[IMPROVEDX] IOM report is released - Diagnosis in actual practice
Amos Cahan
acahan at US.IBM.COM
Thu Oct 15 20:03:25 UTC 2015
If you base the decision to treat PE on a positive angiogram, then for
every 100 angiograms you perform, you will end up treating a false positive
case. So- to justify doing this test in the first place, you need to have
at least 5 patients with PE for every 100 patients examined (this will give
you approximately 1 true positive for every false positive result). This
means you can’t avoid the question of prior probability even if you want
to, because patients do not present to the ER with their angiogram. A
physician has to refer them to the test. And such referral may only be
justified if the prior is at least 5% (and at least in my opinion, this
number is too low). How do you decide when an angiogram is needed?
Amos
Amos Cahan, MD
Research scientist, Clinical Informatics
IBM T. J. Watson Research Center
1101 Kitchawan Road, Route 134
Yorktown Heights, NY 10598
1.914.945.2590
acahan at us.ibm.com
From: "Jain, Bimal P.,M.D." <BJAIN at PARTNERS.ORG>
To: IMPROVEDX at LIST.IMPROVEDIAGNOSIS.ORG
Date: 10/15/2015 02:36 PM
Subject: Re: [IMPROVEDX] [IMPROVEDX] IOM report is released - Diagnosis
in actual practice
If we know that out of 100 patients with pulmonary embolism, 21 have
positive chest CT angiogram, we can say the frequency of positive angiogram
in this group or the probability of positive angiogram in a patient in this
group is 21 percent. This is a statement of likelihood. If , in addition,
we know that out of 100 patients without pulmonary embolism only 1 has
positive angiogram , we have a likelihood ratio of 21. You are right,
likelihood or likelihood ratio is not expressed in terms of odds.
Let us now calculate the probability of pulmonary embolism in a patient
with positive angiogram. For this, we need to know the prior probability of
pulmonary embolism in this patient which is derived from the distribution
of embolism in a group of patients like him. Suppose it is 5 percent as the
presentation is highly atypical. We calculate the posterior probability of
embolism to be 52 percent.
In the probabilistic approach, we are asked to consider this posterior
probability as evidence for embolism in this patient and diagnose it on its
basis. From it , we should diagnose embolism to be indeterminate in this
patient. In practice, we diagnose embolism in this patient with near
certainty in this patient from the strong evidence furnished by likelihood
ratio of 21 alone. In fact, we diagnose embolism with near certainty from
positive angiogram in any patient regardless of prior probability as the
likelihood ratio remains 21 in all patients.
On the other hand, if we are epidemiologists and wish to study the
distribution of pulmonary embolism in the defined group of patients
similar to this patient, the posterior probability of 52 percent would be
evidence which we would correctly use to determine this distribution.
For us clinicians, diagnosis is local, so to speak, as it is performed in a
given individual patient as opposed to epidemiology which is global being
done in a large group or population. The point I have raised is that
probability is evidence only at a global not a local level.
The notion of likelihood is simple but subtle which escaped the notice of
mathematicians of the rank of Gauss and Laplace. It took the genius of
Ronald Fisher to isolate it. It is employed in all forms of statistical
inference including the probabilistic approach. The only problem with this
approach is it is combined with a distribution in the form of a prior
probability making this approach unsuitable locally in a given individual
patient in diagnosis.
Bimal
From: Ely, John [mailto:john-ely at uiowa.edu]
Sent: Wednesday, October 14, 2015 10:59 AM
To: Society to Improve Diagnosis in Medicine; Jain, Bimal P.,M.D.
Subject: RE: [IMPROVEDX] [IMPROVEDX] IOM report is released - Diagnosis in
actual practice
If I have a room with 100 people complaining of sore throat in it and 20 of
them have strep throat and I select one person randomly from that room, the
probability that that person has strep throat is 20%. So in that sense,
the probability in the population is the same as the probability in the
individual.
I think we use the word “likelihood” to mean the same thing as probability.
For example, a likelihood ratio is the ratio of two probabilities (not two
odds, not two frequencies, but two probabilities). Also, The American
Heritage Dictionary of the English Language defines likelihood as “The
state of being probable; probability.”
