ELDER-ICU-DB: A Comprehensive Critical Care Data Resource for International Multicenter Study on Elderly ICU Patients

@article{PhysioNet-elder-icu-db-1.0.0,
  author = {Hao, Sicheng and Rodemund, Niklas and Liu, Xiaoli},
  title = {{ELDER-ICU-DB: A Comprehensive Critical Care Data Resource for International Multicenter Study on Elderly ICU Patients}},
  journal = {{PhysioNet}},
  year = {2026},
  month = oct,
  note = {Version 1.0.0},
  doi = {10.13026/126e-mx94},
  url = {https://doi.org/10.13026/126e-mx94}
}

@article{pollard_physionet_2026, title = {{PhysioNet} as a global platform for biomedical research}, volume = {1}, issn = {3005-0693}, url = {https://doi.org/10.1038/s44360-026-00096-z}, doi = {10.1038/s44360-026-00096-z}, abstract = {Twenty-five years after its first public repository was established, the PhysioNet resource for biomedical and clinical data has enabled tens of thousands of research papers. Further efforts are ongoing to advance secure data sharing, community engagement and scalable research infrastructure.}, number = {8}, journal = {Nature Health}, author = {Pollard, Tom and Moody, Benjamin E. and Lehman, Li-wei H. and Gow, Brian J. and Fernandes, Chrystinne and Xie, Chen and Johnson, Alistair and Mark, Roger G. and Heldt, Thomas}, month = aug, year = {2026}, pages = {792--795},}

Abstract

Critically ill elderly patients, characterized by high mortality and complex comorbidities, pose growing challenges to ICUs while consuming substantial healthcare resources. Current clinical research on this vulnerable population predominantly relies on single-center designs, significantly restricting the generalizability of findings. To address this gap, we present ELDER-ICU-DB, a multinational intensive care dataset specifically tailored for elderly patients, designed to facilitate robust generalizable research. It harmonizes five intensive care databases (MIMIC-III, MIMIC-IV, eICU-CRD, SICdb, and ZhejiangProvinceICUdb) spanning 209 clinical sites across three countries (US, Austria, and China), encompassing 93654 ICU admissions (2001-2022). Notably, in addition to routinely collected clinical data, the dataset integrates geriatric-related variables such as physical frailty, GNRI, and code status, which are particularly relevant for assessing vulnerability in older patients.

Background

The phenomenon of population aging has emerged as a pressing societal challenge that demands urgent attention.[1] Demographic projections indicate that by 2030, approximately one-fifth of the global population will be aged 65 years or older. With the median age of critically ill patients consistently rising, octogenarians now represent a substantial proportion of ICU admissions, comprising 36% of the critical care population.[2] Advanced age is associated with a constellation of physiological changes, including diminished organ function, increased frailty, impaired activities of daily living, reduced mobility, and cognitive decline.[3] Among octogenarians in particular, mortality rates are substantial, ranging from 15-25% in the ICU to 25-45% in the hospital, with one-year mortality potentially reaching 70%.[4] Consequently, elderly patients represent a distinct and vulnerable subgroup within ICU populations.[5,6] These findings underscore the critical necessity for prioritizing geriatric care and focused interventions.

