Predictive Risk Stratification for Cost Control in Healthcare

Predictive risk stratification is shifting cost control in healthcare away from retrospective utilization review toward prospective allocation of clinical and operational resources. Instead of analyzing costs after they occur, this approach assigns each patient a probability of future cost or adverse events, enabling healthcare systems to design tiered intervention strategies. Care delivery is then aligned with anticipated patient needs rather than applied uniformly — a structural shift that has become increasingly urgent as value-based payment models expand across Medicare, Medicaid, and commercial payers. According to the 2025 AHIP survey, participants represent over 271 million people nationwide — 87.5% of covered lives — signaling that APM adoption is now a system-wide operating reality, not a peripheral experiment[i].

From Population Averages to Probability Distributions

Traditional cost management relies on per-member-per-month averages that obscure the fundamental asymmetry of healthcare spending. Predictive stratification reframes this into probability distributions, recognizing that a small share of patients drives a disproportionate majority of costs.

The scale of this concentration is well documented. According to the Agency for Healthcare Research and Quality (AHRQ) Medical Expenditure Panel Survey (MEPS), 5% of the U.S. population consistently accounts for approximately half of all healthcare expenditures. In 2022, the top 1% of spenders accounted for 21.7% of total healthcare expenditures, while the bottom 50% accounted for less than 3%. Among adults in the top 5% expenditure tier, 75.1% had two or more AHRQ-designated priority conditions — predominantly chronic diseases — confirming that clinical complexity is the primary driver of spending concentration[ii][iii].

Modern stratification models extend beyond identifying current high-cost patients to detecting”rising-risk” individuals likely to transition into high-cost categories within the next 6–12 months. The Johns Hopkins ACG System identifies this rising-risk cohort — patients in the middle tier of the health pyramid whose status is gradually worsening — as the segment most likely to be positively affected by effective interventions, as high-cost patients at the apex are already receiving intensive management and are unlikely to return to lower-need categories[iv].

These models integrate clinical data (lab results, comorbidities), claims and utilization history, and social determinants such as housing stability and access to care. The AHRQ’s Social Determinants of Health (SDOH) Database, developed under the Patient-Centered Outcomes Research Trust Fund, makes this integration possible by linking SDOH variables across five key domains — social context, economic context, education, physical infrastructure, and healthcare context — to patient-level data at the county, ZIP code, and census tract levels. Claims-only models that bypass this integration tend to underestimate both clinical severity and future risk trajectories[v].

Mechanism of Cost Control: Targeted Intervention Intensity

The effectiveness of predictive stratification lies not just in prediction but in how it informs tiered intervention strategies:

• • High-risk patients receive intensive management — home visits, pharmacist support, and telemonitoring
  • Medium-risk patients are managed through structured follow-ups
  • Low-risk patients are supported through automated outreach

In hospital readmission scenarios, predictive models generate discharge risk scores based on comorbidities, length of stay, and social conditions. The financial stakes are significant: under CMS’s Hospital Readmissions Reduction Program (HRRP), which has been in effect since FY 2013, hospitals with excess 30-day readmission rates face Medicare payment reductions of up to 3% on all inpatient claims. For FY 2026, approximately 2,400 hospitals face some level of penalty, with about 8% facing reductions of 1% or more. In FY 2023 alone, CMS issued approximately $320 million in HRRP penalties across 2,273 penalized hospitals — 75% of all program participants[vi][vii].

Against this backdrop, predictive models have demonstrated meaningful financial returns. A large language model algorithm (NYUTron) developed at NYU Grossman School of Medicine predicted 80% of readmissions and saved $5 million in costs. AI models have also been shown to outperform traditional screening tools such as the LACE index by 25–40% in predicting 30-day readmissions, particularly for patients with complex multi-system conditions. Population health lakehouse implementations using unified ML-driven risk stratification have demonstrated 19% reductions in 30-day readmissions within 12 months. These results confirm that clinical resources directed at high-risk individuals yield the greatest marginal benefit[viii][ix][x].

Enterprise Implementations Across Healthcare Models

Integrated delivery networks: Kaiser Permanente’s AI-driven alerting system, which uses hundreds of millions of EHR data points to predict patient deterioration, has been documented to prevent more than 500 deaths per year while reducing high-risk readmissions by 10%. Transition-of-care programs using stratification have demonstrated readmission reductions of around 20% by aligning follow-up intensity with predicted risk.