John Ely, MD
University of Iowa
From: Jain, Bimal P.,M.D. [mailto:BJAIN at PARTNERS.ORG]
Sent: Tuesday, October 13, 2015 3:14 PM
To: IMPROVEDX at LIST.IMPROVEDIAGNOSIS.ORG
Subject: Re: [IMPROVEDX] [IMPROVEDX] IOM report is released - Diagnosis in
actual practice
In response to Dr. Brush, when a CPC discussant says ‘ the most likely
diagnosis is ‘ he does not mean ‘ the most probable diagnosis is’ in the
sense of a disease having the highest frequency in a series or group of
similar patients. What he means is that a disease has the greatest
likelihood in that the given presentation or a test result has been
observed to have the highest frequency in this disease compared to other
diseases. He makes a statement of likelihood, not one of probability. The
two concepts are completely different as a disease is fixed in a likelihood
but is a random variable in probability. Despite this difference they are
often mistaken for each other. Even the great Ronald Fisher thought
likelihood was a probability when he first introduced it in 1913. However,
he soon realized his mistake and described its properties clearly as we
them today in his landmark paper Mathematical Foundations of Statistics in
1922.
If we look carefully at the process of diagnosis in CPCs as I have done in
50 CPCs, we note that it proceeds by employing likelihood as evidence. An
extensive differential diagnosis is initially created and the likelihood of
each disease is estimated from given data. The disease with the highest
likelihood is then put forth as the most likely diagnosis. One would think
that if ‘most likely’ is the same as ‘most probable’ as Dr. Brush suggest,
then the phrase ‘ most probable’ would be used, if he is correct in at
least a few among the hundreds of CPCs that have been conducted. But this
phrase has never been used, to the best of my knowledge in any CPC.
Furthermore, a prior probability is not estimated and used as prior
evidence, a hallmark of the probabilistic approach in any7 CPC. This is not
done, I suggest, because the prior probability only indicates the
distribution of a disease in the population from which the given patient
has been drawn. As this distribution has no bearing as evidence for or
against a disease in the given patient, a prior probability is not
estimated.
There is universal agreement that probability is highly accurate in
representing evidence in terms of frequencies or averages in populations or
large groups of similar objects or persons.. Therefore it has been
extensively employed in areas such as epidemiology, life insurance
business, stock portfolio management, statistical mechanics. Incidentally,
some excellent comments on application of probability are to be found in
writings of some great figures in statistical mechanics which include
Maxwell, Boltzmann, Gibbs, Ehrenfests, Einstein. It is in employing
probability to represent evidence in a single, given instance that problems
arise. Here is Charles Sanders Peirce weighing in on this issue. ‘Any
individual inference must be either true or false.....in reference to a
single case, considered in itself, probability can have no meaning.’
There are at least four other methods of statistical inference besides the
probabilistic approach in which evidence is represented by a measure other
than a probability. These are:
1. Fisher’s likelihood method, briefly discussed above and
championed in recent years by AWF Edwards, Richard Royall.
2. Fisher’s test of significance method.
3. Neyman’s error probabilities (Types 1 and 2) method
championed recently by Deborah Mayo.
4. Neyman’s confidence interval method.
One wonders why the founders of EBM (Sackett et al) chose probability as
evidence and not some other measure for evidence in evidence based
diagnosis. I believe, we have a clue to their thinking in this regard in
the title of Sackett’s well known book Clinical Epidemiology. It is
possible they were so impressed by the accuracy of probability in
representing evidence as a stable frequency in a population that they
chose to represent evidence by a probability in the entirely different
domain of a given individual patient in which diagnosis is performed. It
is amazing how rapidly and widely this was accepted without any proof of
its validity. For example, I am still waiting for the first CPC or
clinical problem solving exercise in which a disease is diagnosed
correctly from a probability obtained by estimating a prior probability
and calculating a posterior probability. Furthermore, a probabilistic
approach, in my view promotes diagnostic error in some patients. Thus
what is cognitive bias of representativeness in which only diseases with
prototypical features (high prior probabilities) are suspected and
diseases with atypical variants (low prior probabilities) ignored but a
form of probabilistic reasoning in diagnosis.
What needed to be done and was not done by the founders, in my view, was
to develop evidence based diagnosis in a constructive manner. This means
starting with a close study and analysis of diagnosis in actual practice.
A close examination of even a few CPCs would have revealed to them that
evidence is not assessed by a probability in actual practice. They would
have noted that probability behaves very differently in a given patient
compared to in a population. In the latter, a higher probability always
corresponds to a greater frequency compared to a lower probability which
always corresponds to a lower frequency. In a given, individual patient,
on the other hand, a higher as well as a lower probability may correspond
to presence of a disease. If this initial analysis were done, it is
possible, evidence in diagnosis may not have been represented by a
probability.
In any case, there is such strong dogmatic belief in correctness of
probability is evidence that even a suggestion to the contrary evokes the
same incredulity and astonishment that were displayed by the
Aristotelians when they learnt that uniform motion does not require a
mover which is described by Galileo in his great book Discorsi or
Discourses on two new sciences.
