HiRID, a top time-resolution icu dataset. Anonymization procedure

Posted Variation: 1.0


HiRID is really an easily accessible critical care dataset containing data associated with nearly 34 thousand patient admissions to your Department of Intensive Care Medicine of this Bern University Hospital, Switzerland (ICU), an interdisciplinary 60-bed device admitting >6,500 clients each year. The ICU supplies the complete selection of contemporary interdisciplinary intensive care medication for adult clients. The dataset was created in cooperation involving the Swiss Federal Institute of tech (ETH) ZГјrich, Switzerland in addition to ICU.

The dataset contains de-identified demographic information and a total of 681 regularly gathered physiological factors, diagnostic test outcomes and treatment parameters from very nearly 34 thousand admissions through the duration. Information is kept having an uniquely about time quality of just one entry every 120 seconds.


Critical disease is seen as an the existence or threat of developing organ dysfunction that is life-threatening. Critically sick clients are generally looked after in intensive care units (ICUs), which focus on supplying monitoring that is continuous advanced therapeutic and diagnostic technologies. This dataset ended up being gathered during routine care in the Department of Intensive Care Medicine regarding the Bern University Hospital, Switzerland (ICU), an interdisciplinary 60-bed device admitting >6,500 clients each year. It had been initially removed to guide a research regarding the very very early forecast of circulatory failure into the intensive care device making use of machine learning 1. The latest documents for the dataset is available2.


The HiRID database has a big choice of all routinely gathered data relating to patient admissions towards the Department of Intensive Care Medicine associated with Bern University Hospital, Switzerland (ICU). The information had been obtained from the ICU individual information Management System which will be familiar with register that is prospectively wellness information, dimensions of organ function parameters, outcomes of laboratory tests and therapy parameters from ICU admission to discharge.

Dimensions from bedside monitoring

Measurements and settings of medical products such as for example mechanical air flow

Findings by healthcare providers e.g.: GCS, RASS, urine as well as other output that is fluid

Administered drugs, liquids and nourishment

HiRID has a greater time quality than many other posted datasets, above all for bedside monitoring with many parameters recorded every two minutes.

To guarantee the anonymization of people in the information set, we implemented the procedures effectively sent applications for the MIMIC-IIwe and Amsterdam UMC db dataset, which adopted the ongoing health Insurance Portability and Accountability Act (HIPAA) secure Harbor needs and, when it comes to Amsterdam UMC db, additionally europe’s General information Protection Regulation (GDPR) standards 3,4.

Elimination of all eighteen data that are identifying placed in HIPAA

Times were shifted by way of a random offset so that the admission date lies. We ensured to protect the seasonality, period of time therefore the day’s week.

Individual age, weight and height are binned into containers of size 5. The max bin is 90 years and contains also all older patients for patient age.

Dimensions and medicines with changing devices as time passes had been standardized to your latest device utilized. This standardization ended up being required to create a summary about approximated admission times, in line with the devices found in a certain client, impossible.

Complimentary text had been taken from the database

k-anonymization had been used on patient age, fat, height and intercourse.

Ethical approval and consent that is patient

The review that is institutional (IRB) of this Canton of Bern authorized the research. The necessity for acquiring informed patient consent ended up being waived due to the retrospective and observational nature of this research.

Information Description

The data that are overall for sale in two states: as natural information and/or as pre-processed information. Furthermore you will find three guide tables single parents meet for adjustable lookup.

Guide tables

adjustable guide – guide table for factors (for natural phase)

ordinal guide that is adjustable reference dining table for categorical/ordinal variables for string value lookup

pre-processed adjustable guide – guide dining dining table for factors (for merged and stage that is imputed

Natural information

The raw information was just prepared if it was necessary for patient de-identification and otherwise left unchanged set alongside the initial supply. The foundation information offers the set that is complete of factors (685 factors). It consist of the after tables:

Preprocessed information

The pre-processed information comes with intermediary pipeline phases from the accompanying book by Hyland et al 1. Supply factors representing the exact same concepts that are clinical merged into one meta-variable per concept. The info provides the 18 many predictive meta-variables just, as defined within our book. Two different phases for the pipeline can be found

Merged phase supply factors are merged into meta-variables by clinical ideas e.g. non-opioid-analgesics. The full time grid is kept unchanged and it is sparse.

Imputed phase the information through the merged stage is down sampled up to a five-minute time grid. The full time grid is full of imputed values. The imputation strategy is complex and it is talked about when you look at the initial book.

The rule utilized to create these stages are available in this GitHub repository beneath the preprocessing folder 5.

Which information to make use of?

The pre-processed information is intended primarily as being a way that is quick jump-start a task or even for use within an evidence of concept. We advice utilizing the supply data as much as possible for regular tasks. It’s the many versatile kind and possesses the whole group of factors into the time resolution that is original.

Information platforms

Information is for sale in two platforms: CSV for wide compatibility and Apache Parquet for convenience and gratification.

Considering that the information sets are fairly big, they have been divided into partitions, so that they could be prepared in parallel in a way that is straightforward. The lookup dining dining table mapping patient id to partition id is supplied into the file called combined with the information. The partitions are aligned between your various information sets and tables, in a way that the information of an individual can invariably be located into the partition using the exact same id. Note however, that an individual may well not take place in all data sets, e.g. a patient could be lacking within the preprocessed information, because an individual don’t meet with the demographic requirements to be contained in the research.

Patient ID / ICU admission

The dataset treats each ICU admission uniquely which is extremely hard to spot numerous ICU admissions as originating from the exact same client. For each ICU (re-)admission an original “Patient ID” is created.

Information schemata

The schemata of any dining table are available in the *schemata.pdf* file.

Use Records

Since the database contains detailed information about the care that is clinical of, it should be addressed with appropriate care and respect.

Scientists have to formally request access via PhysioNet. The user has to be a credentialed PhysioNet user, digitally sign the Data Use Agreement and provide a specific research question to be granted access.

Conflicts of Interest

The authors declare no disputes of great interest


Access Policy: Only PhysioNet credentialed users whom signal the specified DUA have access to the files.