In recent years, the emergence of public critical care databases has significantly accelerated research in this field. In North America, the Medical Information Mart for Intensive Care Database (MIMIC) and the eICU Collaborative Research Database (eICU-CRD) have served as foundational resources; MIMIC captures detailed clinical data for patients admitted to the Beth Israel Deaconess Medical Center between 2001 and 2022, while the eICU-CRD covers over 200 sites across the United States from 2014 to 2015.[7,8] Similarly, European research has been bolstered by the Amsterdam University Medical Centers Database (AmsterdamUMCdb), which contains data covering 2003 to 2016, and the Salzburg Intensive Care database (SICdb), which provides high resolution data from Austria.[9] In addition to resources from the US and Europe, China has emerged as a contributor to public critical care data with the release of the ZhejiangProvinceICUdb and the JinhuaNSICU. The ZhejiangProvinceICUdb, released in 2023, originates from a single tertiary care center in China and includes high-resolution data from the comprehensive central ICU and emergency ICU (EICU) between 2012 and 2022.[10] Furthermore, the JinhuaNSICU, published in 2025, specifically targets the neurosurgical intensive care population, covering admissions from 2014 to 2023 at a tertiary hospital.[11] Collectively, these databases have greatly facilitated the investigation of epidemiology and prognosis in elderly critical illness. However, despite this proliferation of resources, most existing ICU prediction models suffer from single-center biases and insufficient external validation, severely limiting their clinical adoption and hindering advancements in personalized critical care.[12,13] Although recent multi-center studies have utilized MIMIC, eICU-CRD, and AmsterdamUMCdb to validate models across US and European cohorts, the applicability of these findings to broader populations, particularly Asian cohorts, remains questionable.[14] Despite their value, these databases present significant accessibility challenges, including large storage requirements, heterogeneous data formats, specialized querying skills (e.g., database management software, structured query language (SQL)), and medical-engineering interdisciplinary background. Therefore, there remains an urgent need for a user-friendly, unified, multinational and public intensive care dataset for elderly populations to enable methodologically sound and widely applicable scientific insights.

Recognizing the limitations in existing researches, we present a comprehensive database for international multicenter studies on elderly critical illness research, integrating harmonized patient records from multiple public intensive care datasets including MIMIC-III, MIMIC-IV, eICU-CRD, SICdb, and ZhejiangProvinceICUdb. By compiling standardized, well-structured data from diverse sources, this initiative aims to address the persistent challenges of single-center bias, limited generalizability, and accessibility barriers that currently hinder progress in geriatric critical care research.

Methods

Methods

The ELDER-ICU-DB dataset integrates multisource data extracted from five critical care databases: MIMIC-III, MIMIC-IV, eICU-CRD, SICdb, and ZhejiangProvinceICUdb [15-19]. Our proposed multinational dataset comprised patients aged ≥65 years, and all hospitalization episodes for each patient were included. To ensure the statistical independence of observations, only the initial ICU admission during each hospitalization was analyzed for each patient. We implemented exclusion criteria encompassing: 1) ICU stays shorter than 24 hours, 2) cases lacking essential demographic (age) or clinical outcome data, and 3) absence of baseline physiological parameters recorded during the first ICU day—specifically heart rate, respiratory rate, mean blood pressure, Glasgow Coma Scale (GCS), temperature, and peripheral oxygen saturation (SpO2).

Data extraction and harmonization

From each database, we systematically extracted and derived a comprehensive set of key variables to facilitate the analysis. This encompassed demographic data such as age, sex, and ethnicity, alongside detailed hospitalization records including ICU and hospital length of stay (LOS), admission type, ICU admission specialty, outcome status, and precise timestamps for admission and discharge. Furthermore, we determined 28-day mortality status and calculated survival times using discharge and mortality records to enable both short- and long-term prognostic assessments. Clinical physiologic data were rigorously captured, comprising laboratory test results recorded from 6 hours before ICU admission through the first 24 hours of the stay, as well as vital signs. For continuous physiological variables recorded over a 24-hour period, we calculated the following metrics: Mean value: The arithmetic average of all measurements obtained within the 24-hour window. Maximum value (Max): The highest recorded value during the observation period. Minimum value (Min): The lowest recorded value during the observation period. Therapeutic interventions, including the use of mechanical ventilation, vasopressors, and code status, were identified based on administration during the first ICU day. The vasopressor assessment focused on four fundamental vasoactive agents such as norepinephrine, epinephrine, dopamine, and dobutamine. Additionally, we calculated established clinical severity scores, including the Charlson Comorbidity Index (CCI), Sequential Organ Failure Assessment (SOFA), Oxford Acute Severity of Illness Score (OASIS), and Simplified Acute Physiology Score (SAPS II). Finally, to address the specific focus of this study, we incorporated geriatric-related information such as physical frailty and the Geriatric Nutritional Risk Index (GNRI).

