Collector-IQ

Portfolio: 3,847 open claims
Today's capacity: 320 touches
Days of operation: Day 87 of 120

CollectorIQ

A post-service intelligence dashboard for healthcare revenue cycle management. It answers a simple question: given a limited number of collectors, which claims should they work today to recover the most revenue?

Five views, one scoring engine:

Powered by a 9-state Markov chain, a restless multi-armed bandit (RMAB) scoring engine, and Thompson sampling on Beta-Binomial posteriors for online learning.

Claim Status Board E Powered by: 9-State Markov Chain
Backend Engine: Claim Lifecycle Model
Modulesim/04-rmab-simplified
States9 (s₀–s₈)
Transition matrixP (9×9)
X12 grounding837, 999, 277, 835
Absorbing statess₇ Paid, s₈ Written-Off
A snapshot of every open claim grouped by where it sits in its lifecycle. Each column is a transient state in the underlying 9-state Markov chain; each card is a claim with its current dwell time and one-touch lift ℓ(s).
  • s₂ Accepted — payer received the 837 and adjudication is in progress; lift is small because most touches don't accelerate this stage.
  • s₃ Pending — payer has requested additional records or information; uploading them moves the claim toward s₅.
  • s₄ Denied — payer rejected the claim with a CO-code; corrected resubmission redirects the claim from the write-off path back into adjudication. Highest lift cell.
  • s₅ Adjudicated — payment determined and awaiting posting; small lift from contract-term review.
s₂ Accepted 847 ?
CLM24891$7,846
United Healthcare
6d in state · Lift: 0.003
CLM23901$3,214
Aetna
2d in state · Lift: 0.003
CLM25102$1,890
Blue Shield
1d in state · Lift: 0.003
+ 844 more claims
s₃ Pending 312 ?
CLM22847$9,120
Cigna
14d in state · Lift: 0.008 · Records requested
CLM23156$6,430
Humana
9d in state · Lift: 0.008
CLM24033$4,871
ACME Health Plan
5d in state · Lift: 0.008
+ 309 more claims
s₄ Denied 524 ?
CLM25149$11,778
Kaiser Permanente
5d in state · Lift: 0.101 · CO-16
CLM23788$10,612
Anthem
11d in state · Lift: 0.101 · CO-197
CLM24317$10,296
United Healthcare
9d in state · Lift: 0.101 · CO-4
+ 521 more claims
s₅ Adjudicated 198 ?
CLM21903$8,340
Blue Shield
3d in state · Lift: 0.016 · Payment pending
CLM24567$5,210
Aetna
7d in state · Lift: 0.016
CLM22891$3,775
Cigna
12d in state · Lift: 0.016
+ 195 more claims

