Chronic Condition Cohorting

Methodology

Step 1 — Define Condition Algorithms

Apply CMS CCW reference algorithms or custom clinical logic. Each condition requires:

• Inclusion ICD-10 codes: Primary and secondary diagnosis positions
• Look-back window: Typically 1-2 years of claims history
• Minimum claim count: Usually ≥ 2 outpatient or ≥ 1 inpatient to confirm chronic presence
• Exclusion logic: Rule out rule-out diagnoses, lab-only encounters

For HCC-based grouping, map ICD-10 codes through the CMS-HCC v28 model crosswalk to assign HCC categories and hierarchical condition interactions.

Step 2 — Apply Severity Staging

Within each condition cohort, assign clinical severity tiers:

• Diabetes: Use HbA1c ranges (<7% controlled, 7-9% moderate, >9% uncontrolled), presence of complications (retinopathy HCC 27, neuropathy, nephropathy HCC 18)
• CKD: Stage by eGFR (G1-G5) per KDIGO guidelines; flag dialysis dependence
• COPD: Stratify by exacerbation frequency, oxygen dependence, FEV1 if available
• CHF: NYHA functional class proxy via utilization intensity and BNP levels

Step 3 — Identify Multi-Morbidity Patterns

Cross-tabulate conditions to detect clinically significant comorbidity clusters:

• Cardiometabolic syndrome (diabetes + hypertension + dyslipidemia)
• Cardio-renal overlap (CHF + CKD stages 3-5)
• Behavioral health comorbidity (any chronic condition + depression/anxiety/SUD)

Calculate Elixhauser or Charlson comorbidity indices as supplementary complexity scores.

Step 4 — Segment by Utilization and Cost

Overlay utilization data on clinical cohorts:

• Rising risk: Patients with worsening lab trends or increasing visit frequency but low current cost
• High utilizers: Top 5% by total cost of care; ≥ 3 ED visits or ≥ 2 inpatient admissions in 12 months
• Stable managed: Controlled metrics, regular PCP engagement, low acute utilization

Step 5 — Apply Attribution and Geographic Filters

• Attribute patients to primary care providers using plurality-of-visits or formal panel assignment
• Apply geographic overlays (HSA, HRR, zip-code clusters) for regional analysis
• Filter by line of business (Commercial, Medicare Advantage, Medicaid, ACA Exchange)

Step 6 — Validate Cohort Integrity

• Compare prevalence rates against CDC national benchmarks (e.g., 11.6% diabetes prevalence nationally)
• Verify cohort sizes are statistically meaningful for planned analyses
• Check for coding artifacts: suspiciously uniform diagnosis patterns, missing demographic segments
• Reconcile against clinical registries or EHR problem lists where available

Step 7 — Document Cohort Definitions

Produce a cohort specification document including:

• Exact ICD-10, CPT, NDC code lists used
• Look-back windows and minimum claim thresholds
• Exclusion criteria and rationale
• Date ranges and data sources
• Cohort size, demographic composition, and prevalence benchmarks

Output Specification

Cohort Report:
├── Cohort Summary Table (condition, N, prevalence, avg RAF, avg cost PMPM)
├── Severity Distribution (tier counts and percentages per condition)
├── Comorbidity Matrix (co-occurrence rates for top 10 conditions)
├── Rising Risk Segment (patients meeting escalation criteria)
├── Provider Attribution Summary (panel sizes, condition density)
├── Geographic Heat Map Data (prevalence by zip/HSA)
└── Cohort Definition Appendix (full algorithm specification)

Analysis Framework

Prevalence Benchmarking

Compare observed prevalence against:

• CDC Chronic Disease Indicators for national/state rates
• AHRQ HCUP data for inpatient condition mix
• Plan-specific historical trends (year-over-year)

Risk Stratification Matrix

Cohort Quality Metrics

• Sensitivity: Percentage of known disease patients captured
• Specificity: Rate of false-positive inclusion
• Stability: Month-over-month cohort membership churn rate (target < 5%)

Examples

Example 1 — Diabetes Cohort for MA Plan Build a Type 2 diabetes cohort for a 45,000-member Medicare Advantage plan. Apply CCW diabetes algorithm (ICD-10 E11.x, look-back 24 months, ≥ 2 outpatient claims). Layer HbA1c severity staging. Identify 5,400 diabetics (12% prevalence). Segment into controlled (58%), moderate (28%), uncontrolled (14%). Flag 756 members with diabetic CKD overlap for nephrology co-management.

Example 2 — Multi-Condition Rising Risk Across a 200,000-member commercial population, identify members with ≥ 2 chronic conditions whose cost trajectory has increased > 20% over the prior 6 months while severity indicators worsen. Produce a list of 3,200 rising-risk members for care management outreach prioritization.

Guidelines

• Always use the most current CMS-HCC model version (v28 for payment year 2025+) for risk-adjusted analyses
• Apply minimum cell-size suppression (n ≥ 11) for any patient-level or sub-group reporting per CMS cell-size policy
• Validate ICD-10 codes against current fiscal year code sets; retired codes produce false negatives
• Separate incident (new-onset) from prevalent (existing) cases when trend analysis requires it
• Document all assumptions, especially around look-back windows and claim-count thresholds

Validation Checklist

• Cohort prevalence rates are within ±20% of CDC/national benchmarks
• No single provider accounts for > 30% of cohort (attribution skew check)
• Severity tiers sum to 100% of cohort membership
• Comorbidity rates are biologically plausible (e.g., T1D + T2D co-occurrence should be near zero)
• Enrollment gaps are handled (continuous enrollment requirement documented)
• Cohort definition is fully reproducible from documented algorithm
• Minimum cell-size suppression applied to all sub-group outputs

HIPAA Compliance

This skill processes Protected Health Information (PHI). All outputs must comply with HIPAA Privacy and Security Rules. Apply minimum necessary standards, de-identify data where feasible using Safe Harbor or Expert Determination methods (45 CFR §164.514), and ensure all patient-level outputs are transmitted and stored in accordance with organizational BAA and data governance policies. Never include direct identifiers (name, SSN, MRN) in analytical outputs without explicit authorization.