Accountable care organizations (ACOs): ACOs embed risk stratification into value-based contracts where financial performance depends on managing the total cost of care. The 2024 Medicare Shared Savings Program (MSSP) results—the strongest year since the program’s inception in 2012—illustrate the cumulative impact. Out of 476 participating ACOs, 75% earned performance payments totaling $4.1 billion, and Medicare realized $2.5 billion in net savings relative to benchmarks. ACOs achieving savings demonstrated consistently lower utilization across hospital discharges, ED visits, and Skilled Nursing Facility stays compared to benchmarks.

Payor-provider platforms: Payor-provider platforms, including those developed by Optum (which expanded its AI-based predictive analytics system with advanced risk assessment tools in October 2025), stratify populations using risk scores, disease gaps, and utilization patterns. These systems generate prioritized intervention lists and incorporate social determinants of health to identify nonclinical cost drivers. At a national scale, analytics vendors such as Inovalon process billions of medical events, improving predictive accuracy and enabling robust risk adjustment across diverse populations.

Cost Levers Enabled by Stratification

Across healthcare enterprises, predictive risk stratification activates several consistent cost-control levers:

• • Early identification and intervention reduce avoidable utilization — hospital admissions, readmissions, and emergency department visits. The National Academy of Medicine identifies avoidable care as a major cost driver; analytics-driven interventions have been estimated by the Mayo Clinic to reduce hospitalizations by 15%–25%[xi].
  • Resource allocation efficiency ensures high-cost interventions are reserved for patients with the greatest expected benefit, minimizing waste from uniform care programs.
  • Risk adjustment accuracy ensures appropriate compensation for managing high-acuity populations and prevents underestimation of expected costs in value-based contracts. The CMS-HCC V28 risk adjustment model, fully phased in as of 2026, recalibrated HCC codes to ICD-10-CM standards, expanding from 86 to 115 HCC categories to better reflect actual clinical complexity and cost trajectories[xii].
  • Care pathway standardization identifies variation in treatment patterns and aligns clinical practices with risk-adjusted outcomes.

Operational Constraints and Implementation Challenges

Despite its growing adoption, predictive risk stratification faces several operational challenges.

Data fragmentation remains a key barrier, as EHRs, claims data, and SDOH sources are often siloed. The AHRQ SDOH Database and CMS’s Mapping Disparities Tool provide foundational infrastructure for integration, but interoperability gaps persist across care settings[xiii][xiv].

Model interpretability is a persistent concern: clinicians require transparency into risk score drivers to act on them. According to ONC’s September 2025 Data Brief (No. 80), while 71% of U.S. hospitals reported using predictive AI integrated with their EHR in 2024 — up from 66% in 2023 — fewer hospitals evaluated predictive AI models for accuracy and bias across all or most of their models, and three-quarters reported that multiple entities were accountable for model governance. Small, rural, independent, government-owned, and critical access hospitals lagged significantly in adoption[xv].

Workflow integration is critical: predictive insights must translate into actionable steps within clinical systems rather than remain confined to dashboards. Evidence indicates that predictive accuracy alone does not lead to cost reduction without corresponding changes in workflows and care delivery models.

Equity considerations are increasingly prominent. AHRQ data show that 78.1% of adults in the top 5% expenditure tier had two or more priority conditions, and that Non-Hispanic White patients and individuals aged 65 and older are disproportionately represented in top spending tiers. Incomplete or improperly applied SDOH data can introduce bias into predictive models, potentially widening rather than narrowing health disparities. CMS’s SDOH mapping tool, updated with the latest available data, provides county- and census-tract-level disparities data to support more equitable model design[xvi][xvii].

Emerging Direction: Continuous and Dynamic Stratification

The field is evolving from static risk scoring toward continuous and dynamic stratification. Real-time or near-real-time scoring systems are replacing periodic assessments, enabling providers to respond to event-driven triggers such as missed appointments or abnormal lab results. The global healthcare predictive analytics market reflects this trajectory: valued at $16.3 billion in 2025, it is projected to grow at a CAGR of 17.7% through 2035, driven by the integration of machine learning, NLP, and real-time EHR data for accurate risk prediction. These systems continuously recalibrate using streaming data, shifting stratification from episodic segmentation to ongoing risk surveillance[xviii].