It is well to remember that many longstanding incorrect beliefs such as
planetary orbits being circular, every motion requiring a mover, time
being absolute were overthrown and corrections made only by careful study
and analysis of actual processes related to them. Similarly, I believe,
progress in improving diagnosis will be made only by studying and
analyzing the process of diagnosis in actual practice. My own view, based
on careful study of CPCs is that evidence in diagnosis in actual practice
is represented by likelihood and not by probability. This needs to be
confirmed or disproved by further study and analysis of actual diagnosis.
In the end, regarding this discussion, I can do no better than leave the
last word for Einstein in the following long quote:
Concepts that have proven useful in ordering things easily achieve such
an authority over us that we forget their earthly origins and accept them
as unalterable givens. Thus they come to be stamped as ‘necessities of
thought’, ‘a priori givens’ etc. The path of scientific advance is often
made impassable for a long time through such errors. For that reason, it
is by no means an idle game if we become practiced in analyzing the long
commonplace concepts and exhibiting those circumstances upon which their
justification and usefulness depend, how they have grown up individually
out of the givens of experience. By these means, their all too great
authority will be broken. They will be removed if they cannot be
legitimized, corrected if their correlation with given things be far too
superfluous, replaced by others if a new system can be established that
we prefer for whatever reason.
I make no apologies for this very long email as the discussion of
diagnosis, which is a complicated topic with many strands, requires many
well thought out arguments that can only be expressed by detailed
statements.
I would like to thank Mark Graber for providing a forum in ListServ for
this important discussion.
Bimal
Bimal P Jain MD
Pulmonary-CriticalCare
NorthShore Medical Center
Lynn MA 01904
From: John Brush [mailto:jebrush at ME.COM]
Sent: Friday, October 09, 2015 1:16 PM
To: IMPROVEDX at LIST.IMPROVEDIAGNOSIS.ORG
Subject: Re: [IMPROVEDX] [IMPROVEDX] IOM report is released - Diagnosis in
actual practice
Mark Ebell’s insightful email points to the absurdity of not
thinking probabilistically. The probabilities that skilled clinicians use
in an intuitive fashion in practice are derived from scientific evidence
and experience (experiential knowledge). I am a big believer in
evidence-based medicine. Without the scientific underpinnings for our
clinical activities, and explicit acknowledgement of where the science is
lacking, we would be adrift.
In David Sackett’s original BMJ article on evidence-based
medicine, he promoted the use of the best available scientific evidence "
especially from patient centred clinical research into the accuracy and
precision of diagnostic tests.” He included with his book a nomogram that a
clinician could use to incorporate likelihood ratios for diagnostic tests
with prior probabilities to yield a posterior probability for a single
patient undergoing a single test. To him and other EBM founders,
probability was the very foundation of statistical inference and
evidence-based medicine.
In my book, I stated "We can make highly accurate actuarial
predictions for populations, but we have trouble even comprehending what
probability means for a single patient or event.” I think this email trail
points to the difficulty of thinking about probability in medicine. I am
concerned that Dr. Jain’s paper only adds to the confusion. I would submit
that in every one of the CPCs that Dr. Jain refers to, the discussant makes
a statement like, “I think the most likely diagnosis is….” The statement
"most likely" is another way of saying "most probable." Dr. Jain’s
statement that probability is only a theoretical consideration and is not
used in the practice of medicine is, I think, absurd. Not acknowledging the
uncertainty in medicine through some statement of probability (which is
simply a way to quantify uncertainty) leads to an illusion of certainty,
arrogance on the part of the practitioner, and unrealistic patient and
family expectations.
In my book, it took me 5 chapters to fully develop the idea of
probability and how it should be used to think about the diagnosis and
treatment individual patients. This is tough to think about and comprehend
and I think it can be misrepresented in an email listserv.
For excellent reading on probability, I would suggest Ian
Hacking’s “An Introduction to Probability and Inductive Logic” or Gerd
Gigerenzer’s “Risk Savvy: How to Make Good Decisions” or “Calculated Risks:
How to Know When Numbers Deceive You.”
Ian Hacking and Gerd Gigerenzer participated in a year long
sabbatical with other philosophers, scientists, and cognitive psychologists
where they explored the meaning of probability. Hacking has also written
“The Taming of Chance” and “The Emergence of Probability.” Girgerenzer has
also written “The Empire of Chance: How Probability Changed Science and
Everyday Life” for more in-depth reading.
