| Vose Software

Industry: Healthcare and Epidemiology
Product: ModelRisk
Application: Hospital Resource Allocation — Stochastic LP for Beds, OR-Time, and Nurses


A 480-Bed Hospital Allocating Beds, OR-Time, and Nurses Under Stochastic Demand

The hospital's annual operating-plan spreadsheet — built on a deterministic linear program — said the 480-bed tertiary facility would run at 88% utilization and comfortably within capacity. The same exposure data fed through a Monte Carlo simulation said census exceeded 480 beds on roughly 3% of days, and the ICU spilled over its 28 staffed beds on 10% of days. The deterministic LP is a beautiful answer to the wrong question — it allocates against the average day. Hospitals do not live on the average day; they live on the joint draw of a flu spike, a multi-vehicle trauma, and a weekend staffing dip, and any plan that does not price that joint draw is mis-pricing the resource base.

Daily inpatient census — 480-bed tertiary hospital

The deterministic LP optimizes against the green line — the mean census of 420. The simulation says the 95th-percentile day sits at roughly 473 patients, and the 99th-percentile day at roughly 497 patients (a 17-bed overflow that must be absorbed by hallway boarding, transfer-out, or elective cancellation). The probability of any-day overflow is ~3% — roughly one day in a month, on average, the hospital is over its staffed bed count, a tail the deterministic plan does not show at all.

A 480-bed tertiary hospital with five service lines (ICU, Cardiac, Surgical, Medical, Maternity) rebuilt its allocation model in ModelRisk. The deterministic LP still runs nightly — it remains the right tool for the daily duty-roster — but its constraints are now stress-tested against a 100,000-iteration demand simulation before the annual capacity plan is signed.

Service-line census is over-dispersed, not Poisson

Three years of hospital census data show variance-to-mean ratios between 1.0 (Maternity) and 2.6 (Surgical). Treating each service's daily census as Poisson — which the deterministic LP implicitly does whenever it uses a single mean — understates the tail. The simulation uses a Negative Binomial per service, calibrated to the historical mean and variance:

  • ICU — mean census 22, P99 ≈ 35 (well above the 28 staffed beds)
  • Cardiac — mean 78, P99 ≈ 113
  • Surgical — mean 125, P99 ≈ 170
  • Medical — mean 155, P99 ≈ 204
  • Maternity — mean 40, P99 ≈ 55

These tails are not academic. The simulation says the total inpatient demand reaches 497 patients on the worst 1% of days — 17 above the staffed bed count — even though the mean total is 420 and the deterministic plan would call that day "well within capacity."

What drives the overflow

Tornado — drivers of daily hospital-overflow probability

The single biggest mover is Medical service mean LOS: shaving the average length of stay from 5.0 to 4.2 days drops the overflow probability by ~2.6 percentage points — a large relative move against a ~3% baseline. Total staffed beds is second — opening 10 additional beds buys roughly 2 percentage points. Surgical demand surge sits third, ahead of ICU mean census. That ranking is what told the operations COO that the highest-leverage spend was a discharge-acceleration program for the Medical service, not a capital project to add ICU beds.

Three allocation strategies, one demand stream

The simulation re-evaluates allocation under common random numbers — every strategy is tested against the same 100,000 demand draws, so the comparison reflects strategy difference, not sampling noise.

Three allocation strategies — daily patients without a designated bed

Strategy A (today) leaves ~24 patients per day at the 95th percentile without a designated service-line bed. Flexing 8 beds from Surgical to Medical (Strategy B) barely moves the P95 — it actually edges it up to ~26, because Surgical demand has a fat tail of its own that the flex exposes. Opening 12 step-down beds (Strategy C) cuts the P95 shortage to about 12 patients/day and the P99 from 39 to 27. The step-down approach is more expensive at standing capacity but materially better in the tail, which is exactly where overflow drives the worst patient-safety incidents.

A flex-staffing rule for ICU

The same simulation evaluated a forecast-triggered flex-staffing rule for the ICU: when the next-day forecast exceeds 30 patients, open up to 6 additional ICU beds with a pre-credentialed flex-pool nurse roster.

ICU overflow days per year — before vs after flex-staffing rule

ICU overflow days fall from a mean of ~37 days/year to ~17 days/year, and the probability of exceeding the 20-day quality-of-care threshold drops from roughly 100% to about 21%. The marginal cost of the flex roster is $540,000/year; the alternative — staffing six additional ICU beds permanently — was costed at $2.1M/year and produced essentially the same risk-reduction profile. The simulation paid for itself in the first quarter of the next budget cycle.

What the model changed

  • Discharge-acceleration program for Medical service funded as the highest-ROI item in the operating plan, targeting a 0.8-day LOS reduction worth ~2.6 percentage points of daily overflow probability.
  • Step-down unit (12 beds) approved ahead of additional ICU build, because the simulation showed step-down beds catch the right tail of Cardiac+ICU joint demand more efficiently per dollar.
  • ICU flex-staffing rule replaces a proposed permanent ICU expansion, saving an estimated $1.6M/year while delivering equivalent overflow reduction.
  • OR scheduling rule tightened. Elective add-ons on days with predicted inpatient census > 460 are deferred — the simulation showed this rule lowers same-day cancellation cost while not affecting the 4-week elective-throughput target.
  • Capacity-plan filings to the regional health authority now cite simulated P95/P99 census and per-service shortage metrics rather than annual averages.

ModelRisk Functionality Used

  • Per-service Negative Binomial census models capturing the 1.0×–2.6× over-dispersion the historical data shows — replacing the implicit Poisson assumption in the deterministic LP.
  • 100,000-iteration aggregation of service-line draws to total daily census, producing the full census distribution rather than a single utilization figure.
  • Scenario engine under common random numbers so the three allocation strategies are compared on identical demand streams, isolating strategy effect from sampling noise.
  • Tornado ranking that placed Medical LOS and total staffed beds above ICU mean census, redirecting capital spend away from an ICU build to a discharge-acceleration program.
  • Forecast-conditioned flex-staffing simulation for the ICU, quantifying the trade-off between $540K/year flex pool and $2.1M/year permanent staffing for equivalent overflow reduction.
  • Annual aggregation by re-sampling daily simulation draws into 4,000 simulated years to get the overflow-days-per-year distribution used in the quality-of-care threshold comparison.

A hospital running at the deterministic mean is a hospital one bad week from a capacity crisis. ModelRisk turns that hidden tail into a number the COO can buy down with the cheapest available unit of capacity — which is rarely the unit the deterministic LP would have spent on first.