Hospital KPI Tree: Patient Days, Not Length of Stay

Every Service Line Got Faster and the Hospital Got Slower

A hospital KPI tree usually breaks at the top node, and the top node is usually average length of stay. The figure is correct. It is also a weighted average, and a weighted average moves for two reasons that look identical on a dashboard.

Take one quarter, 200 staffed beds and 91 days. Two service lines carry the inpatient block. Medicine discharged 2,000 cases at 5.50 days. Surgery discharged 1,500 at 3.40 days. That is 3,500 discharges and 16,100 patient days, so hospital average length of stay was 4.60 days.

Next quarter both lines improved. Medicine ran 5.45 days, surgery ran 3.35, each 0.05 days faster. Discharges held at 3,500. Hospital average length of stay rose to 4.67.

Occupancy rose with it, from 88.46 percent to 89.81 percent on the same 18,200 available bed days. Inpatient contribution fell 752,500 dollars.

Nothing in that paragraph is a rounding artifact. Every number reconciles, and the metric everyone watched moved in the wrong direction for a reason the metric cannot express.

What Is a Hospital KPI Tree?

A hospital KPI tree decomposes an inpatient result into the summable quantities that produce it: discharges, patient days, available bed days and contribution dollars. Length of stay, occupancy and case mix index sit in the tree as derived nodes, each one a quotient of two summed columns rather than a figure stored per row.

The metrics themselves are standard and externally benchmarked. The Agency for Healthcare Research and Quality publishes national inpatient stay and cost statistics through the HCUP Fast Stats tools and its national statistics charts.¹ ² The American Hospital Association publishes bed, admission and utilization counts in its annual Fast Facts release.³ The OECD maintains hospital occupancy and average length of stay as two separate published indicators, which is itself a hint that they are not the same measurement.⁷ ⁸

What none of those sources can do is tell one hospital why its own figure moved. That is a decomposition problem, and it needs the counts underneath the ratio.

Finance teams meet the same problem outside healthcare whenever a plan is built on drivers, which is the subject of the Association for Financial Professionals guide to driver-based models.⁹ The hospital version is harder only because two of its headline metrics share a numerator and neither of them is additive.

Patient Days Are the Spine of the Tree

Three identities carry an inpatient tree, and only the first one is additive.

Patient days equal discharges multiplied by average length of stay, per service line. Available bed days equal staffed beds multiplied by days in the period. Occupancy is patient days divided by available bed days.

Store the left side of each identity, never the right. The summable columns are discharges, patient days, staffed beds, net revenue and direct cost, one row per service line per period. Length of stay, occupancy, revenue per case and contribution per case are all derived by dividing two summed columns at whatever level is being read.

The reason is mechanical. Sum the patient day column across service lines and the total is the hospital's patient days. Average the length of stay column across service lines and the result is 4.45 days, which belongs to no hospital. The correct 4.60 comes from 16,100 divided by 3,500.

That is the same failure that makes a KPI tree in Excel average a stored ratio, and it is why the tree stores days rather than the metric people quote.

Why Did Average Length of Stay Rise When No Service Line Slowed?

Because hospital length of stay is weighted by discharges, and the discharge mix moved toward the longer-stay service. Swapping 200 surgical cases for 200 medicine cases added 420 bed days on its own, since a medicine case occupied 2.10 days more than a surgical one. The two stay improvements returned 175 bed days. The net is 245 more days.

Run the bridge in bed days rather than in days of average, because bed days are the unit that can be staffed.

Second quarter volumes were 2,200 medicine discharges at 5.45 days and 1,300 surgery discharges at 3.35, giving 11,990 and 4,355 patient days, or 16,345 in total against 16,100. The hospital used 245 more bed days to treat the same 3,500 patients.

Split that into two terms. The mix term holds each line's first-quarter stay and moves only the discharge counts: 200 extra medicine cases at 5.50 days is 1,100 bed days, and 200 fewer surgical cases at 3.40 is minus 680, for plus 420. The stay term holds second-quarter counts and moves only the stay: 2,200 at minus 0.05 days is minus 110, and 1,300 at minus 0.05 is minus 65, for minus 175.