In clinical medicine, we use conditional probability every day,
but because the exact numbers are only estimates, we actually use a
heuristic called anchoring and adjusting. We use reason every day, but
because the ultimate truth may be unknown, we use a type of reasoning
called abductive reasoning. (My spell-checker incorrectly changed the
spelling to adductive reasoning in my prior email.) Abductive reasoning,
described by Charles Saunders Peirce, is "reasoning toward the most
plausible hypothesis." When we start the diagnostic process, we are dealing
with multiple hypotheses (plausible conjectures). We work through the
process toward the most plausible hypothesis and, again, the term “most
plausible" implies some concept of relative probability. With abductive
reasoning, we blend both inductive reasoning and causal reasoning to make
an argument (meaning a logical statement) that combines both probability
and pathophysiologic rationale.
As W. Edwards Deming said, "if you don’t understand the process
of what you are doing, you don’t know what your are doing.” It is important
for clinicians to have a better understanding of the process of making a
diagnosis. By developing and using good habits based on a deep
understanding of process, the clinician will have the best chance of making
the correct diagnosis, as reliably as humanly possible.
My apologies about the long email, but I am very serious and
passionate about improving the quality of medical decisions.
John
John E. Brush, Jr., M.D., FACC
Professor of Medicine
Eastern Virginia Medical School
Sentara Cardiology Specialists
844 Kempsville Road, Suite 204
Norfolk, VA 23502
757-261-0700
Cell: 757-477-1990
jebrush at me.com
On Oct 8, 2015, at 2:54 PM, Mark H Ebell <ebell at UGA.EDU> wrote:
So, I should order a chest CT for every patient with cough, to rule out
lung cancer.
And I should order a stress thallium for every single 25 year old with
chest pain that appears to be musculoskeletal, so I don’t miss the rare MI.
And of course I should get a CT or MRI for every patient with a headache,
to not miss the rare (and generally untreatable) CNS cancer.
Do you realize the cost and harm of this approach? The complications of
invasive tests and biopsies and follow-up that go nowhere? The false
alarms? Radiation?
But at least you won’t successfully sue me. I guess that’s all that
matters.
Mark
—
Mark H. Ebell MD, MS
Professor of Epidemiology
University of Georgia
Editor, Essential Evidence
Deputy Editor, American Family Physician
ebell at uga.edu
From: Phillip Benton
Reply-To: Society to Improve Diagnosis in Medicine, "pgbentonmd at AOL.COM"
Date: Thursday, October 8, 2015 at 1:17 PM
To: "IMPROVEDX at LIST.IMPROVEDIAGNOSIS.ORG"
Subject: Re: [IMPROVEDX] Fwd: [IMPROVEDX] IOM report is released -
Diagnosis in actual practice
I am an experienced physician-attorney (Medicine 54 Years, Law 44 years)
teaching Medical Malpractice at a top tier law school for almost 20 years.
Relevant to this discussion is the fact that missed or delayed diagnosis
tops the list of causes for awards to injured plaintiffs at mediation or in
jury trials. Accounts of serious medical error by the Institute of Medicine
(1999) and the Journal of Public Safety (Sept, 2013) document missed or
delayed diagnosis as a leading cause of preventable harm. The IOM estimate
of up to 98,000 preventable deaths per year means one every 5 minutes 22
seconds; Dr. John James' evidence-based 2013 update of up to 440,000
preventable deaths per year equals one every 72 seconds.
One standard approach by the plaintiff''s attorney is to have the
defendant physician or defense expert witnesses agree that it is important
to make a differential diagnosis, and then to first rule out the most
serious and dangerous diagnoses on that list. In other words it is not the
odds but it is the stakes that matter most. Greater experience of a
physician may move this process from System II (rational) toward System I
(intuitive) thinking, but the point is that juries (i.e., patients)
routinely agree that you should always deal with the most important things
first.
A common expression heard when an uncommon disease is misdiagnosed is that
"When you hear hoofbeats you think of horses, not zebras." The savvy
attorney will then then ask "And how do you tell the difference? (pause)
You look!" Adequate testing to first rule out life-threatening conditions,
treatable if caught early, may often allow a successful defense.
Phillip G Benton, MD, JD
Atlanta Georgia
-----Original Message-----
From: Bob Latino <blatino at RELIABILITY.COM>
To: IMPROVEDX <IMPROVEDX at LIST.IMPROVEDIAGNOSIS.ORG>
Sent: Thu, Oct 8, 2015 10:55 am
Subject: Re: [IMPROVEDX] Fwd: [IMPROVEDX] IOM report is released -
Diagnosis in actual practice
I am not a physician nor a clinician, so I come at this issue basically
from the perspective of a patient.
When the physician becomes the patient, what is the expectation of them
towards their care provider in terms of their diagnosis?