Today's Worklist

Claims ranked by expected dollar gain per collector touch E Powered by: RMAB Scoring Engine
Backend Engine: One-Touch Lift Scoring
Modulesim/04-rmab-simplified
Score formulaScore = C × ℓ(s)
Lift ℓ(s)B⁽¹⁾(s,s₇) − B⁽⁰⁾(s,s₇)
Passive kernelP⁽⁰⁾ — no intervention
Active kernelP⁽¹⁾ — one collector touch
Capacity constraintK = 320 touches/day
320
Touches available ?
$187,420
Total expected gain ?
Today's prioritized action plan. Every open claim is scored by expected dollar gain per collector touch; the top K=320 are surfaced for work. The score combines what a claim is worth with how much a single touch is expected to move it toward payment.
  • Score = Contracted × Lift — lift ℓ(s) is the increase in payment probability from one collector touch, computed from the dual-kernel absorption matrix.
  • Recommended Action — state-specific: resubmit corrected claim (denials), upload requested records (pending), review contract terms (adjudicated).
  • Why? — click the link on any row to see the math behind its rank, the state-specific lift explanation, and the backend module reference.
  • Total expected gain — sum of scores across the top 320 claims; the dollar figure captured if every recommended touch lands as expected.
Rank? Claim? Patient? Payer? State? Age? ? Contracted ? Score Recommended Action?
1 CLM25149 Rodriguez, Sofia Kaiser Permanente s₄ Denied 5d $11,778 $1,194 Resubmit corrected claim Why?
2 CLM23788 Williams, Marcus Anthem s₄ Denied 11d $10,612 $1,075 Resubmit corrected claim Why?
3 CLM24317 Patel, Anika United Healthcare s₄ Denied 9d $10,296 $1,043 Resubmit corrected claim Why?
4 CLM24513 Johnson, Terrence Cigna s₄ Denied 8d $9,794 $993 Resubmit corrected claim Why?
5 CLM25031 Nakamura, Yuki ACME Health Plan s₄ Denied 5d $9,630 $976 Resubmit corrected claim Why?
6 CLM22847 O'Brien, Kathleen Cigna s₃ Pending 14d $9,120 $73 Upload requested records Why?
7 CLM23384 Torres, Miguel Humana s₄ Denied 13d $8,642 $876 Resubmit corrected claim Why?
8 CLM25270 Davis, Angela Blue Shield s₄ Denied 4d $8,522 $864 Resubmit corrected claim Why?
9 CLM21903 Kim, Joon-ho Blue Shield s₅ Adjudicated 3d $8,340 $133 Review contract terms Why?
10 CLM21179 Fischer, Emma Aetna s₄ Denied 24d $8,028 $814 Resubmit corrected claim Why?
HFMA MAP Keys Scorecard E Powered by: Portfolio Simulation
Backend Engine: MAP Keys Aggregation
Modulesim/04-rmab-simplified
AR-1, AR-3Claim age >90d / portfolio totals
AR-5s₄ absorptions / remittances
AR-6s₄→s₈ dollars / NPSR
FM-1Expected time from N = (I−Q)⁻¹
FM-2Cumulative s₇ dollars / NPSR
FM-6, FM-7Touches × cost / dollars recovered
HFMA Management Accountability Program (MAP) keys, computed from the live portfolio simulation. Each gauge shows current value, the published HFMA target threshold, and trend versus Day 60. Green / amber / red reflect distance from target.
  • Account Resolution (AR) — aging mix and denial outcomes. AR-1 / AR-3 measure aged A/R as a share of billed and total A/R; AR-5 is the remittance denial rate; AR-6 is denial write-offs as a share of NPSR.
  • Financial Management (FM) — cash conversion and operating cost. FM-1 is net days in A/R from the fundamental matrix; FM-2 is cash collected as a share of NPSR; FM-6 / FM-7 are total and per-area cost-to-collect.
  • Trend arrow — movement since Day 60 of the simulation horizon. ▲ improving, ▼ declining, ● stable.
Account Resolution (AR)?
AR-1?
Aged A/R % Billed A/R
28.4%
0%Target <25%50%
▲ 2.1 pts from Day 60
AR-5?
Remittance Denial Rate
12.7%
0%Target <8%25%
▲ 1.8 pts from Day 60
AR-6?
Denial Write-Offs % NPSR
2.1%
0%Target <3%8%
▲ 0.6 pts from Day 60
AR-3?
Aged A/R % Total A/R
34.2%
0%Target <30%50%
▼ 0.9 pts from Day 60
Financial Management (FM)?
FM-1?
Net Days in A/R
42.3
0Target <4075
▲ 3.1 days from Day 60
FM-2?
Cash Collection % NPSR
96.8%
80%Target >96%100%
● Stable
FM-6?
Cost to Collect
3.8%
0%Target <4%8%
▲ 0.4 pts from Day 60
FM-7?
Cost to Collect by Area
Billing2.1%
Follow-up4.7%
Appeals6.2%
▲ Overall improving
Recovery Intelligence E Powered by: Dual-Policy Simulation
Backend Engine: RMAB vs. Baseline Comparison
Modulesim/04-rmab-simplified
RMAB policyTop-K by Score = C × ℓ(s)
Baseline policyTop-K by claim age (oldest first)
Portfolio2,000 initial + 200/day arrivals
HorizonT = 120 business days
Shared RNG seedSame arrivals, same draws
Side-by-side comparison of the Collector-IQ policy against an aging-only baseline run on the same claim stream with a shared random seed. The only difference between the two runs is which K=320 claims receive active-kernel transitions each day.
  • Cumulative Recovery — total dollars resolved over the 120-day horizon. The gap to baseline is the dollar value the policy delivers.
  • Write-Offs Averted — claims that would have absorbed into s₈ (Written-Off) under baseline but were redirected to s₇ (Paid) under Collector-IQ.
  • Appeal Opportunity Queue — denied claims ranked by recoverable value × resubmission lift, with filing deadlines flagged.
  • Cost to Collect — FM-6 / FM-7 broken out overall, per payer, and per collector. Identifies where touches are expensive relative to dollars recovered.

