportunity. Model reachability alongside clinical benefit.
Calibration Matters More Than the AUC Your Vendor Quotes
Vendors lead with discrimination, usually an area under the curve figure, because it is a single impressive number. Discrimination tells you whether the model ranks patients correctly relative to each other. Calibration tells you whether a predicted twenty percent risk corresponds to twenty percent of those patients actually experiencing the outcome. Tier thresholds depend entirely on calibration, because a threshold set on miscalibrated probabilities assigns the wrong patients to the wrong tier while the AUC stays reassuringly high. Ask for calibration curves, not just discrimination.
Discrimination Versus Calibration
An AUC of 0.90 with poor calibration produces correctly ordered patients at wrong absolute risk levels. Since tiers are defined by thresholds, the tiering will be wrong.
Recalibrate Locally
Calibration is population-specific. A model calibrated on a vendor’s development cohort needs recalibration against your population before thresholds mean anything operationally.
Validate Vendor Models on Your Data
External validation studies of widely used vendor readmission models have found that model complexity can hamper generalizability, with updating warranted for new settings. Assume this applies to you.
Set Thresholds From Capacity, Not Convention
Work backward from how many patients your team can manage. Thresholds should produce a list the size of your actual capacity rather than a clinically tidy percentage.
Recheck Drift on a Schedule
Populations, coding practice, and care patterns shift. Governance for ongoing monitoring belongs in a defined framework, covered in our healthcare AI governance framework.
The Data Foundation Stratification Actually Requires
Stratification quality is bounded by data completeness, and most programs discover their gaps only after the first model produces implausible rankings. Claims give you utilization and diagnosis history with a lag. Clinical data gives you current state without the out-of-network picture. Neither alone is sufficient. Social and behavioral factors frequently matter more than clinical severity for predicting who deteriorates, and are the least well captured. Assembling this into one queryable layer, with identity resolved across sources, is the unglamorous majority of the work.
Claims Plus Clinical, Not Either
Claims capture out-of-network utilization your EHR never sees. Clinical data captures current labs and vitals claims lag by months. Programs using only one systematically misrank.
Identity Resolution Comes First
Patients appear differently across claims, EHR, and ancillary systems. Unresolved identity produces duplicate records that fragment risk history and understate severity.
Social and Behavioral Factors
Housing instability, transportation, food insecurity, and social isolation predict deterioration strongly. Capture is inconsistent, so document what is missing rather than treating absence as absence of risk.
Care Gaps Are Signal
Overdue screening, missed appointments, and unfilled prescriptions are leading indicators available before clinical deterioration appears in labs or utilization.
Build the Layer Once
Stratification, quality reporting, and analytics draw on the same foundation. Our healthcare data lake architecture post covers assembling it.
Clinical Governance, Explainability, and Bias
A stratification model determines who receives clinical attention, which makes it a resource allocation mechanism with equity implications rather than a neutral analytics output. Published research has shown that models using healthcare cost as a proxy for health need can understate illness burden in populations that historically consumed less care, producing systematically lower risk scores for equally sick patients. That failure mode is a design choice about the target variable, not a modelling error, and it is invisible unless you test for it deliberately. Clinical committees are right to ask.
Choose the Target Variable Carefully
Predicting cost is not predicting need. Where access has historically differed across groups, cost-based targets encode that difference as lower clinical risk.
Test Performance Across Subgroups
Evaluate stratification accuracy separately across demographic and payer groups. Aggregate performance can look strong while one population is consistently under-tiered.
Explainable Criteria, Not Opaque Scores
Clinical committees approve logic they can interrogate. A ranked list with the contributing factors visible per patient gets adopted; a black-box score gets overridden or ignored.
Keep the Care Team in Control
Stratification informs prioritization; clinicians decide who receives what. Enrollment decisions and care plans stay with the care team, with overrides captured as data.
Document the Model for Review
Intended use, population, target variable, performance including subgroup breakdowns, and known limitations. This is what a governance committee needs and what accreditation frameworks increasingly expect.
How We Scope and Deliver Risk Stratification Work
We work only in healthcare, and since 2013 we have delivered more than 500 software projects across over 200 organizations, with over 250 healthcare and EHR integrations completed. On stratification specifically, the pattern is consistent: the model is a few weeks of work and the data foundation is a few months. Programs that fail do so because nobody owned identity resolution, or because the output never reached the care management worklist, not because the algorithm was insufficiently sophisticated. We scope accordingly, front-loading data and workflow rather than modelling.
Data Readiness Assessment
We inventory available sources, assess completeness and identity resolution quality, and identify what stratification your current data can actually support. Fixed scope, delivered before any model work.
Stratification Build: $40,000 to $80,000
Data pipeline, tiering model with local calibration, explainable output, and integration into one care management workflow. Suited to a first program or a single line of business.
Platform Build: $80,000 to $200,000
Multi-source ingestion, combined risk and impactability scoring, subgroup performance monitoring, care gap integration, and worklist delivery across multiple programs and populations.
Enterprise Programs: $200,000 and Above
Multi-entity populations, payer and provider data integration, model governance infrastructure, and the documentation and monitoring required for clinical committee oversight at scale.
Ongoing Model Operations
Recalibration, drift monitoring, subgroup review, and threshold adjustment as capacity changes. A stratification model left alone for a year is ranking last year’s population.
Related Clinical AI Work
Post-discharge prioritization is a distinct model on a different population. See our AI post-discharge risk prediction page for that scope.
FAQs
Frequently Asked Questions About Patient Risk Stratification
These questions come up in most stratification conversations, usually after a first vendor demo has quoted an AUC figure and a care management director has asked what it means for their worklist. Answers reflect general practice as of September 2026. Model performance, calibration, and tier thresholds are population-specific, so treat any general guidance, including this, as a starting point rather than a specification.
Segmenting a patient population into tiers by risk level so that finite care management capacity is allocated to the patients most likely to benefit from intervention.
Usually four: high risk, rising risk, stable, and low risk. The right number equals the number of genuinely distinct interventions you can deliver, not a modelling preference.
Rising risk, generally. Those patients are cheaper to help and more responsive to intervention, yet they attract less attention because they do not look alarming in utilization data.
Not on its own. Discrimination tells you ordering; calibration tells you whether predicted probabilities are accurate. Tier thresholds depend on calibration, so ask for calibration curves.
Not safely. External validation studies have found vendor models generalize poorly to new settings, and calibration in particular is population-specific and needs local assessment.
Clinical need. Cost as a proxy for need can systematically understate illness burden in populations that historically used less care, producing inequitable tier assignment.
Model development is weeks. Data assembly, identity resolution, and workflow integration typically dominate, putting a first program at three to six months.