eICU Collaborative Research Database

• eICU Collaborative Research Database is a large, de-identified ICU dataset comprising around 200,000 patient encounters from over 200 US hospitals, enabling center-aware analyses.
• Its relational structure integrates diverse data types—administrative, physiologic, laboratory, and treatment records—facilitating both low-dimensional risk models and high-frequency time-series analysis.
• Methodological applications range from external validation and transportability studies to federated learning and synthetic-data research, underscoring its role in multicenter evaluation.

Searching arXiv for recent and foundational papers that discuss the eICU Collaborative Research Database and its methodological uses. The eICU Collaborative Research Database (eICU-CRD) is a public, de-identified, multicenter intensive care database derived from the Philips eICU telemedicine program. Downstream methodological papers describe the public release as containing about 200,000 ICU admissions from 2014–2015 across over 200 hospitals in the continental United States, with 208 hospitals explicitly enumerated in several studies; one benchmark paper on eICU-CRD v1.0 reports 200,859 patient unit encounters for 139,367 unique patients across 335 ICUs in 208 hospitals, while a transportability study on eICU-CRD v2.0 reports 200,859 distinct unit stays (Bennett et al., 2021, Sheikhalishahi et al., 2019, Johnson et al., 2018). Its research importance is not only scale, but the fact that hospital identity is preserved strongly enough to support center-aware validation, transportability analysis, and multicenter methodological development.

1. Provenance, scale, and access

eICU-CRD is consistently described in the literature as a de-identified dataset spanning critical care units of over 200 hospitals across the continental United States and as a public resource hosted through the usual PhysioNet access pathway. In the ricu ecosystem, access to full eICU requires PhysioNet credentials, while eicu.demo provides a small demo subset without credentialed access; the same paper emphasizes that eICU is the only clearly emphasized multicenter United States dataset among the four ICU databases natively supported by ricu (Bennett et al., 2021).

Several studies restate the database scale in task-specific terms. A public benchmark paper reports eICU-CRD v1.0 as containing 200,859 patient unit encounters, 139,367 unique patients, and 335 ICUs across 208 hospitals (Sheikhalishahi et al., 2019). A later mortality transportability paper describes eICU-CRD v2.0 as containing 200,859 distinct unit stays (Johnson et al., 2018). A delirium external-validation study reports that the raw eICU source contains 200,859 ICU admissions from 208 hospitals, and that study-specific exclusions reduced this to 55,665 ICU admissions from 193 hospitals (Contreras et al., 2024).

This multicenter composition is the property most often foregrounded in methodological work. Studies repeatedly use eICU not merely as a large table collection, but as a proxy for realistic between-hospital heterogeneity, allowing explicit holdout of entire hospitals rather than only random patient-level partitioning (Johnson et al., 2018, Mehrjou et al., 2022).

2. Relational organization and temporal semantics

The database is a relational ICU source with broad coverage of administrative, physiologic, laboratory, treatment, and charted data. In ricu, eICU is exposed as a 31-table source including admissiondrug, admissiondx, allergy, apacheapsvar, apachepatientresult, apachepredvar, diagnosis, hospital, infusiondrug, intakeoutput, lab, medication, microlab, note, nurseassessment, nursecharting, patient, physicalexam, respiratorycare, respiratorycharting, treatment, vitalaperiodic, and vitalperiodic (Bennett et al., 2021). The same paper highlights patient demographics, routine vital signs, laboratory measurements, medication administrations, admission diagnoses, and treatment information as prominent content classes.

A persistent implementation feature is heterogeneous table layout. eICU contains both wide tables and long entity-attribute-value-like tables; vitalperiodic is explicitly given as a wide table, whereas many other sources are handled differently through source-specific item classes (Bennett et al., 2021). This matters for reproducibility because concept extraction in eICU often depends on column-level retrieval rather than item-ID filtering.

Temporal semantics also differ from those of other ICU databases. ricu emphasizes that eICU uses relative time referenced to admission, unlike sources with absolute timestamps, and normalizes this through a common difftime representation (Bennett et al., 2021). METRE adopts the same basic premise at the extraction level: eICU cohort identification is keyed by patientunitstayid, and time-varying data are then converted into harmonized hourly representations for cross-database modeling (Liao et al., 2023).

Identifier structure imposes a major analytical constraint. ICU admissions within a hospital stay can be linked, but patients cannot be identified across hospital stays (Bennett et al., 2021). This does not prevent within-stay longitudinal modeling, but it limits patient-level longitudinal follow-up across encounters and complicates repeated-admission analyses.

