Hospital-Free Days: Define Death and Readmission Before Counting

Consider this synthetic protocol scenario before database build. Your primary endpoint is hospital-free days (HFD), also called days alive and out of hospital (DAOH), within a 28-day window. A participant presents through the emergency department (ED), is an inpatient on Days 1–6, and leaves at the end of Day 6. After eight days at home, the participant is readmitted on Days 15–19, returns home on Days 20–24, and dies on Day 25. What score enters the primary analysis?

There is no answer until the endpoint conventions are specified: 0 if any death within the window overrides previous free days; -1 if death is ranked below a survivor hospitalized throughout; 13 if actual pre-death free days are retained and readmissions deducted; or 22 under an index-admission-only formula using a supplied index length of stay of six days. Death is an observed event; its effect on the score is a design choice. Define the death rule, follow-up scope, start point, facility classes, and day boundaries before programming the endpoint.

A whole-window day count measures cumulative time alive outside qualifying hospital care. That question differs from time to first hospitalization, which records when the first event occurs rather than the duration of all stays. ESCAPE (2005) used days alive out of hospital during the first six months as its primary endpoint. SALT-ED (2018) used hospital-free days to Day 28, with outcome ascertainment limited to the index hospitalization. The shared label does not make these outcomes interchangeable.

Also distinguish HFD from related constructs whose facility boundaries or event definitions differ:

• Hospital-free days / DAOH: qualifying days alive outside hospital care. Whether nursing, rehabilitation, ED, or observation days qualify depends on the definition.

• Institution-free days / home time: may exclude additional institutional settings. Auriemma and colleagues distinguish these from HFD and describe a healthy-days-at-home measure that also discounts home health or hospice services; these labels need their own definitions.

• Ventilator-free days: time alive and free from mechanical ventilation. A participant may accumulate ventilator-free days while still hospitalized; this is not a hospital-free day.

Simply subtracting hospital days from the window can reward early death: a one-day terminal admission would produce 27 under a naive 28-minus-stay formula. Schoenfeld and Bernard (2002) studied ventilator-free days by jointly considering ventilation duration and mortality, warning that interpreting the measure without mortality could mislead. Their work concerns ventilation in acute lung injury, not a universal rule for HFD. It supports examining survival and time free from care together, rather than assuming any improvement in a composite establishes a survival benefit.

Shackleton et al. (2025) reviewed 110 studies of adult critical-care randomized trials from 2000–2024, including trial registrations, protocols, and published reports. The full-text Mortality section reports these contrasting conventions:

• Assign zero for death during follow-up: 32 studies (29.1%). Previously accumulated free days are overwritten. A survivor hospitalized throughout also scores zero. The review suggests this rule pending further patient input; SALT-ED applied zero to deaths during the index hospitalization.

• Assign minus one: three studies (2.7%). Death occupies an ordered category below a survivor who remains hospitalized and scores zero. This value is a scoring convention, not a measured negative duration.

• Retain actual days alive and out of hospital before death: 16 studies (14.5%). The review notes that five of these censored follow-up at hospital discharge, so this percentage does not imply complete readmission ascertainment.

• Leave death handling unspecified: 57 studies (51.8%) in the full-text section. Two other studies used different rules, so the four categories above are not an exhaustive percentage distribution.

These figures describe reporting in adult critical care, not an optimal score or a frequency estimate for every indication. The review's abstract reports 56 studies (50.9%) with unspecified death handling, whereas its full-text Mortality section reports 57 (51.8%); the figures above follow that section. Separately, Auriemma et al. (2021) illustrate a patient who accumulates 76 HFD before dying on Day 100, retaining prior free days rather than assigning zero to every death. Retaining time before death can let a participant who dies outscore a survivor with a long admission. Zero assignment avoids that ordering for positive-score survivors but ties death with prolonged hospitalization. Trial teams should justify the clinical value judgment rather than describe either rule as universally correct.

This is an illustrative construction, not trial data. For the whole-window calculations, Day 1 is the index ED presentation date; any part of a day in hospital makes that date a hospital day, including discharge dates. The death date earns no free day. The intervals are:

• Days 1–6: six index hospital days; discharge occurs at the end of Day 6.

• Days 7–14: eight days alive at home.

• Days 15–19: five readmission days; discharge occurs at the end of Day 19.

• Days 20–24: five further days alive at home.

• Days 25–28: the death date and three subsequent dates; none contributes a free day.

The first three rows below are illustrative applications of death rules documented in the systematic review. Row D is a deliberate ascertainment failure, not a recommended endpoint. Row E applies the published index-stay construction described by SALT-ED and Auriemma et al. to an assumed six-day index length-of-stay input. It does not reproduce SALT-ED's timestamp or partial-day derivation.

Facility classification determines which dates count as free. A participant can move among an acute ward, ED, observation unit, long-term acute-care hospital (LTACH), skilled nursing facility (SNF), rehabilitation facility, and hospice during one follow-up window. Admission billing status alone may not describe the care setting required by the endpoint.

