Associated Parameter Selection – by Rob Donald

Abstract
This discussion paper gives an description of the techniques that will be used to identify the
associated parameters to be included in the Bayesian Artificial Neural Network (BANN) train-
ing database.

1 Introduction

The initial research carried as part of WP1 (HypoPredict Parameter Identification) has been fo-
cused on investigating various definitions currently being used for hypotensive events [Donald,
2008a,b]. We now have a group of ∼ 2000 events that will be used in the next stage of the re-
search which is to identify parameters which appear to be associated with these events. Together
the events and their associated parameters will form the “Training Database” which is a starting
point for WP2 (HypoPredict Engine Design)

2 Event Density

From the research into the event definition which resulted in the group of events that we are us-
ing, it is apparent that the events are often grouped together in “Episodes” of hypotensive activity.
An analysis of the characteristics of these episodes will be carried out. This analysis will be used
to decide on the timespan that will be used in calculating the derived parameters. Generation
of “Episode” data will be investigated by varying the parameter “Inter Event Time” (IEvT) which
will be varied from 0 to 30 mins in steps of 5 mins. This range being chosen as sensible given the
duties of clinical staff in the intensive care unit (ICU).

3 Associated Parameters

The BrainIT database contains supporting information in six tables. For each event I will be scan-
ning through these six tables looking for supporting data around the time an event started.
The issue of missing data will certainly arise during this phase of the research. I intend to
attempt to record a value for all of the measurements that have some data represented in the
BrainIT database. This will undoubtedly result in certain fields have a percentage of missing
values. However by at least attempting to fill in the fields, e.g. ETCO2 from the Physiological
table, we will gain a quantitative understanding of the missing data problem and we can then
decide if we want to use an imputation technique to fill in the data or reject events which do not
have complete supporting data sets.

3.1 Physiological Table

3.1.1 Measured Parameters
In my previous reports I have noted that the parameters ICP, CPP, TC, HRT, SaO2 from the Phys-
iological data table appear to be available for most events whilst the coverage for RR, CVPm and
ETCO2 are less often available. As described above, I will record the values for each of these read-
ings at the appropriate times (e.g. -15,-30,-45 mins) before the event occured.

3.1.2 Derived Parameters
For each of the measured parameters from section 3.1, I will calculate the following derived pa-
rameters over some previous timespan that comes out the research in section 2:
• Mean
Standard deviation
• Slope coefficient from linear regression
• R2value as a measure of linear regression fit

3.1.3 Principal Components Analysis

From section 3.1.2 on derived parameters, it can be seen that a large number of parameters will
be generated for each event. I will use a standard statistical technique called Principal Com-
ponents Analysis (PCA) to combine these parameters into a smaller number of representative
indices (typically 2 or 3). These PCA indices will then also form part of the training database for
each event.

3.2 Demographic Table

In contrast to my earlier thoughts on a combined injury score, I have decided to use the nine
“injury_*” parameters as individual binary variables and let the model decide on the weightings
that contribute to event prediction.

I will be including as many as possible of the fields which contain data e.g. age, sex, type of
trauma, influence of alcohol, CT scan classification, etc. I will also be deriving some values from
data contained in the Demographic table e.g. time from trauma, time to NSH, etc.

3.3 Event Data

There are four event data tables, TargetTherapies, LabResults, NeurologicalStatus and
OtherClinicalEvents. I will be trying to match up treatment data from these tables to identify
associations. This will require some investigation to see how far pre or post event a particular
measurement can be considered as having a possible effect. I will need the advice of the clinicians
on this question.

Bibliography
Rob Donald. Event definition analysis, preliminary results. Technical report, Avert-IT Project,
2008a.
Rob Donald. Event definition analysis using EUSIG hypotension definitions. Technical report,
Avert-IT Project, 2008b

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