Predicting Acute Hypotensive Episodes: The PhysioNet/Computing in Cardiology Challenge 2009

Introduction

Among the most critical events that occur in intensive care units (ICUs), acute hypotensive episodes require effective, prompt intervention. Left untreated, such episodes may result in irreversible organ damage and death. Timely and appropriate interventions can reduce these risks. Determining what intervention is appropriate in any given case depends on diagnosing the cause of the episode, which might be sepsis, myocardial infarction, cardiac arrhythmia, pulmonary embolism, hemorrhage, dehydration, anaphylaxis, effects of medication, or any of a wide variety of other causes of hypovolemia, insufficient cardiac output, or vasodilatory shock. Often the best choice may be a suboptimal but relatively safe intervention, simply to buy enough time to select a more effective treatment without exposing the patient to the additional risks of delaying treatment.

Of the 2320 patients whose monitored waveforms and accompanying clinical data were included in the MIMIC-II Database as of December 2008, arterial blood pressure was recorded in 1237 (53%); among these 1237 patients, 511 (41%) experienced recorded episodes of acute hypotension (as defined below) during their ICU stays. The mortality rate for these 511 patients is more than twice that of the MIMIC II population as a whole. To the extent that one might forecast acute hypotensive episodes in the ICU, there is a possibility of improving care and survival of patients at risk of these events.

This year's challenge is the tenth in the annual series of open challenges hosted by PhysioNet in cooperation with Computers in Cardiology. The goal of the challenge is to predict which patients in the challenge dataset will experience an acute hypotensive episode beginning within the forecast window.

Definitions

Acute Hypotensive Episode: The challenge dataset includes, for each case, a time series of mean arterial blood pressure (MAP) at one-minute intervals. Each sample of the series is an average of the blood pressure measured in the radial artery over the previous minute. Given such a time series, an acute hypotensive episode (AHE) is defined for the purposes of this challenge as any period of 30 minutes or more during which at least 90% of the MAP measurements were at or below 60 mmHg.

Forecast window: This is defined as the one-hour period immediately following a specified time T₀. In the test sets, the forecast window (and indeed all data following T₀) are withheld, and the forecast must be made using only information available before T₀.

The Challenge Dataset

The MIMIC II project has collected data from about 30,000 ICU patients to date. MIMIC II patient records contain most of the information that would appear in a medical record (such as results of laboratory tests, medications, and hourly vital signs). About 5,000 of the records also include physiologic waveforms (typically including ECG, blood pressure, and respiration, and often other signals as well) and time series that can be observed by the ICU staff. The intent is that a MIMIC II record should be sufficiently detailed to allow its use in studies that would otherwise require access to an ICU, e.g., for basic research in intensive care medicine, or for development and evaluation of diagnostic and predictive algorithms for medical decision support.

The challenge dataset consists of selected patient records from the MIMIC-II Database. In the training set, the records include all available data before and after T₀. In the test sets, the records are truncated at T₀; the data recorded after T₀ in each case will be made available for study only after the conclusion of the challenge. (Update 2 September 2009: The previously withheld data are now available; see the test set home page.)

Not all MIMIC II records include all of the data elements needed for this challenge. Records chosen for the challenge dataset include, at a minimum:

• At least 10 hours of data before T₀, and at least one hour of data after T₀. (As noted, data collected after T₀ for the test set records will be withheld until after the conclusion of the challenge in September, 2009.) Most MIMIC II records are significantly longer, and many include a week or more of data.
• ECG and arterial blood pressure (ABP) signals sampled at 125 Hz. Records in the training set may include one or two additional signals, and those in the test set may include as many as six additional signals. (Note that two records in the training set do not include these signals, however.)
• Time series of vital signs sampled once per minute (in the training set) and once per second (in the test sets). These include heart rate and mean, systolic, and diastolic ABP. Most records include a variety of additional vital-signs time series, most often including respiration rate and SpO₂.
• Clinical data entered into the ICU medical information systems (records of observations, measurements, and interventions performed in the ICU). These include intravenous medications and fluids as well as other medications administered. Note, however, that some of this information is manually entered by the ICU staff at times when it is possible to do so without compromising patient care, so the associated timestamps may be imprecise.
• Results of laboratory tests, records of medications ordered, and other data gathered in the hospital but outside of the ICU.

MIMIC II records meeting the criteria above are assigned to a group (H or C) and a subgroup (H1, H2, C1, or C2):

• Records in group H contain an episode of acute hypotension beginning during the forecast window (the one-hour period following T₀).
  • Records in subgroup H1 come from patients who received pressor medication.
  • Records in subgroup H2 come from patients who did not receive pressor medication (i.e., those in group H but not in subgroup H1).
• Records in group C contain no episodes of acute hypotension within the forecast window.
  • Records in subgroup C1 come from patients with no documented acute hypotensive episodes at any time during their hospital stay.
  • Records in subgroup C2 come from patients who had AHE before or after the forecast window (i.e., those in group C but not in subgroup C1).