We derived a categorical variable named “code status” from the chartevents/careplangeneral table of the databases, indicating the patient’s emergent treatment preferences—specifically distinguishing between "full code," which denotes permission for all resuscitative and life-sustaining measures, and limited options such as partial code (Do Not Resuscitate, Do Not Intubate) or comfort care. The MIMIC and eICU-CRD databases systematically document and evaluate code status, while, owing to inherent limitations in their respective source datasets, SICdb and ZhejiangProvinceICUdb lack the assessment of this parameter.

The intensive care setting employs several commonly used clinical scoring systems-including the OASIS, SAPS II, and SOFA-which serve as critical tools for evaluating disease severity. Our meticulously curated dataset integrates computationally derived prognostic indices, which are computed based on physiological parameters recorded during the initial 24-hour intensive care unit stay.

• GCS: Glasgow Coma Scale (GCS), a neurological scale assessing level of consciousness using eye, verbal, and motor responses. Score ranges from 3 (deep coma) to 15 (fully alert).
• Oxford Acute Severity of Illness Score (OASIS): OASIS is a severity-of-illness scoring system designed for use in critical care settings without relying on laboratory values. It includes variables such as age, heart rate, systolic blood pressure, temperature, respiratory rate, oxygenation, level of consciousness, and ICU admission type.
• Simplified Acute Physiology Score (SAPS II): SAPS II is a severity-of-illness score, which incorporating physiological variables measured during the first 24 hours of ICU stay, along with age, severe morbidity and admission type.
• Sequential Organ Failure Assessment score (SOFA): The SOFA score evaluates organ dysfunction across six systems: respiratory, cardiovascular, hepatic, coagulation, renal, and neurological. Each system is scored from 0 (no failure) to 4 (severe failure), yielding a total score ranging from 0 to 24.

Recent studies have demonstrated that certain specific clinical characteristics, particularly the Geriatric Nutritional Risk Index (GNRI) and frailty status, exhibit robust associations with prognostic outcomes in elderly populations. Our quantitative assessment of these parameters aims to facilitate more comprehensive geriatric research. To systematically assess physical frailty, we extracted and defined best state of activity-including sitting, standing, and bed rest—from events documented during the 24 hours before and after ICU admission in the chartevents/ careplangeneral tables. Using conditional aggregation, we defined three binary indicators for each patient’s activity status:

• Stand: Assigned a value of 1 if the patient was documented as able to stand or walk independently (e.g., "Ambulate", "Walk - 10+ Steps").
• Sit: Assigned a value of 1 if the patient could sit up independently but was chairbound (e.g., "Chair", "Transfer to chair", "Lift to chair/bed").
• Bed rest: Assigned a value of 1 if the patient was restricted to bed rest with limited mobility (e.g., "Bedrest", "Only lying", "Turning in bed").

The presence of any qualifying event during the observation period resulted in a positive flag for that activity level. An additional variable (activity_eva_flag) was created to indicate whether any mobility assessment was documented.

GNRI is a nutritional assessment tool used to evaluate the risk of malnutrition and associated adverse outcomes in elderly patients. It incorporates serum albumin levels and body mass index (BMI), or when BMI is unavailable, it uses the ratio of actual body weight (ABW) to ideal body weight (IBW). The GNRI was calculated using the following formula:

GNRI = (14.89 × Albumin [g/dL]) + (41.7 × min(IBW/ABW, 1))

Where:

Serum albumin levels were obtained from laboratory measurements recorded during the initial ICU stay.

IBW was estimated using the Devine formula:

IBW(kg)=45.5 + 0.9 × (Height[cm] − 152) for females

IBW(kg)=48.0 + 0.9 × (Height[cm] − 152) for males​

A GNRI score < 92 indicates high nutritional risk, while scores ≥ 98 suggest low risk.

All data extraction and preprocessing steps were implemented using Structured Query Language in PostgreSQL (version 13) and Python (version 3.9).