3. Cohort construction and feature extraction conventions

Published work on eICU-CRD exhibits recurring cohort-construction patterns: restriction to adults, early-stay feature windows, and explicit filtering for support or task-specific label validity. In a mortality transportability study, the source database was eICU-CRD v2.0, but the analytic cohort was reduced to 50,106 ICU stays from 46 hospitals after excluding non-ICU stays, excluding ICU stays without an APACHE IVa score, and removing hospitals with fewer than 500 stays (Johnson et al., 2018). In a large delirium external-validation study, exclusions included age under 18 years, non-first ICU admissions, ICU stay shorter than 24 hours, death within 48 hours, delirium or coma in the first 24 hours, and missing first-24-hour EHR data, leaving 55,665 ICU admissions from 193 hospitals (Contreras et al., 2024).

Feature extraction is usually organized around an early observation window. A canonical first-day mortality setup defined a fixed window of length $W = 24$ hours starting on ICU admission, with vital signs summarized over $[t_{i,w}-W, t_{i,w}]$ and blood gases and laboratory variables over $[t_{i,w}-W-24, t_{i,w}]$; 82 features were extracted using simple “first,” “last,” or “sum” summaries (Johnson et al., 2018). By contrast, the delirium study harmonized 81 predictive features common across three databases and rendered first-24-hour structured EHR into templated text, using minimum and maximum values for temporal variables and total first-24-hour dose for medications before passing the serialized report to a clinical LLM backbone (Contreras et al., 2024).

Other studies exploit finer temporal granularity. A sepsis forecasting paper used vitalPeriodic and diagnosis to construct 5-minute trajectories for mean blood pressure, heart rate, and respiration rate around sepsis or septic shock diagnosis, retaining data up to 9 hours prior to diagnosis offset and creating 6-hour input plus 3-hour forecast windows (Bhatti et al., 2023). A synthetic-data study likewise extracted four vital signs from vitalPeriodic—heart rate, respiratory rate, oxygen saturation, and mean blood pressure—at 5-minute cadence over the first 24 hours, producing sequence length $L = 276$ for 62,453 ICU stays (Tian et al., 2023).

These studies collectively show that eICU supports both low-dimensional structured risk modeling and higher-frequency time-series analysis, but the exact extraction logic remains highly task dependent.

4. Benchmarking and hospital-level generalization

A central use of eICU-CRD is as a public benchmark for ICU prediction tasks. One benchmark suite defined four tasks on eICU-CRD v1.0— in-hospital mortality, remaining length of stay, phenotyping, and physiologic decompensation—using a cohort of 73,718 adult patients with at least 15 records in their ICU admission, corresponding to 4,564,844 clinical records (Sheikhalishahi et al., 2019). That benchmark built hourly time series from patient, lab, nurseCharting, and diagnosis, and reported that a BiLSTM with entity embeddings generally performed best, especially for mortality and decompensation (Sheikhalishahi et al., 2019).

The database is also used specifically to quantify hospital-level transportability. A logistic-regression mortality study trained on pooled hospitals and evaluated on held-out hospitals rather than random patient splits. In that design, 5-fold cross-validation on the pooled training hospitals yielded AUROC $= 0.854$ and SMR $= 0.998$, while held-out hospital performance ranged from AUROC 0.688 to 0.933 and SMR 0.560 to 1.43 (Johnson et al., 2018). The same study defined the calibration statistic as

$\mathrm{SMR} = \frac{\sum_i y_i}{\sum_i \hat p_i},$

and concluded that cross-validation on a multicenter training set gave a reasonable, though optimistic, estimate of performance in new centers (Johnson et al., 2018).

That paper also examined transfer versus local retraining at a single held-out hospital. The transferred pooled model retained AUROC 0.849 after recalibration, whereas a local model trained only on that hospital’s accumulating data improved from AUROC 0.614 to 0.796 after 2,400 patients and still did not overtake the transferred model (Johnson et al., 2018). In practical terms, eICU’s multicenter structure enabled a direct empirical demonstration that pooled cross-hospital development can outperform initially local-only modeling.

5. Cross-database validation and data harmonization

eICU-CRD has become a standard external-validation target for ICU machine learning systems trained elsewhere. In DeLLiriuM, model development occurred entirely on a single University of Florida hospital, while eICU and MIMIC-IV served as out-of-distribution test sets; eICU contributed 55,665 ICU admissions from 193 hospitals and yielded AUROC 0.84 (95% confidence interval 0.83–0.85), outperforming all reported structured and text baselines in that external cohort (Contreras et al., 2024). In a sepsis-associated acute kidney injury study, eICU supplied an 8,547-patient external validation cohort, and the primary XGBoost model achieved external AUROC 0.720 (95% confidence interval 0.708–0.733), lower than the internal MIMIC-IV AUROC of 0.878 but still used as evidence of partial transportability (Chen et al., 25 Feb 2025).

