Risk Stratification as a Data Pipeline: Turning RAF Into a Worklist

Risk stratification sounds like a clinical strategy, but for the engineers who build it, it's a ranking pipeline. You take a population, score each member's expected need, sort, and hand a care team a prioritized list. This post is about how that pipeline is actually wired in a Medicare Advantage context.

The goal, stated as a function: stratify(population) -> ranked_list_of_members_by_expected_need. The hard parts are choosing the score, making it explainable, and refreshing it on a cadence that's useful to humans.

The Inputs

A useful stratification model blends several signals. In Medicare Advantage, the natural backbone is the same data that drives funding accuracy:

• RAF (Risk Adjustment Factor) - the normalized expected-cost score built from demographics and HCCs (CMS-HCC V28).
• HCC composition - which conditions, not just the total. Two members with the same RAF can need very different interventions.
• Utilization signals - recent admissions, ED visits, polypharmacy.
• Gaps - suspected-but-undocumented conditions (a stratification model that ignores gaps under-ranks the sickest, least-documented members).

This is exactly the approach walked through in this Medicare Advantage risk adjustment case study, where risk data became a concrete outreach list.

A Minimal Scoring Sketch

def member_score(m):
    """Synthetic, illustrative weighting only."""
    score = m.raf * W_RAF
    score += m.recent_admissions * W_ADMIT
    score += len(m.open_gaps) * W_GAP
    score += m.polypharmacy_flag * W_RX
    return score

ranked = sorted(population, key=member_score, reverse=True)

The weights are policy decisions, not engineering ones - surface them in config, don't bury them in code. Your clinical leadership will want to tune them, and your auditors will want to see them.

Explainability Is Not Optional

A stratification score that a care manager can't interrogate is a score they won't trust. For every ranked member, emit the contribution breakdown:

{
    "member_id": "SYNTH-10293",
    "score": 3.8,
    "drivers": {
        "raf": 2.1,
        "recent_admissions": 1.0,
        "open_gaps": 0.5,
        "polypharmacy": 0.2
    }
}

Now "why is this member #1?" has a one-glance answer.

Refresh Cadence and Stability

• Recompute on a schedule (often monthly) and snapshot each run.
• Watch for churn. If members thrash in and out of the top tier week to week, your weights are too sensitive - care teams need stability to actually act.
• Track outcomes back to the score. The point isn't the ranking; it's whether outreach to high-ranked members changed anything.

Engineering Guardrails

• Synthetic data only in dev and test. Generate illustrative populations.
• Version the model. A stratification run should be reproducible later.
• Separate scoring from action. The pipeline ranks; humans decide. Don't auto-trigger interventions off a raw score.

The Payoff

Done well, stratification connects the backward-looking RAF to forward-looking care: instead of reconciling last year's risk, you're deciding who to reach this month. For the full data-side walkthrough that complements this pipeline view, see the Medicare Advantage risk stratification guide.

VBC Risk Analytics. Educational only - not coding, billing, or clinical advice; verify against the current CMS Rate Announcement. Synthetic data only.