Trevor you are correct. The deterministic approach forces a continuous variable to be considered as a series of small increments each of which can have a different interpretation.  As you note, if we are considering the interpretation of two tests (say Free T4 and TSH) we could have 20 increments for each test and therefore 400 different combinations. This becomes even more complicated if we are dealing with three or more variables.  This approach is possible using "brute force" --- in our first attempt at interpretive reporting back in the late 1970s we went this way.  However, it becomes unmanageable.  The AI shell that I use (Acquire) provides tools for structuring and displaying the Knowledge Representation.

A curious observation when using such a tool is that it forces the "expert" to make a clear decision when presented with a specific combination of results.  I have found that in these borderline areas my interpretation actually changes depending upon whether I feel aggressive, confident, tired, etc.  Although the machine is no better than I am ... it is certainly more consistent.


Mike

Michael McNeely MD FRCPC
Vancouver BC Canada
[log in to unmask]

-----Original Message-----
From: [log in to unmask]
[mailto:[log in to unmask]]
Sent: December 13, 2001 8:19 AM
To: [log in to unmask]
Subject: Re: a challenge for interpretative comments ....


Mike,

My understanding of a deterministic system is that probability is always 1
or 0 that is there is no uncertainty.

Many decisions are best made in a series of 'either/or' steps and, as such,
each step is deterministic (just as medical treatments are all or none).

A deterministic approach could certainly improve the order in which data is
collected and in many instances negate the need for collection in the first
place.

It seems to me that steps within a decision tree are guided by the
evaluation of risk (together with moderation by reward to the individual)
and that risk may best be described by the net outcomes of many traverses of
the decision tree.

I should be interested to learn how a deterministic approach deals with the
problem of continuous variables, especially in combination because this
would appear to increase the size of the arrays that are necessary for every
result combination.

Trevor



> -----Original Message-----
> From: McNeely, Michael [SMTP:[log in to unmask]]
> Sent: 13 December 2001 15:05
> To:   [log in to unmask];
> [log in to unmask]
> Subject:      RE: a challenge for interpretative comments ....
>
> Trevor
>
> What you are describing is the situation where an expert system is based
> upon some mathematical algorithm (for example, sensitivity and
> specificity).  However, these do not work well (as you point out) because
> the "a priori" probability of the disorder for a given patient is rarely
> (or never) known. Production Rule based systems and Neural Nets have the
> serious problem that we can never be certain that they won't come up with
> some unexpected (and absurd) result.  The most secure approach is
> "Deterministic" where every combination of results leads to a specified
> comment.  This creates a large number of potential combinations that must
> be reviewed but with a clever "AI Shell" this task can be managed.  The
> Deterministic approach is more intuitive for the Expert who must manage
> the system for it emulates the way the Expert actually produces
> interpretations "manually".
>
> Mike
>
> Michael McNeely MD FRCPC
> Vacouver BC Canada
> [log in to unmask]
>
> -----Original Message-----
> From: [log in to unmask]
> [mailto:[log in to unmask]]
> Sent: December 13, 2001 2:00 AM
> To: [log in to unmask]
> Subject: Re: a challenge for interpretative comments ....
>
>
> Obviously one requires data in order to define a set but so often we fail
> to
> use that data.
>
> If, for example, one knows the value of a urea one can define a set
> bounded
> by that urea value plus or minus the measurement uncertainty (say the sum
> of
> the analytical and biological variations).
>
> The new set may or may not differ in proportionate composition from the
> parent set from which is was derived (that is there may or may not be
> information gain) but it is the new set that defines prior probability.
>
> Extracting relevant sets is not that difficult provided the data is
> properly
> structured and the process can be made very fast.
>
> Most of the expert systems that I have seen described are, in truth,
> little
> more than risk algorithms using pre-defined data alone.
>
> Trevor Tickner
>
> > -----Original Message-----
> > From: Dr Henry Chandler [SMTP:[log in to unmask]]
> > Sent: 12 December 2001 19:42
> > To:   [log in to unmask]
> > Subject:      Re: a challenge for interpretative comments ....
> >
> > In message <[log in to unmask]
> > norwich.thenhs.com>, TICKNER TREVOR (RM1) Norfolk and Norwich NHS Trust
> > <[log in to unmask]> writes
> > >
> > >Similar objections can be raised to many 'expert systems'. Only rarely
> do
> > >such systems make any attempt to divide patient into prior-data sets
> > except
> > >where that data is known to correlate with diagnostic assignment and
> that
> > >'gold standard' assignment is no more than the opinion of an expert
> > panel.
> > >Trevor Tickner
> > >Norwich
> >
> > The trouble in designing "expert systems" that divide patients into
> > prior data sets is that you need a test to make that division which, of
> > course, has a likelihood ratio ...
> >
> > Do I hear the Reverend Dr Bayes thundering up behind me with a problem
> > of ever decreasing circles, but again perhaps Pascal was right.
> > --
> > Dr Henry Chandler
> >
> >
> --------------------------------------------------------------------------
> -
> The information contained in this e-mail is confidential and is intended
> only for the named recipient(s). If you are not the intended recipient you
> must not copy, distribute, or take any action or reliance on it. If you
> have
> received this e-mail in error, please notify the sender. Any unauthorised
> disclosure of the information contained in this e-mail is strictly
> prohibited.
>
>
>
>
>
---------------------------------------------------------------------------
The information contained in this e-mail is confidential and is intended
only for the named recipient(s). If you are not the intended recipient you
must not copy, distribute, or take any action or reliance on it. If you have
received this e-mail in error, please notify the sender. Any unauthorised
disclosure of the information contained in this e-mail is strictly
prohibited.