Identifying Parameters Indicating Hypotension

On June 27th Rob published to partners a report on his finding during the Work Pack 1 research.  This is a summary of the report.

1 Introduction
The initial research carried out as part of WP1 (HypoPredict Parameter Identification) has been
focused on investigating various definitions currently being used for hypotensive events [Donald,
2008a,b]. This report details progress of the ongoing research using the group of
∼ 2000 events, that were identified in the earlier reports. This research is aimed at trying to identify parameters that appear to be associated with this group of hypotensive events.

Together the events and their associated parameters will form the initial “Training Database”
which is a starting point for WP2 (HypoPredict Engine Design).

The research has divided into two phases. An initial phase to look at event densities and
then a follow on stage to look for associated parameters. The first phase has been completed
however the second phase has become more complex and although started will continue to be
investigated as we build up the training database in WP2.

2 Investigative Method
For the first part of the investigation on event densities, I have continued to use the research
specific computer program called “Data Set Generator” (DSG), described in the Heidelberg report
[Donald, 2008a]. The main modifications were to introduce a system to allow the investigation of
inter event times and episodes of hypotensive activity.

For the second part of the research I have begun to develop another project specific computer
program called “TrainingData”. This program takes the event list produced by the DSG program
and will perform scans of the supporting tables looking for associated parameters.

5 Discussion

The results presented in this report are part of the ongoing research effort aimed at producing
a system that can provide some early warning of a hypotensive event. This report is the third
in a series which has documented the analyses carried out as the Avert-IT task “Work Pack 1”
Rob Donald: 2008-06-27 Page 9 of 11 Avert-IT Confidential Technical Report Associated Parameters Selection

( WP1). This task was officially entitled “HypoPredict Parameter Identification”. In fact the initial
two reports [Donald, 2008a,b], covering some four months of activity, detailed the considerable
work carried out in producing a defensible definition for what actually constitutes a hypotensive
event. This work was an essential first step and the insights gained in working with the clinical
teams and in examining the data within the BrainIT database should benefit the long term aim
of the project.

The last two months of WP1 have concentrated on looking for “associated parameters” in
preparation for progressing to the next stage of the research. This next stage will investigate
Bayesian neural network techniques applied to the task of classifying a set of medical data in-
puts as a possible precursor to a hypotensive event. This classification/modelling task is ex-
pected to involve highly nonlinear relationships which of course is the motivation for considering
a Bayesian neural network.

Although I have carried out some simplistic univariate and multivariate linear modelling in an attempt to establish relevant “associated parameters” these have not shown strong relationships which in a sense is to be expected. I feel that it is important however not to exclude apparently unrelated parameters from the next stage of the research and therefore I have begun to produce a custom computer program ( TrainingData ref: section 2) that can be easily modified to produce training sets which can incorporate as much of the BrainIT data as we would like to investigate.

Reblog this post [with Zemanta]
This entry was posted in Uncategorized and tagged , , , , , , , . Bookmark the permalink.