- From: Kashyap, Vipul <VKASHYAP1@PARTNERS.ORG>
- Date: Wed, 27 Sep 2006 06:57:53 -0400
- To: <helen.chen@agfa.com>
- Cc: <public-semweb-lifesci@w3.org>
- Message-ID: <2BF18EC866AF0448816CDB62ADF6538104C167C1@PHSXMB11.partners.org>
Here's a potential solution, though I am not completely happy with this
solution.
Define a state called "TwoOrMoreConditions" and define the following rules:
IF Temp > 100F AND Pulse > 100 THEN TwoOrMoreConditions = true
IF Pulse > 100 AND Decreased Breath Sounds THEN TwoOrMoreConditions = true
IF Decreased Breath Sounds AND Rales THEN TwoOrMoreConditions = true
....
Then define a rule which checks for this state to be true:
IF "TwoOrMoreConditions" is true THEN order an X-Ray
This could be represented as an OWL definition
You may want to start of with defining the following XML Schema datatypes
TempGreaterThan100
PulseGreaterThan100
RespiratoryRateGreaterThan100
and the following OWL definitions
equivalentClass(PatientWithDecreasedBreathSounds, intersectionOf(Patient,
Restriction(breathingIntensity, hasValue(reduced))))
equivalentClass(PaitentWithRales, intersectionOf(Patient, Restriction(condition,
hasValue(Rales))))
equivalentClass(TwoOrMoreConditions, unionOf(intersectionOf(TempGreaterThan100,
PulseGreaterThan100),
intersectionOf(PulseGreaterThan100, PatientWithDecreasedBreathSounds)
....)
DL Reasoners support datatype reasoning....
The problem with this approach is the combinatorial complexity of authoring the
rule base...
So for this case, you would need 5C2 rules, i.e., 10 rules to model this...
Another approach to reduce the rule authoring load is to define a "macro"
Atleast K are true
and then expand this macro behind the scenes. Of course need to understand
This is one of the motifvations for going to statistical/linear programming
approaches.
Feedback, suggestions welcome.
---Vipul
________________________________
From: helen.chen@agfa.com [mailto:helen.chen@agfa.com]
Sent: Wednesday, September 27, 2006 5:05 AM
To: Kashyap, Vipul
Cc: public-semweb-lifesci@w3.org; public-semweb-lifesci-request@w3.org
Subject: RE: CAP use case - Reasoning on Weighted Condition and Fuzzy Reasoning?
Hi, Vipul
Strongly agreed with two points you made here.
The problem, however, is first on the modelling side of such "weighted
conditions". In the example,
> 3: "Obtain chest X-ray, especially if patient has two or more of these
> signs:
> Temp > 100F
> Pulse > 100
> Decreased breath sounds
> Rales
> Respiratory rate > 20
these 5 signs are all indicative to a diagnosis of pneumonia, but carry
different weights in decision making. The weight you assign to a symptom here
is local to the context (suspected pneumonia). When multiple guidelines are
taking into consideration to form a pathway for a patient, these weights will
likely to change, isn't it? How your KB handle it?
Kind regards.
Helen
"Kashyap, Vipul" <VKASHYAP1@PARTNERS.ORG>
Sent by: public-semweb-lifesci-request@w3.org
09/26/2006 05:42 PM
To
Helen Chen/AMPJB/AGFA@AGFA, <ogbujic@bio.ri.ccf.org>
cc
<public-semweb-lifesci@w3.org>
Subject
RE: CAP use case - Reasoning on Weighted Condition and Fuzzy Reasoning?
This is a very good use case which can easily be done using the
statistical/linear programming approach
where you create equations out of various decision variables The two downsides
of this approach are as follows:
1. Lack of explanation capabilities: A key feature for clinical decision
support is that physicians like to get
explanations for the recommendations proposed by the system.
2. Lack of "knowledge visibility": The biggest downside is from the KM
perspective, what if one of the conditions changes? We need this to be visible
so that we can have KM processes handle these changes.
As far s handling fuzzy situations using rule bases, a "hackers solution" would
be to have specialized global variables keeping
track of the number of conditions matched, etc Of course some variant of this
has been implemented in the MYCIN type of
expert systems.... one needs to tread carefully:
BTW, current rule engines support rule priorities and one could model the most
specific rule as having higher priority than others, for instance
IF A1 and B1 Then C1 will have a higher priority then
IF A1 Then C2 or
IF B1 Then C3....
Of course these requirements need to make it into SW standards ...
=======================================
Vipul Kashyap, Ph.D.