The same logic motivates database-harmonization pipelines. METRE was explicitly designed to work on MIMIC-IV and eICU, extracting static, vital, and intervention tables with matched feature ordering. Under default settings, METRE extracted 126,448 ICU records from eICU and reported cross-database AUC changes that could be as small as $+0.019$ or $-0.015$, although some intervention-defined tasks transferred much worse (Liao et al., 2023). METRE also makes explicit how harmonization remains imperfect: 15 variables were not found in eICU and were filled with all NaN values to preserve dataframe shape (Liao et al., 2023).

eICU also functions as a target-population covariate distribution in transfer learning for decision rules. In a sepsis individualized treatment rule study, MIMIC-III provided covariates, treatment, and outcome, whereas eICU-CRD supplied only target-population covariates. After baseline restriction, the study used 21,995 unique patients from eICU and 13 common covariates, with calibration weighting toward eICU increasing the estimated rule value from approximately 0.7223 to approximately 0.7616 (Wang et al., 3 Jan 2025). This use is conceptually distinct from ordinary external validation: eICU is not only a test set, but the formal target population.

6. Advanced methodological uses

Because hospitals can be treated as natural clients, eICU has also become a benchmark for privacy-preserving and distributed learning. A federated learning case study treated each hospital as a client, used APACHE variables together with age from the patient table to predict survival of each ICU stay, and systematically varied the number of local epochs $E$, participation fraction $[t_{i,w}-W, t_{i,w}]$0, and batch size $[t_{i,w}-W, t_{i,w}]$1 under FedAvg (Mehrjou et al., 2022). The study concluded that larger local epoch counts often improved AUC and reduced communication cost, while small hospitals were the main source of instability (Mehrjou et al., 2022).

Synthetic-data research uses eICU to evaluate whether generative models can preserve utility while reducing privacy risk. In the TimeDiff study, eICU sequences were constructed from vitalPeriodic with four vital signs, 5-minute sampling, sequence length 276, 10.5% missingness, and mortality rate 10.63%; TimeDiff achieved discriminative score $[t_{i,w}-W, t_{i,w}]$2, predictive score $[t_{i,w}-W, t_{i,w}]$3, runtime 8.7 hours, NNAA $[t_{i,w}-W, t_{i,w}]$4, and MIR $[t_{i,w}-W, t_{i,w}]$5 on eICU (Tian et al., 2023). The paper explicitly modeled missingness as a discrete sequence rather than suppressing it.

Dynamic survival modeling has likewise been demonstrated on eICU. A respiratory-failure study validated a cumulative hazard function–based autoregressive hidden Markov model on 4,391 eICU samples using the first 24 hours of ICU physiological data, reporting AUROC values from 0.80 to 0.83 and a Day-5 AUCPR of 0.44, about 4% above the best baseline (Yin et al., 2021). These uses show that eICU is not confined to classical tabular benchmarking; it supports federated, generative, and time-varying survival formulations.

7. Limitations and interpretive caveats

The same properties that make eICU attractive also create recurring methodological caveats. First, hospital-level heterogeneity is real rather than incidental. Mortality transportability work shows substantial center-level variation in both discrimination and calibration, even when a pooled model transfers reasonably well on average (Johnson et al., 2018). Second, patients cannot be identified across hospital stays, which constrains patient-level longitudinal analyses (Bennett et al., 2021). Third, variable availability differs across hospitals, and some bedside-recorded data may still be absent because of technical collection limitations (Bennett et al., 2021).

External-validity claims also require care. Although eICU is multicenter, one mortality paper notes that all hospitals came from the same overall tele-ICU program and data vendor, so validation within eICU is not equivalent to validating on entirely different EHR systems or care networks (Johnson et al., 2018). Task labels can be especially fragile when they depend on structured assessments or interventions. In DeLLiriuM, delirium incidence in the final eICU cohort was only 2.1%, far lower than in the other two study datasets; the paper notes that this may reflect documentation practices, assessment frequency, patient mix, or missing or under-recorded CAM and RASS observations (Contreras et al., 2024). In the SA-AKI external validation study, the authors attribute the drop from internal to external performance partly to inter-database population differences, institutional practice variation, and data-collection differences (Chen et al., 25 Feb 2025).

A broader methodological implication is that eICU should not be treated as a frictionless substitute for single-center ICU data. Its value lies precisely in the fact that center effects, schema heterogeneity, temporal irregularity, and documentation variation remain analytically visible. This suggests that eICU-CRD is most informative when used for center-aware evaluation, harmonization-aware feature construction, and explicit transportability analysis rather than for indiscriminate pooled train-test splitting (Johnson et al., 2018, Liao et al., 2023).