Overview

Table of contents

1. Format
2. Split
3. Data variables
4. Pre-processing
5. Outcomes
6. Tables
   1. Static tables
   2. Time-varying tables

Format

Data are provided as CSV files.

Split

Sets of three datasets (training, validation and test) were created from all inpatient encounters which gave rise to an admission to the general internal medicine (GIM) ward.

• The training set is defined by all encounters where the admission time to GIM (GIM_START_TS) is strictly less than December 1, 2017 00:00:00.
• The validation set consists of all encounters that occur after December 1, 2017 00:00:00 and strictly less than December 1, 2018 00:00:00.
• Finally, the test set is defined as all encounters that occur after December 1, 2018 00:00:00.

Data variables

The top occurring/most important variables were selected in consultation with a GIM staff physician. Variables included in the dataset include:

• Numeric values
  • 9 vital signs
  • 100 labs
  • 7 shift assessment variables
  • 7 intake-outtake variables
  • 1 ulcer variable, 1 alcohol scale, 1 diabetes variable
• Clinical orders (165)
  • Imaging, Telemetry, Consults, cardio, diet, respiration, activities, codes, protocols, transfusions, wound care, neuro
• Medication administrations
  • Medications are grouped into medication AHFS classes.

Pre-processing

The dataset is provided in both a preprocessed format and as raw data tables. Pre-processing is described below.

• Time-varying data is binned into 8-hour windows starting from the time the patient enters the GIM ward. The timeline ends when the patient experiences one of the outcomes.

• For numeric data (e.g., labs, vitals), we take the average measurement within each window. We then add an indicator for whether or not the variable was measured in that window. We also add a time since last measured variable that keeps track of how long it has been since the variable was last recorded. Numeric data rae then trimmed and normalized between 0 and 1.

• For orders, we created an indicator (1 vs 0) for whether or not the order was active during that window.

• For medication administrations, we grouped medications into classes and then created an indicator for whether or not the medication class was administered in that time window.

• Missing data: For numeric data, we used last observation carried forward with mean imputation until a first observation is recorded. For orders, we impute a zero.

Outcomes

Outcome is defined as the first occurrence of any of the following events (i.e. if a patient transfers to the ICU and then dies, their outcome will be ICU transfer). The following patient outcomes will be available at the patient encounter level:

1. ICU transfer - Patient transfers from GIM to ICU.
2. Death - Patient dies while on GIM.
3. Palliative entry - Patient gets transferred to palliative care.
4. Palliative discharge - Patient gets discharge to a palliative care unit
5. Discharged - Patient is discharged from hospital

We also capture the following outcomes in a separate table.

• ICD diagnosis
• Sepsis (sepsis defined based on SIRS criteria)
• Respiratory failure (respiratory failure based on ventilator order)

Tables

Where applicable, we provide both raw and pre-processed forms of the data.

Static tables

Time-varying tables