The training set consists of 60 records (including data after T₀):

• 15 from subgroup H1 (AHE treated with pressors)
• 15 from subgroup H2 (AHE not treated with pressors)
• 15 from subgroup C1 (no AHE)
• 15 from subgroup C2 (AHE outside the forecast window)

Test set A consists of 10 records (excluding data after T₀):

• 5 from subgroup H1 (AHE in subjects receiving pressors)
• 5 from subgroup C1 (no AHE, in subjects receiving pressors)

Test set B consists of 40 records (excluding data after T₀):

• Between 10 and 16 from group H
• Between 24 and 30 from group C

Changes in the Events and Group Definitions

Originally, subgroup H1 was defined as patients who received pressor medication in response to their AHE, and subgroup C1 included some patients who did not receive pressors. These definitions were used to construct the training set. Although not explicitly stated as a selection criterion, none of these patients received pressors before T₀, so the administration of pressors was not a clue that could have been used to classify them.

In selecting the test sets, however, it became apparent that cases meeting the original criteria for subgroup H1 were less common than anticipated. By including cases in which pressors were already being given, we were able to obtain a sufficient number of cases in subgroup H1, but now the problem was that if they were included, it would be possible to classify them simply by observing that pressors were being given. The solution was to redefine subgroups H1 and C1 to include only records of patients who received pressors, so that, as in the training set, the presence of pressors per se does not indicate to which group a record belongs.

Challenge Events

Event 1 focuses on distinguishing between two groups of ICU patients who are receiving pressor medication: patients who experience an acute hypotension episode, and patients who do not. These two groups represent extremes of AHE-associated risk. Designing successful methods for separating these disjoint populations may lead to finding indices that are prognostic of AHE in these individuals.

To enter event 1, design and implement an automated method to identify which of the records in test set A belong to subgroup H1.

Event 2 aims to address the broad question of predicting AHE in a population in which about a third of the patients experience AHE (as in the MIMIC II Database as a whole). It is likely that a variety of methods can be used to identify different subsets of the patients at risk; for example, those who have had previous documented AHE (especially if more than once) may be relatively easy to identify, on the basis of a priori knowledge of their pathophysiology or of their response to medication. The potential benefits of finding AHE predictors for even a modest subset of the at-risk patients may be significant, if improvement in outcome can be shown to follow from increased vigilance and preparation for effective intervention in these patients.

To enter event 2, design and implement an automated method to identify which of the records in test set B belong to group H.

Entering the Challenge

The Challenge has concluded, but it is still possible to attempt the challenge problem, since the data will remain available. Follow the links to the correct classifications at the top of this page in order to determine the accuracy of your predictions. The remainder of this page describes the rules for official participants in the Challenge.

We recommend studying the training set records as preparation for the Challenge itself. The opportunity to see what happens after T₀ in these records will be invaluable in designing successful strategies for predicting acute hypotensive episodes in the test set records.

Download challenge entry forms for event 1 and event 2, then follow the instructions on the forms to fill in your algorithm's classifications and return them for scoring. Be sure that your entry form(s) include the email address where you wish to have your score(s) sent. In each event, the score is simply the fraction of correct classifications (a number between 0 and 1; higher scores are better).

To be eligible for an award:

1. Design and implement an automated method for classifying the records in one or both test sets, and record your method's classifications in an entry form.
2. Submit at least one valid entry (to either event) no later than noon GMT on Wednesday, 6 May 2009.
3. Submit an abstract (about 300 words) describing your work on the Challenge to Computers in Cardiology (CinC). Please select "PhysioNet/CinC Challenge" as the topic of your abstract, so that it can be identified easily by the abstract review committee. The deadline for submitting abstracts is 8 May 2009.
4. Attend Computers in Cardiology in Park City, Utah (USA), 13-16 September 2009.

Entries for event 1 must assign exactly 5 records from test set A to subgroup H1, and entries for event 2 must assign 10 to 16 records from test set B to group H. Entries that do not follow this rule are invalid and will not be scored. You may enter each event up to four times. These restrictions are intended to limit the opportunity for a good score based on a lucky guess, or a series of deductions as in Mastermind. The deadline for revised entries is noon GMT on Monday, 31 August 2009.

If your abstract is accepted, you will be expected to prepare a four-page paper for publication in Computers in Cardiology, and to present a talk or poster about your work at CinC.

