| Vose Software

Industry: Healthcare and Epidemiology
Product: ModelRisk
Application: Disease progression


Twenty-Year Markov Model of Chronic Kidney Disease: Where the Lifetime Cost Really Hides

End-stage renal disease costs the US payer system roughly $95,000 per patient per year for in-centre dialysis and about $145,000 for a transplant in the first year, dropping to $32,000 per year post-transplant. CKD-3a, the entry point of clinical CKD, costs less than $9,000 per patient per year. The whole game in chronic kidney disease economics is where a cohort lives in the state space over 20 years — how much time it spends in the cheap early stages versus the brutally expensive renal-replacement compartments. That trajectory is a Markov chain, not a single number, and the chart below is the actuarial picture of it.

Mean state occupancy over 20 years — CKD cohort

A regional integrated-care payer modeled a 5,000-patient cohort entering at CKD-3a, with six states (3a, 3b, 4, dialysis, transplant, death) and 20 one-year cycles. Each annual transition probability was itself drawn from a Beta distribution to reflect the uncertainty inherent in calibrating against published USRDS and KDIGO cohort data. ModelRisk turned a deterministic CKD progression table into a stochastic Markov model that quantified the lifetime-cost distribution per patient and let the payer compare interventions on an actuarial footing.

Stochastic transition probabilities, not point estimates

A point-estimate Markov table is the standard health-economic representation: 8% per year 3a → 3b, 12% per year 3b → 4, 22% per year 4 → 5 (dialysis), and so on. That table produces one expected trajectory. The trouble is that every entry has a confidence interval — many of them quite wide — and the lifetime cost is non-linear in those probabilities.

In ModelRisk each annual transition probability was modeled as Beta with an effective sample size of 200, anchored on the point estimate. That gives a 90% CI of roughly ±3 percentage points on each probability — enough to move the cohort's renal-replacement exposure across a meaningful range over the 20-year horizon. The Beta prior is the right choice for a probability: bounded on [0, 1], conjugate, and matched to the actual literature sources from which the means came.

Per simulation iteration, the engine:

  1. Draws an annual transition probability for each of the ten transition arcs from its Beta prior.
  2. Runs a 300-patient sub-cohort through 20 one-year cycles using those draws.
  3. Records mean cost per patient, year of death, and state occupancy per year.

Five thousand iterations produce the lifetime-cost distribution.

Where the cohort lives, year by year

The stacked-area chart above is the actuarial picture of a CKD cohort. Year 1: everyone in CKD-3a. By year 10: about 52% are still in early-stage CKD (3a/3b), roughly 5% are on dialysis or post-transplant, and about 36% have died. By year 20: about 19% remain in early-stage CKD, around 8% are on dialysis or post-transplant — the high-cost compartments — and 69% have died. That 8% in renal replacement therapy at the 20-year mark is small in headcount but disproportionate in spend, because a dialysis-year costs more than ten early-CKD-years.

Lifetime cost — the distribution, not the number

Lifetime cost per CKD patient — 20-year horizon

The deterministic table said mean lifetime cost per patient was around $135,000. The Monte Carlo distribution says the mean of cohort means is closer to $213,000, with a 90th percentile cohort mean of $248,000 and a 99th percentile of $277,000. For a payer with 5,000 such patients on the books, that lifts the central 20-year reserve estimate to roughly $1.06 billion, with a P90 estimate near $1.24 billion — about $175 million of difference between budgeting to the mean and budgeting to the P90. The probability that a randomly drawn cohort's mean per-patient cost exceeds $250k is roughly 8% — a tail the deterministic model is structurally unable to see.

What actually drives the lifetime cost

What drives mean lifetime cost per CKD patient

Two drivers dominate. The CKD-4 → dialysis transition probability is the single largest, moving mean lifetime cost by about ±$7k under its plausible range, with dialysis annual cost (the per-state cost parameter) close behind at about ±$4k. The dialysis-to-transplant rate is a distant third at about ±$2k — counterintuitive, because transplant is the "good outcome," but actuarially it only matters once a patient is already on dialysis, by which point most of the cost has been incurred.

The payer's response to the ranking was unambiguous: invest in delaying CKD-4 → dialysis transitions in early-stage patients, where unit cost of intervention is low and the lifetime cost saved per averted transition is high.

The SGLT2-inhibitor scenario

Recent trials of SGLT2 inhibitors (empagliflozin, dapagliflozin) have shown roughly 30% reductions in CKD progression rate in diabetic and non-diabetic CKD-3 populations. Modeling that intervention as a 30% reduction in the 3a → 3b and 3b → 4 annual probabilities:

SGLT2-inhibitor adoption — 30% progression-rate reduction at early CKD

Mean lifetime cost per patient falls from about $213k to about $182k — a $31k average saving per patient over the 20-year horizon, with the P90 dropping from $250k to $212k. At a 5,000-patient cohort that is roughly $155M of 20-year savings against an annual drug cost of about $5,000/patient ($500M over 20 years if every patient stayed on therapy). The actuarial case is positive only when adherence and durability of the progression-rate reduction are factored in — and those parameters are exactly what the next simulation cycle priced explicitly.

What changed

  • Reserves repriced to P90, not mean — 20-year CKD reserves raised from a deterministic ~$675M basis to about $1.24B at P90, eliminating a structural underprovision that had been growing year over year.
  • SGLT2 adoption funded through pharmacy benefit — the simulator paid back the drug-cost investment under adherence > 65% across patient years 1–10.
  • Care-management targeted at CKD-3b cohort — the tornado identified 3a/3b → 4 progression as where the highest-leverage cost is averted, redirecting nurse-care-manager hours from CKD-4 to CKD-3b.
  • Outcomes contracts repriced — capitated arrangements with nephrology groups moved from cohort means to risk-adjusted bands that recognized the upper-decile cost exposure.

ModelRisk functionality used

  • Annual cycle Markov chain with six states (3a, 3b, 4, dialysis, transplant, death) implemented as a transition matrix per simulated cohort.
  • Beta priors on every transition probability with effective sample size 200 — bounded on [0, 1], appropriate for probabilities, with parameter-uncertainty width matched to the published literature confidence intervals.
  • 5,000 outer Monte Carlo iterations × 300 inner patients × 20 years — the two-level structure separates parameter uncertainty (which Beta is drawn) from individual-trajectory variability (which patient does what).
  • Per-state cost vector with first-year-of-transplant cost differentiated from steady-state post-transplant cost — the kind of detail that fits naturally in an Excel table and is one of the harder things to maintain in standalone Markov code.
  • Tornado ranking that placed dialysis-cost and CKD-4 → dialysis progression as the top two cost drivers, redirecting care-management investment to where it actually buys cost reduction.
  • Intervention rerun with SGLT2-inhibitor progression-rate multipliers, producing a comparable lifetime-cost distribution against standard care for the formulary committee.

A point-estimate Markov table tells the payer what the average patient costs. The Monte Carlo rebuild tells them what the distribution of cohort costs looks like, where the cost actually sits in the state space, and which intervention bends the curve the most — three answers the deterministic version cannot give.