ICGrid: Intensive Care Grid

Marios PapaMarios Papa11, Demetrios Zeinalipour, Demetrios Zeinalipour--YaztiYazti 11, Marios D. Dikaiakos, Marios D. Dikaiakos11

George Panayi George Panayi 22, Theodoros Kyprianou, Theodoros Kyprianou 22

1 Dept. of Computer Science 1 Dept. of Computer Science -- University of Cyprus,University of Cyprus,Nicosia, CyprusNicosia, Cyprus

2 Intensive Care Unit 2 Intensive Care Unit –– Nicosia General Hospital, CyprusNicosia General Hospital, Cyprus

http://http://grid.ucy.ac.cygrid.ucy.ac.cy//

2

Introduction

Healthcare Industry:among the world's largest, fastest-growing and most information intensive industries.

Complexity:i) vast amounts of data; ii) varied data quality; iii) privacy constraints, iv) analysis and storage of real signals, v) finding interesting data correlations and dependencies.

Grid:i) tens of thousands of computers, ii) trillions of commands per second, iii) petabytes of storage

=> GRID: the right place to solve the challenges!

3

Motivation• Intensive Care Units (ICUs) monitor patients that are in a

critical (life-threatening) physiological state.

• Patients are connected to a very large number of monitoring devices that continuously acquire the state of the respective patient.

• Clinically Interesting Episodes (CIE), e.g. (temp>X and press>Y) =>cerebral emergency 95% represent a minority in the acquired signals, but are of critical importance.

• Proactively mining the local CIE history log is not meaningful (due to the small size). Therefore, doctors can only reactively respond to alerts coming from monitors.

4

Goal

Create a (distributed) tool that enables the seamless

integration, correlation and retrieval of clinically

interesting episodes across Intensive Care Units.

5

Projected Features

• Storage / Archiving of Interesting Episodes• Browsing Other Related Episodes • Automatic identification of similar episodes

using (high performance) timeseriessimilarity methods: Given a real signal, find other signals with a similar temporal movement.

• Other High Performance Data Mining Tools:Predicting the future value of a signalClustering Similar Patient States

All Computationally/Storage Intensive Tasks!

0 20 40 60 80 100

Point Correspondense, Similarity [δ=10,ε =0.3]

= 0.25

6

Architecture: Local

A) Local Data Acquisition B) Storage and Indexing C) Sharing Filtered Data on the Grid.

7

Architecture: Grid

D) Storage/Replication using LCG.E) Information Processing and AggregationF) Query Interface Modules

8

CyGrid @ Univ. of Cyprus (UCY)

• The Grid Authority in Cyprus (est. 2002)• TestBed: 72 CPU site, 1TB Storage Element SEE

Resource Broker, SEE Information Index,• 38,000 job submissions between Mar05–Jun06• Related Projects: EGEE (2004-current),

Healthware (2005-2008), gEclipse (2006-2008), eScience-CY (2004-2008), CoreGrid (2004-2008), Older: Emispher, CrossGrid, APART

9

Nicosia General Hospital (NGH)

• The largest (500-beds) and most technologically-advanced medical premise on the island.

• It covers a wide range of medical specialties.• Intensive-Care-Unit (ICU): 17 beds, each equipped

with a Phillips Intellivue Monitor, Blood Gas Analyzer, Mechanical Ventilation, Infusion Pumps, many other devices.

10

Bilateral Collaboration BenefitsBenefits for UCY • Apply the GRID technology in the local Life Sciences.• Develop Techniques for Temporal Data Management, build

libraries for the high performance analysis, storage and retrieval of real signals.

Benefits for NGH• Be the first operating ICGrid site (thus stimulate the

evolution of the system).• Build the required tools for: i) Early Diagnosis, ii)

Education and iii) Defining Early Warning Systems / Safety Thresholds (identify when a human life is jeopardy),

• Acquire technological know-how that will be necessary due to the imminent installation of the local Clinical Information System.

11

Overview

ALERTS

Query GUI

EGEE

Local

0 20 40 60 80 100−1.5

−1

−0.5

0

0.5

1

1.5

2

2.5

3

3.5 Applying the Minimum Bounding Envelope (MBE) to the Query

x di

men

sion

Time

1 1 25

20

1

LCSS(MBEQ

, A) = ( 1 + 1 + 1 + 25 + 20) = 48

MBEQ

Trajectory AQuery Q

12

Screenshots

13

Screenshots

14

Screenshots