Data Description

This database aggregates data from a total of 209 distinct clinical sites located across three different countries, specifically the United States, Austria, and China. In terms of scale, it encompasses a comprehensive collection of 93,654 ICU admissions that were recorded between the years 2001 and 2022. Specifically, the dataset is composed of 48,708 admissions from MIMIC, 37,166 admissions from eICU-CRD, 6,371 admissions from SICdb, and 1,409 admissions from ZhejiangProvinceICU. The ELDER-ICU-DB dataset consists of 4 CSV files, each storing data extracted and organized from the 4 source databases. A brief summary of each file is presented in Table 1.

Table 1: Summary of data files

The source-specific identifiers listed in the table enable cross-referencing with the original databases for users who have obtained proper access permissions, but they cannot be used to link patients across different databases. For detailed information on variable availability across source databases, users can refer to the Variable_Availability_Matrix.csv file provided in the data package.

Demographics

Patient and icu_admission identifiers: The original unique patient and admission identifiers from source databases are preserved within the ELDER-ICU-DB to facilitate cross-referencing with primary datasets for supplemental clinical information. It is important to clarify that these identifiers are pseudonymised and pose no risk of direct patient identification; furthermore, their uniqueness is strictly limited to the scope of each specific source database. Consequently, while they allow researchers to retrieve additional information from the original database context, they cannot be utilized to link records of the same individual across different databases. Distinct identifier schemes across databases are detailed in Table 2.

• id: Our curated composite identifier that uniquely specifies each ICU admission instance.
• age: Records exact chronological age in years.
• gender: Coded dichotomously as F (female) or M (male).

Table 2 Patient and admission identifiers across databases

Admission characteristics

• weight: Weight on admission (in kilograms).
• height: Height on admission (in centimeters). Height is uniformly recorded in centimeters (cm) across all source databases, with the majority of values falling within expected physiological ranges (121–226 cm) in MIMIC.csv and eICU-CRD.csv. However, several outliers were identified in other source files: SICdb.csv contains 4 records with height <100 cm (ranging 60–95 cm), and ZhejiangProvinceICU.csv has 1 record with height of 53 cm. These outliers are retained to allow researchers to apply their own exclusion criteria based on their specific analytical needs.
• anchor_year_group: Admission year group identifier.
• admission_type: Type of hospital admission (elective, urgent, emergency and observation).
• first_careunit: The first ICU unit the patient was admitted to within the current hospital stay.
• los_icu_day: Length of stay in the ICU during the current admission (in days).
• los_hospital_day: Total length of hospital stay during the current admission (in days).
• death_hosp: Indicator of in-hospital mortality (1=died in hospital, 0=survived to discharge).
• deathtime_icu_hour: Time elapsed from ICU admission to death (in hours). Null if patient survived.
• pre_icu_los_day: Length of hospital stay prior to ICU admission (in days).
• survival_time_28d (28-day Survival Time)：Quantifies patient survival duration in calendar days post-ICU admission. The variable is right-censored at 28 days, such that: Survivors beyond the 28-day observation window are recorded as 28 days; decedents are annotated with their exact survival duration (≤28 days). Note that a value of 0 indicates the patient died within 0–1 days after ICU admission (i.e., on the same calendar day), which is due to the fact that date of death (dod) in the source MIMIC database is recorded only to the day precision, while ICU admission time (icu_intime) is recorded to the second. Additionally, null (NaN) values have been assigned to records where the survival time could not be reliably determined due to source data errors. Users are advised to treat these as data artifacts and apply their own exclusion criteria as needed.
• death_28d (28-day Mortality): A binary indicator of all-cause mortality within the 28-day observation period.
• hospitalid, teachingstatus, region (eICU-CRD Database Only): The hospitalid serves as a unique numerical identifier for each hospital. The teachingstatus field indicates whether the facility operates as an academic teaching hospital, while the region denotes the hospital's geographical location within the United States, comprising four distinct areas: Northeast, Midwest, South, and West.

Vital signs

A total of eight vital sign variables were included. Detailed variable names and definitions are presented in Table 3. Temperature at ICU admission is recorded in degrees Celsius (°C) across all source databases, with the majority of values falling within expected physiological ranges (31–40°C) in MIMIC.csv and eICU-CRD.csv. However, SICdb.csv contains 483 records with temperature <34°C and 12 records with temperature >42°C. These outliers are retained to allow researchers to apply their own exclusion criteria based on their specific analytical needs.