Auriemma et al. (2021) count acute hospital, ED, observation, and LTACH days as hospital days, while counting nursing-facility, rehabilitation, and inpatient hospice days as hospital-free. They argue that access to home care and the potential value of rehabilitation complicate measures that penalize every institutional day. This is a proposed definition, not a regulatory mandate.

Shackleton et al. (2025) instead suggest counting nursing-home days as hospital days. That disagreement matters even when both papers use the HFD label. Choose the setting boundary to match the trial objective, justify it, and record it in the eCRF and analysis specification. The comparison below uses Auriemma’s HFD boundary and an illustrative broader institution-free definition; the latter is not a universal definition of home time.

In the review’s full text, 34 studies (30.9%) included readmissions, 18 (16.4%) censored data collection at index discharge, and 58 (52.7%) did not specify readmission handling. Separately, 69 studies (62.7%) did not report hospital length of stay. That last figure should not be read as 62.7% omitting every component. The review recommends reporting mortality and length of stay over the same follow-up period.

Shackleton et al. recommend against calling a discharge-censored outcome HFD. An index-only score can be a reverse-coded index length-of-stay measure with a death rule, but it does not verify time alive outside hospital across the whole window. Name that limitation explicitly rather than imply complete post-discharge observation.

SALT-ED was a single-center pragmatic multiple-crossover trial of fluids in noncritically ill adults. Its primary outcome combined index in-hospital death with length of stay, and its report acknowledges index-hospitalization-only outcome ascertainment. Auriemma et al. criticize that construction for treating live discharge followed by early death or readmission like live discharge followed by sustained recovery. This is a difference in measurement scope, not evidence that the investigators accidentally overlooked a specified outcome.

For a whole-window definition, translate the endpoint into a feasible ascertainment plan. The following are operational recommendations derived from the need to observe the specified events:

• Readmissions outside the recruiting hospital: combine appropriate record linkage, participant or proxy follow-up, and record verification. A quiet site EHR is not proof of no admission elsewhere.

• All-cause or cause-specific admissions: state which count. A disease-specific definition needs reproducible classification rules and a justified review process; an adjudication charter can document those decisions.

• Overlapping admissions and transfers: deduplicate calendar dates before summing. A transfer between two hospitals on Day 12 is one hospital day under an any-part-of-day rule.

• Incomplete follow-up: distinguish verified hospital-free dates, verified hospital dates, and unknown dates. Prespecify missing-data handling; do not turn an unverified interval into free days by default.

Ninety and 28 days were the most common horizons in the 2025 review: 43 studies (39.1%) and 31 (28.2%), respectively. Other studies used shorter, longer, multiple, or unspecified horizons. ESCAPE used six months; six calendar months should not automatically be converted into exactly 180 days. Auriemma et al. support follow-up through at least Day 90, particularly in critical illness and acute lung injury, consistent with an ANZICS trial-group recommendation. This is not a cross-indication regulatory minimum.

Window length changes which deaths and readmissions enter the variable. For a synthetic survivor with 18 hospital days and no later events, an any-part-of-day count gives 10 free days in a 28-day window and 72 in a 90-day window. Under a zero-for-death rule, a death during the respective window scores zero. A longer window therefore changes both the attainable range and the observed events; it should follow the clinical question and feasible follow-up, not simply maximize apparent separation.

Lock the anchor and day boundary with the same care used for assessment windows:

• Anchor: randomization, index ED presentation, ICU admission, or discharge can imply different periods at risk. State Day 0 versus Day 1 and the final included date. An admission anchor can include time before randomization.

• Unit: choose calendar dates in a defined time zone or elapsed 24-hour periods. Specify any-part-of-day versus midnight occupancy; neither follows automatically from the phrase “hospital day.”

• Partial days: admission on June 1 and discharge on June 6 spans five date differences but touches six calendar dates. Under an any-part rule both admission and discharge dates are hospital days; do not then label discharge day free.

• Boundary cases: test same-day admission and discharge, a short ED visit, a transfer, death after discharge, and an event on the final date. A boundary change need not shift every participant by the same amount, especially with repeated admissions.

Assigning a terminal score such as 0 or -1 for death incorporates that event into the outcome definition. This can be framed as a composite-variable strategy under ICH E9(R1). That is an application of the generic framework, not an ICH endorsement of HFD or a particular death score. Retaining actual days alive outside hospital asks a different question from overwriting those days. Specify the full estimand, including the population, treatment conditions, variable, handling of relevant intercurrent events, and population-level summary. Our estimands guide explains the wider framework.

ICH E9(R1), Section A.5.1, explains that incorporating an intercurrent event into a composite can avoid assumptions about outcomes after that event, while leaving an interpretation problem. It does not remove assumptions about missing vital status or hospital records. An improvement in the overall score alone cannot identify whether treatment reduced mortality, shortened stays, or reduced readmissions; component effects can point in different directions.

