Background and issues… Neural nets are based
on the CAS of the brain; how the human mind functions. Can neural
net technology improve health care decision making? Can neural net
technology be used by normal clinicians and managers to improve day-to-day
decision making?
Can neural net (NN) technology improve health care
decision making?… The "Dooley criteria" for use of neural nets
over other decision support methods... If health care decisions can
be described as: (1) a finite number of categories that represent
the "answer," (2) a complex set of characteristics that change across
the categories, and (3) there exists a good amount of data to teach
what belongs in the categories... then neural nets is a good approach.
Neural nets are overkill if the pattern is linear... other traditional
methods are better for linear situations. If the "answer" is a continuous
variable rather than a category, non-linear statistical methods are
just as good.
Suggested area where NNs might be useful: predicting
readmission rates in CHF population, cardiology, cardiac surgery,
OB.
We have examples of Dooley's three criteria…
Plsek post... A good illustration of Kevin's
three conditions is the application of neural nets to the problem
of recognizing handwriting. By analogy, we can extend this to the
healthcare application of mammography reading.
(1) finite number of discrete "categories" that
represent "the answer." In handwriting recognition there are only
26 letters, along with numbers and a few other special marks. Every
character written has to be one of those. The question is "given
the character in front of me, which one is it?"
So, in mamography reading are there a finite number
of "answers" that one can give? A good indication of this would
be if there exists automated dictation report generators for it.
All the doc has to do is say a keyword and prewritten text appears
on the report. I know that these exist for some areas of radiology
and pathology.
(2) complex set of characteristics which change
across categories. In handwriting, the details of just exactly how
I make an "a" and how someone else makes an "a" can be amazingly
complex and subtle. I may even have trouble making an "a" exactly
the same twice. But, still, there is something recognizable about
an "a" versus a "b"
In mamography reading, how much skill is involved
in discerning the patterns that lead to a conclusion? Is it a pretty
simple task; seen one seen 'em all? Or is there lots of nuance?
(3) good amount of data to teach it what belongs
to each category. The neat thing about neural nets is that they
are not "programmed" by an expert. They actually learn in much the
same way that a child learns (or a med student!). In handwriting
analysis, we can feed the neural net 1000 examples of what the letter
"a" looks like when written by many different people. This is referred
to as "training" the neural net.
In mamography reading, could someone assemble many
different examples of each of the possible outcome-answers to be
used as a training set for the neural net? An indicator of this
might be to look at medical education in a given area. To what extent
do we actually train young docs by basically giving them lots of
examples and telling them when they get it right and when wrong?
Steve Larned post... I mentioned at our Princeton
meeting that we have met with George Davis and Walter Lowell of
the Augusta Mental Health Institute (AMHI), who have used ANNs to
model psychiatric LOS. This problem, and many clinical problems,
meet Kevin's criteria:
1. finite number of discrete categories representing
the "answer" or "output" (in their example, LOS in the state hospital)
2. complex set of characteristics which change across categories
(in their example diagnostic categories, medication on admission,
treatment in the hospital, medication in the hospital, admitting
physician, community to which pt discharged, discharge medication,
family milieu therapy, etc.) 3. sufficient data (in their example,
the entire database pertaining to patients admitted at AMHI in recent
years).
And we also have more explanation of NN technology…
Dooley post... Let me be more specific.
A NN is a classic "black box". It takes numerical inputs and turns
them into a single (typically) numerical output. People are often
wowed by such magic. Like other black boxes though, it suffers from
two typical problems: you cannot figure out WHY it works (if it
does), and if you put the "wrong" data in, you can't get a good
answer out (garbage in, garbage out).
The input to the NN is called the feature vector.
The choice of features that you choose is paramount--and this task
is often shortchanged. You can think of the output as a diagnosis.
Before I get into mechanics, let me clarify the
"numbers" issue: even though the output of a NN is a number, NN's
in general are not nearly as good at predicting numerical outcomes
as they are at predicting classification-type outcomes. If a NN
is used for numerical prediction, one is essentially doing nonlinear
regression (and one would be better off doing that explicitly).
Let me give three parallel examples:
Outputs/Diagnoses/Predictions: Case A--Does the
patient have cancer or not? Case B--Should a low, medium, or high
dosage of medicine be given? Case C--Is this EEG "normal"?
As I said, the output of the NN is a number, between
0 and 1. We need a simple interpretive filter that changes the number
into a category. Call the number Y:
Case A--If Y<0.5, cancerous; if Y>0.5, not. Case
B--If Y<0.33, low; .33<.66 medium, otherwise high Case C--If
Y<0.5, normal; if Y>0.5, abnormal
This is what I meant by "finite number of discrete
categories".