Cumulative Recovery: Collector-IQ vs. Baseline?

$80M $60M $40M $20M $0 Day 0 Day 60 Day 120 Collector-IQ Aging Baseline +$5.9M

Write-Offs Averted?

5,000 3,750 2,500 1,250 0 Day 0 Day 60 Day 120 1,237 fewer Collector-IQ Aging Baseline

Appeal Opportunity Queue?

Denied claims ranked by recoverable value × resubmission lift
CLM25149 Rodriguez, Sofia $11,778 CO-16: Lack of information Filing deadline: 12 days
CLM23788 Williams, Marcus $10,612 CO-197: Auth not obtained Filing deadline: 8 days
CLM24317 Patel, Anika $10,296 CO-4: Modifier inconsistency Filing deadline: 45 days
CLM24513 Johnson, Terrence $9,794 CO-16: Lack of information Filing deadline: 38 days
CLM25031 Nakamura, Yuki $9,630 CO-197: Auth not obtained Filing deadline: 52 days

Cost to Collect (FM-6 / FM-7)?

Overall ?
3.8%
Target: <4.0%
Per Payer Avg ?
4.2%
Range: 2.1% – 6.8%
Per Collector ?
3.4%
Range: 2.8% – 5.1%

By Payer

ACME Health Plan 2.1%
Blue Shield 3.5%
United Healthcare 4.3%
Cigna 5.3%
Kaiser Permanente 6.8%
Learning Dashboard E Powered by: Thompson Sampling Posterior
Backend Engine: Beta-Binomial Posterior Updates
Modulesim/04-rmab-simplified (planned)
Conjugate priorBeta(1, 1) per cell
Stratificationstate × payer × reason_code
Exploration policyThompson draw on each ranking pass
Holdout fraction5% of would-be-touched
How the system is learning the transition probabilities that drive every score on the Worklist. Each stratified cell (state × payer × reason_code) has a Beta-Binomial posterior on its progress probability α. Touches are allocated across exploit, explore, and holdout slices so posteriors tighten over time without abandoning return.
  • Value of Information — incremental recovery from re-scoring under today's posterior versus the day-0 prior. The dollar amount the learning program has delivered.
  • Capacity Allocation — today's K=320 split into exploit (top-K under posterior mean), explore (Thompson draws that surface wide-CI cells), and holdout (random passive controls).
  • Posterior Estimates by Cell — per-cell α̂ with 95% credible interval, sorted by uncertainty width descending. The widest rows are where today's exploration budget is going.
  • Convergence Trajectory — selected high-volume cells over time. The credible band narrows as samples accumulate; the line approaches the true α.
  • Coverage Heatmap — sample count per cell on a log color scale. Gray cells are blind spots the policy has never touched — the diagnostic for systematic exploration gaps.
  • Active Passive Holdouts — claims that would have been touched but were randomly assigned to passive observation. These pin the no-touch baseline B⁽⁰⁾, without which lift cannot be measured.
Value of Information?
+$184,720
Incremental recovery vs. day-0 prior · Day 87
Capacity Allocation?
Today (320 touches)
Cumulative (Day 1–87, 26,847 touches)
Exploit Explore Holdout
Posterior Estimates by Cell — sorted by uncertainty?
State? Payer? Reason? ? Touches (n) ? Progressed (k) ? α̂ 95% credible interval · prior? Status?

Convergence Trajectory?

Posterior mean α̂ with 95% credible band · selected high-volume cells

Coverage Heatmap?

Touch samples per cell · gray = unobserved · hover for n
0 1–9 10–99 100–299 300+
Active Passive Holdouts — controls for measuring B⁽⁰⁾?
Claim? Cell? ? Day in holdout ? Contracted Filing deadline? Outcome?
Lift is B⁽¹⁾ − B⁽⁰⁾. Holdouts are claims that would have been touched under the policy but were randomly assigned to passive observation. Carve-outs apply: filing deadline < 30 days excluded.
Learning Algorithm
A user-friendly walkthrough of Thompson sampling, the algorithm Collector-IQ uses to learn the per-cell touch transition probabilities α from observed outcomes. The full mathematical treatment is in thompson-sampling.md; this page covers the intuition.
  • The problem — explore vs. exploit, the basic dilemma.
  • What we're learning — α, the touch transition probability.
  • Beta distributions — how we represent uncertainty.
  • The algorithm — four lines that do the work.
  • Worked example — UHC vs. Kaiser, side by side.
  • Why not greedy — the failure mode this avoids.
  • Daily pipeline — how it runs each business day.