During a plenary session of Computers in Cardiology in September, four awards will be presented to the eligible participants in attendance with the best final scores as follows:

1. Best open source entry in event 1
2. Best entry in event 1
3. Best open-source entry in event 2
4. Best entry in event 2

Participants may enter one or both events, and open source entries are eligible for the overall awards as well as for the open source awards. If the best results in any category are achieved by two or more entries, the first of these entries to be submitted will receive the award in that category.

Important! The challenge is to design an automated method for classifying the records. You are welcome to submit an entry based on your personal interpretation of the data, but it will not be scored until the conclusion of the challenge, and it will not be eligible for an award. (If you do very well, however, your achievement will be recognized on PhysioNet.) Obviously this rule is difficult to enforce; please respect the spirit of the challenge.

Entering the Open Source Division

As in previous years, the Challenge includes an open source division. You may enter the open source division by sending the source code for your classifier by email, before noon GMT on Tuesday, 1 September 2009, to PhysioNet. Use the subject line "Challenge 2009 entry source", and be sure to include:

• Your name and email address
• All sources needed to produce a working version of your software (except for readily available standard libraries and header files)
• A note describing how to produce a working version of your software (a commented Makefile is ideal), and how to run your software

Each source file submitted should begin with a comment block containing the names of its authors and a reference to the open source license you have chosen for it, if any; for example:

     /* predict_ahe.c - forecast acute hypotension using artificial intuition
        Copyright (C) 2009  Herman Foobar <hbar@uncertain.org>
        This software is released under the terms of the GNU General
        Public License (http://www.gnu.org/copyleft/gpl.html).
     */

Source files in C, C++, Fortran, or Matlab m-code are preferred; other languages may be acceptable, but please ask first. Do not submit any code that cannot be freely redistributed. Following the conclusion of the Challenge, selected entries will be posted, with full credit to their authors, on PhysioNet.

Challenge Results

The Challenge has concluded, and the correct classifications for event 1 and event 2 are now available. In addition, the c records (the data following T₀ for each of the test set records, which have been withheld for the duration of the Challenge) are now available in the same locations as the a and b records that have been available since April.

Thanks to all of the Challenge participants, many of whom will discuss their work during dedicated scientific sessions of Computers in Cardiology next week. A number of participants were able to classify all 10 cases in event 1 without errors; many were able to classify at least 80% of cases in event 2, and the best result achieved in event 2 was correct classification of 93% (37 of 40) cases. The final scores, and the open source software developed and contributed by participants in the open source division of the Challenge, will be posted shortly after CinC.

Special thanks to Franco Chiarugi, whose invaluable feedback at every stage prompted corrections in the training set and improvements in the design of the challenge that contributed significantly to its success.

Papers presented at CinC 2009 by challenge participants, scores, and sources for the open-source entries, are now available.

Papers

The first of these papers is an introduction to the challenge topic, with a summary of the challenge results and a discussion of their implications.

Predicting Acute Hypotensive Episodes: The 10th Annual PhysioNet/Computers in Cardiology Challenge
GB Moody, LH Lehman

The remaining papers were presented by participants in the Challenge, who describe their approaches to the challenge problem.

Forecasting Acute Hypotensive Episodes in Intensive Care Patients Based on a Peripheral Arterial Blood Pressure Waveform
X Chen, D Xu, G Zhang, R Mukkamala

Prediction of Acute Hypotensive Episodes Using Neural Network Multi-models
JH Henriques, TR Rocha

Predicting Acute Hypotensive Episodes from Mean Arterial Pressure
P Langley, S King, D Zheng, EJ Bowers, K Wang, J Allen, A Murray

A Rule-Based Approach for the Prediction of Acute Hypotensive Episodes
MA Mneimneh, RJ Povinelli

Predicting the Occurrence of Acute Hypotensive Episodes: The PhysioNet Challenge
F Chiarugi, I Karatzanis, V Sakkalis, I Tsamardinos, Th Dermitzaki, M Foukarakis, G Vrouchos

Acute Hypotension Episode Prediction Using Information Divergence for Feature Selection, and Non-Parametric Methods for Classification
PA Fournier, JF Roy

A Biosignal Analysis System Applied for Developing an Algorithm Predicting Critical Situations of High Risk Cardiac Patients by Hemodynamic Monitoring
D Hayn, B Jammerbund, A Kollmann, G Schreier

Smoothing and Discriminating MAP Data
K Jin, NL Stockbridge

Computers in Cardiology / Physionet Challenge 2009: Predicting Acute Hypotensive Episodes
F Jousset, M Lemay, JM Vesin

Utilizing Histogram to Identify Patients Using Pressors for Acute Hypotension
TCT Ho, X Chen