Federating Distributed Clinical Data for the Prediction of Adverse Hypotensive Events

1
Federating Distributed Clinical Data for the
Prediction of Adverse Hypotensive Events

• Anthony Stell, National e-Science Centre
• AHM 2008, Edinburgh, UK

2
Overview

• Adverse hypotensive events
• Current monitoring systems
• Avert-IT Project
• Threshold definitions
• Avert-IT centre definitions
• EUSIG definitions
• Data grid
• Hypo-Predict engine
• Progress
• Future work

3
Adverse hypotensive events

• Abnormally low blood pressure
• Various numerical definitions
• BPs gt 100 5mins
• BPm gt 70 5mins
• Etc

4
Current monitoring systems

• Odin browser/monitor
• Database browser of information about secondary
  insults following head injury
• BioSign
• Produces index predicting cardiovascular
  instability based on several vital signs (Heart
  rate, respiration rate, etc.)
• Philips Event Surveillance Monitor
• Manual correlation of patient parameters into
  discrete events

5
Avert-IT
No single system that can predict the onset of a
hypotensive event over a useful timescale (e.g.
half an hour in advance)
6
Avert-IT Consortium
Grant number FP7-217049 1,780,000 over 3 years

• Technical
• C3 Amulet, Dingwall, UK
• National e-Science Centre, Glasgow, UK
• Clinical
• Southern General Hospital, Glasgow, UK
• Uppsala University Hospital, Sweden
• University of Heidelberg, Germany
• Ospedale San Gerardo, Monza, Italy
• Kaunas University of Technology, Kaunas,
  Lithuania
• Universidad Autonoma de Barcelona, Barcelona,
  Spain
• Philips Medizin Systeme Boblingen GmbH,
  Boblingen, Germany
• Administration
• PERA Innovation Ltd, Bellshill, UK

7
BrainIT

• http//www.brainit.org
• An internet-based group set up to work
  collaboratively on standards for the collection
  and analyses of data from brain-injured patients
• Avert-IT builds on the work of BrainIT
• Many causes of events but naturally focus will be
  on those related to Traumatic Brain Injury (TBI)

8
Threshold definitions by Avert-IT centre
9
Threshold definitions EUSIG

• Edinburgh University Secondary Insult Grade
  (EUSIG)

10
Data grid 1 architecture

• Lightweight client at clinical end
• Parses output data, packages in SOAP
• Connects to grid provider through WS

Data grid provider deconstructs WS and uploads to
central repository
End developer connect to this database and use to
populate the user interface and the BANN
11
Data grid 2 clinical end-points

• Many different formats
• Philips DocVu (ASCII text)
• BrainIT (Access database)
• CMA ICU Pilot (XML)
• Health-Level 7 (HL7)
• Parameters used so far
• Blood pressure (BPs, BPd, BPm)
• Intra-cranial pressure (ICPm)
• Cerebral perfusion pressure (CPP)
• Heart rate (HRT)
• Temperature (TC)
• Blood oxygen saturation (SaO2/SpO2)

12
Data grid 3 security and performance

• Security
• Encryption of data in transit and at rest
• Locally linked patient identifiers
• Contextual demographic information
• Episodic activity could allow inference
• Need feedback mechanism to capture consent
• Performance
• Database optimisation
• Records for a single patient over 2 days number
  8000 for BrainIT, 6000 for ICU-Pilot
• Need smart storage solutions
• WS thresholds
• May need other connection options?

13
Hypo-Predict engine

• Decision support tools
• Look-up tables, case-based reasoning (CBR),
  genetic algorithms (GA), Bayesian belief networks
  (BBN), artificial neural networks (ANN)
• Bayesian approach to ANN (BANN)
• Accounts for probabilities of causative effects
• Experience within BrainIT of setting up BANNs
• Tim Howells (Uppsala), author of Odin software
• Train the Hypo-Predict engine on unified data
• then turn into a product and give to individual
  centres.

14
Progress

• Clinical research
• Threshold definition agreed upon (lt 905mins)
• Parameter list for data established
• Implementation
• Parsers for four out of the seven clinical
  sources
• BrainIT, Glasgow, Heidelberg, Monza
• Connections established between end-developers
  (C3) and data grid provider (NeSC-Glasgow) using
  JDBC
• User interface prototypes developed

15
Future work

• Distribute parsers and automate WS
  connections/uploads (subject to validation)
• Involves negotiating with sysadmins at various
  clinical centres gt takes ages
• Development of the BANN on the data available
• BrainIT database already available
• Example sources available for centres already
  parsed (still awaiting full, real-time data)
• Consent capture loop to be implemented
• Dissemination of work and Hypo-Predict product

16
Avert-IT Resources

• Website http//www.avert-it.org
• Wiki http//wiki.avert-it.org/wordpress
• Dissemination
• Wiki - http//frontofficebox.com/knowledgebase/ind
  ex.php?titleAvert-IT_Portal
• Blog http//avertit.wordpress.com
• Network http//avertit.ning.com
• Contact
• Steve Reeves (C3 Amulet)
• Ian Piper (Southern General Hospital)
• David Keirs (PERA)
• Richard Sinnott, Anthony Stell, Jipu Jiang (NeSC)

17
Demonstration

• At 10.05am on Thursday morning
• Avert-IT
• EuroDSD
• VOTES / Vanguard