1. The problem we're solving

Every business day we have 320 collector touches to allocate across 3,847 open claims. Some claims are slam dunks — high contracted dollars, denial reason we know how to fix, payer that responds well to resubmission. Others are uncertain — small sample size, unclear whether intervention pays off.

This is the classic explore vs. exploit dilemma:

  • Exploit — spend the budget on cells we already know recover well. Safe, but we never learn whether unfamiliar cells might recover even better.
  • Explore — spend some budget on uncertain cells to gather data. Informative, but every exploration touch is a touch not spent on a known winner.

Thompson sampling resolves this tradeoff automatically and in proportion to uncertainty — without an ε knob to tune, without ignoring what we don't know.

2. What we're learning: α (the touch transition probability)

Each business day, every open claim either makes progress or it doesn't. Some progress happens on its own — the payer adjudicates a pending claim, an accepted claim moves into payment. Other progress requires a collector to step in — a denial gets corrected and resubmitted, records get uploaded for a pending request.

α answers one specific question:

α = if a claim was going to stay stuck in its current state today, what's the probability that a collector touch flips it into progress instead?

That's it. α is a number between 0 and 1, and it depends on the cell.

A concrete example: s₄ Denied

Take a denied claim. On any given day, untouched, it has these natural transition probabilities (the passive kernel, hardcoded in the simulation engine):

Passive — no touch
0.68 stays denied s₄ s₆ 0.20 progresses remaining 0.12 to paid / written-off (not shown)
Active — one touch (α = 0.60)
0.27 ↓ was 0.68 s₄ s₆ 0.61 ↑ was 0.20 paid / written-off probabilities unchanged

The touch doesn't add new edges or change the paid / written-off rates. It shifts probability mass from "stuck" to "progress", by an amount equal to α times the dwell probability. Row-by-row:

Outcome for s₄No touchWith touchChange
Stays denied (s₄)0.680.27−0.41
Progresses (s₆)0.200.61+0.41
Paid (s₇)0.020.02
Written off (s₈)0.100.10

The arithmetic: with α = 0.60 and dwell = 0.68, the touch redirects 0.60 × 0.68 = 0.41 of probability mass. That 0.41 comes off the self-loop and lands on the progress edge. So the size of α literally is the size of the redirection (as a fraction of dwell), which is why we call it the touch transition probability.

Why α varies by cell

Different claims behave differently under a touch. α depends on three things:

  • State — touching a pending claim (uploading records) does something different than touching a denied claim (resubmitting). Pending and denied have different α.
  • Payer — different insurance companies respond to resubmission and inquiry at different rates. Cigna, UHC, Kaiser are not the same.
  • Reason code (for denials) — CO-16 (lack of information) is fixable by sending the missing record; CO-50 (medical necessity) often needs an entire appeal. Same state, very different α.

Each (state, payer, reason_code) combination is a cell. Collector-IQ keeps one α estimate per cell, learned from observed touch outcomes.

Today the system ships with pooled engineering priors — α = 0.10 for s₂, 0.50 for s₃, 0.60 for s₄ and s₅, identical across all payers and reason codes. That's a launch fudge. Real Cigna CO-16 denials don't behave like real Anthem CO-197 denials. Replacing pooled priors with cell-specific learned values is the entire point of the learning program.

3. Representing uncertainty: the Beta distribution

How do you talk about a quantity you don't fully know? Not with a single number — with a probability distribution over what you think the value might be. The right distribution for binary outcomes (touch worked / touch didn't) is the Beta distribution.

A Beta distribution has two shape parameters, written Beta(a, b):

  • a−1 ≈ the count of touches you've seen succeed (claim progressed).
  • b−1 ≈ the count of touches you've seen fail (claim stayed put).
  • Mean: a / (a + b) — your best point estimate.
  • As n = a + b grows, the distribution narrows.

Three illustrative shapes:

Wide distribution = uncertain. Narrow distribution = confident. Each cell on the Learning Dashboard's posterior table is a Beta distribution like one of these.