Table 3 Vital Sign Variables and Definitions

Labs

A total of 27 laboratory variables were included. Detailed variable names and definitions are presented in Table 4.

Table 4 Laboratory Variables and Definitions

ABG data

Blood gas-related variables comprise 7 parameters. For detailed variable names and definitions, please refer to Table 5.

Table 5 Blood Gas Variables and Definitions

Treatment

• electivesurgery: Indicator of whether the ICU admission was scheduled following an elective surgical procedure (1=yes, 0=no).
• vent: Indicator of whether the patient received mechanical ventilation (1=yes, 0=no).
• code_status: Patient’s emergent treatment preferences (1=partial code, 0=full code).
• code_status_eva_flag: Flag indicating whether the patient's code status was evaluated or recorded.

Medication

• norepinephrine: Indicator of whether norepinephrine (a vasopressor) was administered (1=used, 0=unused).
• epinephrine: Indicator of whether epinephrine (a vasopressor/inotropic agent) was administered (1=used, 0=unused).
• dopamine: Indicator of whether dopamine (a vasopressor/inotropic agent) was administered (1=used, 0=unused).
• dobutamine: Indicator of whether dobutamine (an inotropic agent) was administered (1=used, 0=unused).

Physical frailty

• gnri: Geriatric Nutritional Risk Index.
• activity_stand: Flag indicating whether the patient is able to stand up and walk independently (mobile and not restricted to bed or chair).
• activity_sit: Flag indicating whether the patient is able to sit up independently but remains primarily chairbound (unable to walk independently).
• activity_bed: Flag indicating whether the patient is able to move only in bed and is considered bedbound (unable to sit up or stand without assistance).
• activity_eva_flag: Flag indicating whether the patient's activity level was evaluated or recorded.

Clinical scores

• cci_score: Charlson Comorbidity Index (CCI) score.
• oasis: Oxford Acute Severity of Illness Score
• saps: Simplified Acute Physiology Score
• sofa: Sequential Organ Failure Assessment score

Usage Notes

This dataset is designed to support multicenter studies on prognosis, mortality prediction, and risk stratification in elderly ICU patients, as well as comparative effectiveness research across different healthcare systems (US, Europe, Asia). It also enables validation and benchmarking of established clinical scoring systems—including OASIS, SAPS II, SOFA, and CCI—specifically in geriatric populations. Researchers can leverage the harmonized data structure to conduct cross-database analyses without the need for separate data preprocessing across individual sources.

However, users should be aware of several important limitations. The dataset includes only the first ICU admission per hospitalization, excluding subsequent transfers or readmissions, and the 24-hour observation window for physiological variables captures only the initial ICU day. Additionally, data completeness and variable availability differ across source databases—for example, code status is unavailable in SICdb and ZhejiangProvinceICUdb. Notably, before applying for access to this derived dataset, data users must first become certified users of PhysioNet, submit a project application with a specific research question for approval by the dataset contributors, and accept the required data use agreement.

Release Notes

This is the initial release of the ELDER-ICU-DB. This first version aims to provide a valuable resource for researchers focusing on critical care gerontology studies. Future updates may incorporate more comprehensive longitudinal patient hospitalization data.

Ethics

ELDER-ICU-DB adheres to the same access policies, licenses, and data use requirements as its parent credentialed datasets. No additional protected health information (PHI) has been introduced, and all data are de-identified in accordance with the de-identification standards applied by the original database providers. For the ZhejiangProvinceICUdb and SICdb source databases, the access requirements for ELDER-ICU-DB follow the same procedures as those of the original source datasets. To obtain access, users must first become credentialed PhysioNet users, submit a project application with a specific research question for approval by the dataset contributors, and accept the required data use agreement.

Conflicts of Interest

No conflicts of interest to declare.