Now onto the feature vector. These must be numbers
also. You can have a lot of features. But the more you use, the
longer you have to train the NN.
For case A, what I could do is take an X-ray and
represent it as a rectangular grid of pixels. Each pixel would then
be assigned a "gray scale" number representing the particular shade:
0 would be completely white, and 1 would be completely black, and
numbers in-between would represent different shades of grayness.
A 12"x12" x-ray may have thousands of pixels. Thus, the feature
vector is very long. NN tend to be very good at such "picture recognition".
For case B, I would have to use my expertise to
come up with "features" that matter. Let's say I have (a) white
blood cell count, (b) results from yes/no test A, (c) temperature,
(d & e) blood pressure. Obviously (a), (c), and (d,e) are numbers
already. They simply need to be scaled so that they range between
0 and 1 (no problem). Let (b) be 0 if the test is negative, and
1 if the test is positive. Our feature vector is size 5.
For case C, I have a time series of data. What I'll
need to do is somehow characterize that into relevant features.
I might take as features (a) the mean, (b) the std. deviation, (c)
kurtosis, (d1-d10) the first 10 values of the auto-correlation function,
(e) the value of the Lyapunov exponent, (f1-64) the values of a
64-resolution power spectrum, etc. My feature vector would be quite
large.
Obviously, in each case, the features allow me to
predict the outcome. When I said "wide range of features that vary
across outcome", what I meant was (a) if they don't vary, then outputs
cannot be associated with inputs, and (b) if their relationship
is simple rather than complex, a NN is overkill.
The final step is training. I need to feed the NN
several hundred cases in order for it to learn how to predict. Each
case in a feature vector and the corresponding "right answer" (output).
After it learns, then I can feed it a feature vector (a new case)
and it will predict the answer.
Examples of Neural Net Technology Use to Improve
Healthcare Decision Making… This is a topic that is already being
explored in healthcare. There are several examples. Our role might
be simply to review the literature here. Some examples and published
papers are cited below. Could probably get more with a literature
search.
- As mentioned in Steve Larned's post above, The
Augusta Mental Health Institute has used NNs to model psychiatric
LOS.
- Mammogram reading (company: Natural Selections
reports that ROC area under curve was >90%; good for a diagnostic
test). Reference: Fogel DB, Wasson EC, Boughton EM, Porto VW.
A step toward computer-assisted mamography using evolutionary programming
and neural networks. Cancer Letters 119(1997), 93-97.
- Locating cardiac hypertrophy on the basis of ECG
information alone. This is difficult for expert human diagnosticians
to do. Reported results are an "Overall diagnostic accuracy rate
of about 70%… We consider the neural network superior both to computer-based
electrocardiography and to a well-trained cardiologist." Reference:
Ouyang N, Yamauchi K. Using a neural network to diagnose the hypertrophic
portions of hypertrophic cardiomyopathy. MD Computing 15(2), 1998,
106-109.
- Determining objective diagnostic probabilities
of insulin-dependent diabetes mellitus. Reported a 9% better prediction
accuracy when compared to classical statistical methods (regression).
Notes "…our findings corroborate other research showing that neural
networks perform better than classical statistical methods." (5
references given to support this) Reference: Armoni A. Use of
neural networks in medical diagnosis. MD Computing. 15(2), 1998,
100-104.
- Some aspect of COPD or diabetes (company: MedAI,
subsidiary of Adventist Health; demonstrated NN was better than
traditional multiple regression at predicting patients at risk for
hospitalization; Al Herzog and Jacquie Byers getting more info).
This may be the same info as the two articles cited previously.
- Outside healthcare… There is a story on CNN about
a genetic algorithm based computer system at the Vovlo truck plant
in Dublin, Virginia. The web version of the report is on the CNN
website (where you can also download the entire 2 minute, 30 second
report on VXtreme video). Here is the URL for the story: http://www.cnn.com/TECH/science/9802/26/t_t/artificial.intelligence/
(Note: if you're typing that in by hand, there is an underscore
character between the two t's as in ....26/t_t/artificial.....)
- Holland's new book has good material on NNs...
Potential Demonstration Project… Steve Larned
at Maine Medical Center has purchased NN software called "BrainMaker"...
will make available to researchers in cardiology, cardiac surgery,
OB... hope to build clinical "flight simulators" to explore non-linear
connections between observed clinical end-points and factors in clinical
data bases. They also hope to involve Gerry O'Connor and the Northern
New England Cardiovascular Study Group in this project.
Ultimate Goal of This Line of Thinking… Improved
clinical diagnostic and "prediction" tools based on NN technology;
leading to better care and better outcomes for patients
.