4. The algorithm in four lines

Thompson sampling is one rule, applied each day:

For every cell, draw a single random sample from its Beta distribution. Call it α̃ ("alpha-tilde"). This is the cell's "candidate α for today."
For every claim, compute Score = Contracted Amount × α̃ using its cell's draw.
Rank claims by score and work the top K=320.
After Δt days, observe which touched claims actually progressed. For each cell, increment a by the count that progressed and increment b by the count that didn't. The Beta posterior tightens.

That's it. The exploration happens automatically: cells with wide distributions occasionally produce high random draws that push their claims into the top K, even when their mean isn't the highest. Cells with narrow distributions produce predictable draws that act like exploitation.

5. Worked example: UHC vs. Kaiser

Two cells from today's posterior table. Both are s₄ Denied with reason code CO-16. Suppose both have a $10,000 claim awaiting work and we have one open touch slot.

UHC · CO-16
Beta(281, 133)
n = 412 k = 280 α̂ = 0.679
Tight bell around 0.68. A draw is almost always between 0.63 and 0.73. Reliable.
Kaiser · CO-16
Beta(12, 4)
n = 14 k = 11 α̂ = 0.750
Wide hump from ~0.50 to ~0.93. A draw could be 0.55, could be 0.88. Volatile.

Kaiser's empirical mean (0.75) is higher than UHC's (0.68). A greedy policy would always pick Kaiser. Thompson doesn't.

On any given day, both cells produce a random α draw. UHC's draws cluster tightly. Kaiser's draws spread widely. Most days, UHC wins because its draws are reliably high. Some days, Kaiser draws above 0.85 and wins — in proportion to how plausible it is that Kaiser's true α really exceeds UHC's. Each Kaiser win produces a new touch outcome and tightens its posterior.

After enough touches, Kaiser's distribution narrows around its true value — whatever that turns out to be. If it really is 0.78, Thompson plays Kaiser preferentially. If it's actually 0.55, Thompson abandons Kaiser. Either way, the policy self-corrects.

6. Why not just pick the highest mean?

The simplest policy — "always pick the cell with the highest empirical mean" — has a clear failure mode. Imagine a brand-new cell that has been touched twice and progressed both times. Empirical mean: 1.0. A greedy policy would pour the entire 320-touch budget into this cell. But two samples is two samples — the true α might be 0.4, and we'd waste the day.

Three approaches handle this differently:

Greedy
Always picks the highest mean. Locks in fast. Failure mode: commits to high-variance cells based on lucky early draws; never recovers.
ε-greedy
(1−ε) of the time picks the highest mean; ε of the time picks uniformly at random. Better than greedy. Failure mode: exploration is uniform — spends as much budget on obviously-bad cells as on uncertain-but-promising cells.
Thompson
Each round, draws one random sample per cell from its belief distribution. Cells with wider distributions explore more; cells with narrower distributions exploit more. Exploration is proportional to plausibility of being optimal. No ε knob.

7. How it runs each business day

The full pipeline as a single loop:

Each open claim looks up its cell (state, payer, reason_code).
Sample α̃ from the cell's Beta(a, b) posterior.
Compute Score = contracted amount × α̃.
Sort claims descending by score, take the top 320 → today's Worklist.
Carve out 5% of the would-be-touched claims as passive holdouts (no touch). These pin the no-touch baseline B⁽⁰⁾.
Touch the remaining claims. Collectors do their work.
After Δt days, observe outcomes. Did each touched claim progress? For each cell, update a and b by the counts.
Repeat tomorrow with updated posteriors.
Today's Worklist is the result of steps 1–4. The Learning Dashboard exposes the state of every cell's (a, b) pair, the convergence trajectories as they tighten over time, the coverage map showing where touches concentrate, and the cumulative dollar value the learning has produced over the day-0 prior.

Further reading

  • thompson-sampling.md — full mathematical treatment, including extensions, pitfalls, and theoretical regret bounds.
  • learning-dashboard.md — design rationale and implementation plan for the Learning Dashboard panel.
  • Russo, Van Roy, Kazerouni, Osband, Wen (2018) — A Tutorial on Thompson Sampling. Free PDF on arXiv.
  • Lattimore & Szepesvári (2020) — Bandit Algorithms. Free PDF from the authors. Chapter 36.