Plus 420 and minus 175 is plus 245, with no residual. The mix term is simply 200 cases multiplied by the 2.10 day gap between the two case types.

Discharges were flat, both service lines improved, and the hospital consumed 245 more bed days. A tree that stores only average length of stay reports the 0.07 day rise and can attribute none of it.

Occupancy Rose and Nothing Improved

Occupancy inherits the same numerator, so it inherits the same problem and adds one of its own.

Available bed days were 18,200 in both quarters: 200 staffed beds across 91 days. Occupancy went from 16,100 divided by 18,200, or 88.46 percent, to 16,345 divided by 18,200, or 89.81 percent. The 1.35 point gain is the same 245 bed days, restated as a percentage of a denominator that did not move.

A hospital that treated no additional patients and earned less money posted a higher utilization rate. Occupancy rose because cases got heavier, not because capacity got busier in any sense worth funding.

The denominator deserves its own scrutiny. Licensed beds, staffed beds and physically available beds are three different counts, and only staffed beds correspond to capacity a hospital can actually fill this quarter. The OECD publishes bed and occupancy series on defined bases for exactly this reason.⁷

Store staffed beds as a column with the period attached. A tree whose capacity denominator is a licensed bed count is measuring against beds nobody can open.

Where Did the 752,500 Dollars Go?

Into the case mix, not into length of stay. Two hundred discharges moved from a service contributing 6,400 dollars per case to one contributing 1,850, a swing of 910,000 dollars. The shorter stays returned 45 dollars per case across 3,500 discharges, or 157,500. The two terms sum to the reported fall exactly.

Under case-rate payment the asymmetry is structural. Medicare's inpatient prospective payment system pays a rate per discharge derived from the case's diagnosis-related group, not a rate per day.⁵ Revenue therefore accrues per case while a large share of direct cost accrues per day. A day removed from a stay is a cost saving, not a revenue loss, and a case removed from the schedule is both.

In the worked quarter medicine contributed 1,850 dollars per case and surgery 6,400. Each line saved 45 dollars of variable daily cost from its 0.05 day improvement, lifting contribution per case to 1,895 and 6,445.

The same two-term split used on bed days applies to the dollars, and both columns reconcile against the same volumes. It is the volume and rate structure of a price, volume and mix bridge, with discharges as the volume.

Case Mix Index and GMLOS Are Ratios Too

Two more hospital figures sit one layer down, and both repeat the pattern.

Case mix index is a discharge-weighted average of the relative weights assigned to each diagnosis-related group. It is the same arithmetic object as average length of stay, so it moves when the mix moves and no clinical practice changes. Store the summed relative weights and the discharge count, then derive the index. A case mix index stored per row cannot be rolled up or regrouped without recomputing from its parts.

The benchmark side carries a subtler trap. The inpatient prospective payment system tables publish a geometric mean length of stay for each diagnosis-related group alongside an arithmetic mean, and the geometric figure is the one that governs the post-acute care transfer payment calculation.⁵ ⁶ A hospital that computes its own arithmetic average and compares it to a geometric benchmark is comparing two different statistics of two different distributions.

Pick one basis, record which, and keep it on the node. Length of stay distributions are heavily right-skewed, so the two means diverge most where the cases are most expensive.

Four Tests Before You Trust a Hospital KPI Tree

1. Patient day additivity test. The sum of service line patient days must equal hospital patient days for every period, before any ratio is displayed. Aggregate agreement can conceal two lines that offset, which is the case the MECE checks exist to separate.

1. Recomputation test. Regroup the tree from service line to division and confirm each length of stay node recomputes from summed days over summed discharges rather than averaging its children. If the hospital total equals the simple average of the lines, the tree is storing the ratio.

1. Denominator basis test. Confirm the capacity denominator uses staffed beds on a stated basis, with the period length attached, and that the basis has not changed between the two quarters being compared.

1. Payment basis test. Confirm whether each service line is paid per case or per day before any stay reduction is scored as a saving or a loss. Mixed contracts need the split stored as a column.