Senior Medical Informatician
Clinical Informatics R&D, Partners HealthCare System
Phone: (781)416-9254
Cell: (617)943-7120
http://www.partners.org/cird/AboutUs.asp?cBox=Staff&stAb=vik
<http://www.partners.org/cird/AboutUs.asp?cBox=Staff&stAb=vik>
To keep up you need the right answers; to get ahead you need the right questions
---John Browning and Spencer Reiss, Wired 6.04.95
________________________________
From: public-semweb-lifesci-request@w3.org
[mailto:public-semweb-lifesci-request@w3.org] On Behalf Of helen.chen@agfa.com
Sent: Tuesday, September 26, 2006 11:05 AM
To: ogbujic@bio.ri.ccf.org
Cc: public-semweb-lifesci@w3.org
Subject: Re: CAP use case - Reasoning on Weighted Condition and Fuzzy Reasoning?
Hi, Chimezie
Yes, let's discuss in detail of possible approaches at our F2F meeting next
week.
I was also considering something similar to your following proposal. But one
obvious drawback of this approach is that the weights you calculated or assigned
are very much local context dependent, also, could lead to non-monotonic
characterises of your KB (i.e. adding new facts could change your weights
assigned to variables). This could seriously compromise the benefit of SW
technology in this area.
Helen.
>The result of the targetted analysis is a multi-variable risk factor
>equation (with a very high level of predictive accuracy) that takes:
>- a set of weights for each variable (the weights are 'built-in' to the
equation)
>- raw patient data
>The equations result in outcome plots that indicate the likelyhood of
>survival (or the resumption of a particular symptom, effect on ability to
>work, etc..) at some point in time (by a percentage).
>
Chimezie Ogbuji <ogbujic@bio.ri.ccf.org>
09/26/2006 09:49 AM
To
Helen Chen/AMPJB/AGFA@AGFA
cc
cebarr01@yahoo.com, aziz@boxwala.com, sam.brandt@siemens.com,
THONGSERMEIER@partners.org, davide@landcglobal.com, DAN.RUSSLER@oracle.com
Subject
Re: CAP use case
On Tue, 26 Sep 2006 helen.chen@agfa.com wrote:
> Notice in the step 3, it says:
>
> 3: "Obtain chest X-ray, especially if patient has two or more of these
> signs:
> Temp > 100F
> Pulse > 100
> Decreased breath sounds
> Rales
> Respiratory rate > 20
>
> Now we are facing the new problem of modelling "two or more" facts of a
> necessary condition for "order chest X-ray" in the knowledge base.
This is definately a problem, I can't see how you would model that using
either qualified cardinalities in DL or a rule function - most of the
examples where I've seen counts within a rule LHS involves counts of
items in an RDF collection.
Perhaps we should consider having a liaison with the Rule Interchange
Group (http://www.w3.org/2005/rules/wg) for requirements such
as these? I think they would benefit from the specific usecase and we
would benefit from the additional expertise.
> Furthermore, doctors will likely tell you that no only they need to
> express "at least two or more", they also want to express "fact A carries
> more weight or more indicative to a diagnosis than fact B". If we were to
> model these "weighted condition", we are opening a whole can of new worms,
> and I don't think any SW reasoners now can do reasoning on this.
Definately, can't manage uncertainty or do any fuzzy reasoning in SW.
However, there is an alternative approach to managing uncertainty that I
was hoping to discuss during the Fact-to-Fact, is mentioned in the CABG
usecase, and is the way we go about clinical research here.
Primarily we conduct targetted studies coordinated between our
biostatisticians and resident physicians. The physicians identify the
relevant data points that they believe are primary factors in a
particular clinical pathway and the statisticians are responsible for the
statistical merits of the study (minimize noise, ensure all the
relationships between the variables are covered, etc..).
The result of the targetted analysis is a multi-variable risk factor
equation (with a very high level of predictive accuracy) that takes:
- a set of weights for each variable (the weights are 'built-in' to the
equation)
- raw patient data
The equations result in outcome plots that indicate the likelyhood of
survival (or the resumption of a particular symptom, effect on ability to
work, etc..) at some point in time (by a percentage).
Such an approach limits the uncertainty factors and weights to the
'black-box' equation - which results from a targetted statistical / domain
analysis - such that the remaining pattern matching can be handled by a
rule-based system. The suggestion is that factors of uncertainty are
better managed as the result of a targetted (and therefore responsible)
statistical analysis that results in a mathematical model than as part of an
adaptable clinical pathway or protocol. The caviat ofcourse is that the
rule-system the adaptable
clinical pathway & protocol is built on must support a logic that includes
functions in it's syntax.
Chimezie Ogbuji
Lead Systems Analyst
Thoracic and Cardiovascular Surgery
Cleveland Clinic Foundation
9500 Euclid Avenue/ W26
Cleveland, Ohio 44195
Office: (216)444-8593
ogbujic@ccf.org
Received on Wednesday, 27 September 2006 10:58:16 UTC