References

1. Partridge L, Deelen J, Slagboom PE. Facing up to the global challenges of ageing. Nature. 2018;561(7721):45-56.
2. Vincent JL, Creteur J. Appropriate care for the elderly in the ICU. J Intern Med. 2022;291(4):458-68.
3. Brunker LB, Boncyk CS, Rengel KF, Hughes CG. Elderly Patients and Management in Intensive Care Units (ICU): Clinical Challenges. Clin Interv Aging. 2023;18:93-112.
4. Guidet B, Vallet H, Boddaert J, de Lange DW, Morandi A, Leblanc G, et al. Caring for the critically ill patients over 80: a narrative review. Ann Intensive Care. 2018;8(1):114.
5. Guidet B, de Lange DW, Boumendil A, Leaver S, Watson X, Boulanger C, et al. The contribution of frailty, cognition, activity of daily life and comorbidities on outcome in acutely admitted patients over 80 years in European ICUs: the VIP2 study. Intensive Care Med. 2020;46(1):57-69.
6. Jung C, Flaatten H, Fjølner J, Bruno RR, Wernly B, Artigas A, et al. The impact of frailty on survival in elderly intensive care patients with COVID-19: the COVIP study. Crit Care. 2021;25(1):149.
7. Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 2023;10(1):1.
8. Pollard TJ, Johnson AEW, Raffa JD, Celi LA, Mark RG, Badawi O. The eICU Collaborative Research Database, a freely available multi-center database for critical care research. Sci Data. 2018;5:180178.
9. Rodemund N, Wernly B, Jung C, Cozowicz C, Koköfer A. Harnessing Big Data in Critical Care: Exploring a new European Dataset. Sci Data. 2024;11(1):320.
10. Jin S, Chen L, Chen K, Hu C, Hu S, Zhang Z. Establishment of a Chinese critical care database from electronic healthcare records in a tertiary care medical center. Sci Data. 2023;10(1):49.
11. Chen L, Huang Z, Liu X, He G, Ren B, Yang F, et al. JinhuaNSICU, an open accessible Neurosurgical Intensive Care Database. Sci Data. 2025;12(1):1633.
12. Ma R, Zhao J, Wen Z, Qin Y, Yu Z, Yuan J, et al. Machine learning for the prediction of delirium in elderly intensive care unit patients. Eur Geriatr Med. 2024;15(5):1393-403.
13. Takkavatakarn K, Oh W, Chan L, Hofer I, Shawwa K, Kraft M, et al. Machine learning derived serum creatinine trajectories in acute kidney injury in critically ill patients with sepsis. Crit Care. 2024;28(1):156.
14. Liu X, Hu P, Yeung W, Zhang Z, Ho V, Liu C, et al. Illness severity assessment of older adults in critical illness using machine learning (ELDER-ICU): an international multicentre study with subgroup bias evaluation. Lancet Digit Health. 2023;5(10):e657-e67
15. Johnson A, Pollard T, Mark R. MIMIC-III Clinical Database (version 1.4). PhysioNet. 2016. RRID:SCR_007345. Available from: https://doi.org/10.13026/C2XW26
16. 16. Johnson A, Bulgarelli L, Pollard T, Gow B, Moody B, Horng S, Celi L A, Mark R. MIMIC-IV (version 3.1). PhysioNet. 2024. RRID:SCR_007345. Available from: https://doi.org/10.13026/kpb9-mt58
17. Pollard T, Johnson A, Raffa J, Celi L A, Badawi O, Mark R. eICU Collaborative Research Database (version 2.0). PhysioNet. 2019. RRID:SCR_007345. Available from: https://doi.org/10.13026/C2WM1R
18. Rodemund N, Kokoefer A, Wernly B, Cozowicz C. Salzburg Intensive Care database (SICdb), a freely accessible intensive care database (version 1.0.8). PhysioNet. 2024. RRID:SCR_007345. Available from: https://doi.org/10.13026/8m72-6j83
19. Jin S, Chen L, Chen K, Zhang Z. Establishment of a Chinese critical care database from electronic healthcare records in a tertiary care medical center (version 1.0). PhysioNet. 2023. RRID:SCR_007345. Available from: https://doi.org/10.13026/901c-yv54