Avasant Viewpoint: From Analytics to Operating Model

From an Avasant perspective, predictive risk stratification is transitioning from a standalone analytics capability to a core operating model for healthcare enterprises. Its economic value lies not merely in identifying high-cost patients, but in systematically reallocating clinical effort based on predicted marginal impact.

Three implications follow from this shift:

1. 1. Stratification must be embedded within clinical workflows. Standalone analytics platforms deliver limited value without integration into care pathways, staffing models, and incentive structures. ONC data confirm that the most common use of predictive AI in hospitals is predicting health trajectories or risks for inpatients — but the fastest-growing uses are administrative: billing and scheduling. Operationalizing stratification for care delivery, rather than administrative efficiency alone, remains the higher-value opportunity[xix].
   2. Rising-risk cohorts represent the primary opportunity for cost control. AHRQ data confirm that high-cost patients are often already consuming intensive resources; the rising-risk middle tier is where intervention yields the greatest return on investment. As Johns Hopkins ACG analysis shows, the goal is to prevent upward migration in the risk pyramid, not just manage patients already at the apex[xx].
   3. Integration across payer and provider data is essential. Fragmented risk assessments lead to misaligned interventions. The record MSSP performance in 2024 — $2.5 billion in net Medicare savings — reflects the maturation of ACO-level data integration; organizations that replicate this model at the payer-provider interface are best positioned to capture similar value[xxi].

Organizations that operationalize predictive stratification as a continuous layer within care delivery are more likely to achieve sustained cost control under value-based care models, including CMS’s emerging ACCESS, TEAM, ASM, and LEAD models, which embed real-time clinical outcome thresholds and expect participants to leverage AI and digital tools as core delivery infrastructure.

References

[i] https://www.ahip.org/news/articles/new-survey-demonstrates-health-plans-continued-commitment-to-value-based-care-models

[ii] https://meps.ahrq.gov/data_files/publications/st560/stat560.shtml

[iii] https://meps.ahrq.gov/data_files/publications/st556/stat556.shtml

[iv] https://www.hopkinsacg.org/wp-content/uploads/2024/06/Johns-Hopkins-ACG-System_Risk-Stratification-Handbook_v062024.pdf

[v] https://www.ahrq.gov/sdoh/data-analytics.html

[vi] https://legalclarity.org/cms-readmission-rates-calculation-and-penalties/

[vii] https://www.symplr.com/wp-content/uploads/2024/08/DataVision-HRRP-Infographic.pdf

[viii] https://www.medicaleconomics.com/view/reducing-costs-and-readmissions-through-data-analytics-in-health-care

[ix] https://www.zynix.ai/blog-predictive-analytics-population-health

[x] https://www.zymr.com/case-study/population-health-lakehouse-19-readmission-reduction

[xi] https://pmc.ncbi.nlm.nih.gov/articles/PMC7467834/

[xii] https://veradigm.com/veradigm-news/cms-2024-risk-adjustment-model-changes/

[xiii] https://www.ahrq.gov/sdoh/data-analytics.html

[xiv] https://data.cms.gov/tools/mapping-disparities-by-social-determinants-of-health

[xv] https://www.healthit.gov/data/data-briefs/hospital-trends-use-evaluation-and-governance-predictive-ai-2023-2024/

[xvi] https://meps.ahrq.gov/data_files/publications/st556/stat556.shtml

[xvii] https://data.cms.gov/tools/mapping-disparities-by-social-determinants-of-health

[xviii] https://marketgenics.co/reports/healthcare-predictive-analytics-market-20644

[xix] https://www.healthit.gov/data/data-briefs/hospital-trends-use-evaluation-and-governance-predictive-ai-2023-2024/

[xx] https://www.hopkinsacg.org/wp-content/uploads/2024/06/Johns-Hopkins-ACG-System_Risk-Stratification-Handbook_v062024.pdf

[xxi] https://www.cms.gov/files/document/fact-sheet-ssp-py24-financial-quality-results.pdf

By Samkit Jain, Lead Analyst, Avasant