Prespecify component reporting alongside the main estimate: mortality at the horizon, index length of stay, and readmission frequency or days. Interpret summaries restricted to survivors cautiously because survival may itself differ by treatment. If secondary analyses use time to discharge or hospitalization, choose methods aligned to their probability or rate question; competing-risks methods may be needed when death precludes the event. There is no single required estimator for every secondary question.

The FDA June 2019 draft on heart-failure drug-development endpoints is marked draft, not for implementation, and contains non-binding recommendations. Its heart-failure scope does not establish HFD counting conventions for other indications or guarantee acceptance of a proposed primary endpoint.

Within that scope, the draft lists distinct ways to quantify hospitalization:

1. Binary endpoints (whether or not a participant was hospitalized during follow-up);

2. Number of hospitalizations over the follow-up period;

3. Time-to-initial hospitalization (the traditional time-to-first-event construction);

4. Time alive (or at home) and out of the hospital; and

5. Time to recurrent hospitalizations.

The draft encourages discussion of all-cause versus cause-specific hospitalization. Discuss the proposed construction in the relevant development program rather than infer acceptance from the endpoint label. A 2026 editorial record by McCaw et al. confirms continuing methodological interest; its publicly available record has no abstract. Its keywords include censoring and recurrent-event methods, but they do not establish a preferred estimator.

Auriemma et al. describe bounded HFD distributions that may concentrate near the maximum and, in high-mortality populations, near zero. Bimodality is therefore a possibility to investigate, not a universal distributional law. Align the analysis and sample-size assumptions with the chosen score and expected population:

• Ordinal analysis: SALT-ED used multivariable proportional odds. This is a precedent, not an automatic choice. Justify the common-odds assumption and the interpretation of the treatment estimate for the planned score.

• Count or two-part approaches: Auriemma et al. discuss models that accommodate bounded counts or separate any positive HFD from the amount among positive scores. A zero score can include prolonged hospitalization as well as death. Simulation can assess power under plausible mortality and stay patterns.

• Clinical importance: the 2021 perspective identified MCID estimation as an evidence need. The 2025 review reports proposed MCIDs ranging from 1.0 to 7.3 days in nine of 12 primary-outcome studies. These varying values do not establish a universal threshold across indications or horizons.

Keep a single executable specification across clinical operations, data management, and statistics. The following flow illustrates whole-window follow-up after a team has selected the facility and partial-day rules. Unknown observations require their own analysis plan:

flowchart TD
  A["Define anchor, horizon, facility classes and partial-day rule"] --> B{"Vital status known through the required horizon?"}
  B -- "No" --> M["Use prespecified missing-data plan"]
  B -- "Yes" --> C{"Death in window with a prespecified override?"}
  C -- "Yes" --> D["Assign chosen terminal score: zero or minus one"]
  C -- "No" --> E["Classify eligible alive dates using records for index stay and later encounters"]
  E --> F{"Hospital status known for each relevant date?"}
  F -- "No" --> M
  F -- "Yes" --> G["Count qualifying hospital-free dates once each; exclude death date and later dates"]
  D --> H["Analyze chosen score and report mortality and hospital-use components"]
  G --> H
  M --> H

Use the convention decision table below as a pre-build verification checklist when reviewing protocol Section 3 (Objectives and Endpoints) and the statistical analysis plan (SAP) endpoint definitions locked before unblinding:

The following is an illustrative operational definition, not a validated protocol or a required convention. It deliberately uses zero for death and an any-part-of-day hospital rule. A team choosing another question should revise the whole specification consistently:

Endpoint: HFD through Day 90. Day 1 is the local calendar date of randomization; Day 90 is 89 calendar dates later. Among participants alive through the end of Day 90, count dates with no acute-care inpatient, ICU, ED, observation, or LTACH encounter at any time. Deduct qualifying encounters from any hospital, including readmissions; overlapping encounters count once. A qualifying encounter on the discharge date makes that date a hospital day, including same-day visits.

Death rule: death on any included date, in or out of hospital, assigns a total score of zero. Nursing-facility, rehabilitation, and hospice-facility stays count as free unless a separately qualifying acute-care encounter occurs on that date. These are selected conventions for this example, not universal facility rules.

Ascertainment: reconcile site records with structured participant or proxy follow-up and verification of outside encounters through Day 90. Record event dates and source coverage; keep unverified dates and unknown vital status identifiable. Specify missing-data assumptions and same-estimand sensitivity analyses in the SAP before analysis.

Analysis: select and justify the population-level treatment summary, estimator, covariates, and power assumptions separately. Report mortality, index length of stay, and readmissions alongside the primary estimate. Alternative death or facility rules define supplementary questions; under ICH E9(R1), sensitivity analysis examines assumptions for the same estimand.

Changing a death override, readmission scope, or horizon changes the endpoint question and may change its estimand. Document substantive revisions, their rationale and timing, what results were accessible, and the corresponding protocol, SAP, registry, and data-collection changes. Review the impact on bias, interpretation, and error control with the responsible trial team. The endpoint-adaptation guide discusses that wider decision; this article does not establish a blanket rule that every post-start modification has the same statistical consequence.