Run test 1 first because it fails loudly. Test 4 decides whether the tree's dollars mean anything.

Where a Hospital KPI Tree Is the Wrong Instrument

Four situations defeat this decomposition.

Clinical quality questions. A tree explains why bed days moved. It cannot say whether a shorter stay was appropriate for the patient, and readmission consequences sit outside the arithmetic entirely.

Unstable service line definitions. If a service transferred between lines mid-year, the mix term is measuring a reporting change. Restate both periods on one definition first.

Small case volumes. A line with 40 discharges a quarter has a length of stay that swings on two outliers. Aggregate to a level where the count supports the ratio.

Capacity constrained by staffing, not beds. When nurse availability sets the ceiling, staffed beds are an output of rostering and the occupancy node describes the roster rather than demand. Model the roster separately.

The first case is out of scope. The other three are fixable inputs, the same discipline OEE counting requires before minutes mean anything.

Margin pressure makes the distinction expensive rather than academic, which is why operating performance is tracked monthly in reports such as the National Hospital Flash Report.⁴

Building the Tree in kpitree.io

kpitree.io builds this decomposition from one uploaded CSV. The file needs one row per service line per period and six summable columns: discharges, patient days, staffed beds, days in period, net revenue and direct cost.

Inside the tree, parent to child relationships are addition and subtraction, which is what lets medicine patient days plus surgery patient days produce hospital patient days as a real arithmetic identity rather than a layout convention. Every rate is a derived node that divides two summed columns at the level being read: patient days over discharges gives length of stay, patient days over available bed days gives occupancy, contribution over discharges gives contribution per case. Regroup from service line to division to hospital and each rate recomputes against its own denominator instead of averaging an average.

CSV upload is the evidenced ingest path today. The tree computes the arithmetic. It does not narrate it, and it will not tell you which service line to expand.

The smallest useful version is four rows. Take two service lines across two quarters, six columns, and decompose the length of stay figure your operations meeting already argues about.

Frequently Asked Questions

Why does hospital length of stay differ from the average of my service lines? Because the hospital figure weights each line by its discharge count. The simple average of 5.45 and 3.35 is 4.40, while the discharge-weighted figure is 4.67. Only the second one is the hospital's.

Should the tree store average length of stay at all? Store it as a derived node, never as a column. A stored ratio cannot be regrouped, and it cannot be bridged into mix and rate terms.

Does reducing length of stay increase revenue? Not under case-rate payment, where the rate attaches to the discharge rather than the day.⁵ It reduces variable daily cost, and it frees bed days that only convert into revenue if additional cases fill them.

Which bed count belongs in the occupancy denominator? Staffed beds, on a single stated basis, with the period length attached.⁷

Is case mix index a driver or an outcome? Both, which is why it should never sit alone. Keep the summed relative weights and the discharge count beneath it so a move can be split into the same mix and rate terms.

Closing: Store Days and Discharges

Average length of stay and occupancy are the two figures a hospital board reads, and neither one can explain itself. They share a numerator, they are both quotients, and both move when the case mix moves even though no clinician, ward or process changed.

Store six summable columns per service line per period. Derive every rate by dividing two of them at the level being read. Bridge any move into a mix term and a rate term, in bed days first and dollars second, and check that the two terms sum to the reported change with nothing left over.

Done that way, a quarter where both service lines improved and contribution fell 752,500 dollars stops being a contradiction. It becomes 200 cases that changed type, priced at the 4,550 dollar gap between them, next to a stay improvement worth 157,500 that was real and was never going to be enough.

Upload one CSV with two service lines, two quarters and six summable columns, and decompose your own length of stay in kpitree.io.

Sources

• HCUP Fast Stats Data Tools
• HCUP-US National Statistics Charts
• Fast Facts on U.S. Hospitals, 2026
• National Hospital Flash Report, April 2026 Metrics
• Acute Inpatient Prospective Payment System
• FY 2027 IPPS Final Rule Home Page
• Hospital beds and occupancy, Health at a Glance 2025
• Length of hospital stay
• FP&A Guide to Driver-based Models and Plans