Physicians themselves would obviously be more critical of their peer's
diagnosis when their lives are involved, because they are 'insiders' and
know the probing questions to ask about how the diagnosis was derived.
What are those questions? What should the non-clinical patient be asking
of their doctors when they provide a diagnosis?
I am in the investigation business and work in aviation, nuclear power,
military and other potentially life-threatening businesses. Many in these
businesses have to make spur of the moment decisions (diagnosing the
problem) and then quickly act on it.
I will take pilots for instance. I know healthcare has taken an interest
in Crew Resource Management (CRM) from the training that pilots receive
about effective cockpit communications and teamwork. They too have to
quickly make a diagnosis and act on it accordingly.
The difference between a pilot and a doctor in these situations is that the
pilot and crew's lives are at stake (along with the passengers) as well,
based on the accuracy of their diagnosis, decisions and actions.
Given this informative debate about probabilities and looking at
situations/patients singularly versus as a population, how does a pilot
make their quick assessment versus a doctor and their diagnosis? Does the
fact the pilot's life is at stake differ in their decision as opposed to a
doctor, whose life is not likely at stake based on their decision? Does it
matter? Should it?
Robert J. Latino, CEO
Reliability Center, Inc.
1.800.457.0645
blatino at reliability.com
www.reliability.com
From: Jason Maude [mailto:Jason.Maude at ISABELHEALTHCARE.COM]
Sent: Thursday, October 08, 2015 10:09 AM
To: IMPROVEDX at LIST.IMPROVEDIAGNOSIS.ORG
Subject: Re: [IMPROVEDX] Fwd: [IMPROVEDX] IOM report is released -
Diagnosis in actual practice
Although this is a very stimulating debate, I am struggling to understand
how relevant it actually is to the diagnosis of individual patients as a
key additional variable will always be the personal consequences of a wrong
decision. The key difference with a probabilistic approach in life
assurance or similar versus diagnosis of a particular patient has to be the
consequences of getting it wrong. This means that nobody is likely to
follow a purely probabilistic approach if they know the patient might die
if they didn’t check for something even if it was a lower probability. The
odds of winning the lottery are ludicrously bad but because the prize is so
big (upside consequences) people still try their luck. Personal
consequences will always seriously affect rational calculations of
probability.
Jason Maude
Founder and CEO Isabel Healthcare
From: "Jain, Bimal P.,M.D." <BJAIN at PARTNERS.ORG>
Reply-To: Society to Improve Diagnosis in Medicine <
IMPROVEDX at LIST.IMPROVEDIAGNOSIS.ORG>, "Jain, Bimal P.,M.D." <
BJAIN at PARTNERS.ORG>
Date: Thursday, 8 October 2015 12:13
To: "IMPROVEDX at LIST.IMPROVEDIAGNOSIS.ORG" <
IMPROVEDX at LIST.IMPROVEDIAGNOSIS.ORG>
Subject: Re: [IMPROVEDX] Fwd: [IMPROVEDX] IOM report is released -
Diagnosis in actual practice
We can only comment on and critically evaluate material that is published.
I find it simply amazing that a probabilistic approach in which probability
is evidence has not been employed in even a single amidst hundreds of
published CPCs and clinical problem solving exercises. Dr. Brush dismisses
CPCs as artificial, pedagogical exercises employing System 2 thinking over
days or weeks. This is all the more reason to employ a probabilistic
approach as the discussants then have plenty of time to estimate prior
probabilities and calculate posterior probabilities. This is not done
simply because this approach has not been found useful for diagnosis. Some
time back, I carefully examined 50 consecutive CPCs in NEJM from July 2013
to OCTOBER 2014. I found the word probability mentioned only once in these
50 CPCs. If Dr. Brush thinks this approach is suitable only for System 1
thinking in diagnosis ,Croskerry has pointed out the danger of such
thinking in causing diagnostic errors. At present, the emperor does not
appear to have any clothes with regard to probabilistic approach to
diagnosis in these exercises. What is needed ,I think are head to head
observational or experimental studies comparing usual to probabilistic
approach in real patients.
The adage ‘Common things are common’ is useful only in indicating chance of
a disease in a given patient. Certainly, we should look for a common
disease first as it has the greatest chance of being found. The problem
arises when a frequency or probability is taken as evidence for a disease.
There is little doubt in my mind, diagnostic errors due to failure to
suspect a disease in patients with atypical presentation in studies of
Hardeep Singh and John Ely arose from interpreting low prior probability as
absence of evidence for the disease.
In discussion about STEMI, Dr. Brush rightly deals with all patients with
STEMI regardless of prior probability in the same manner by taking them
all for cardiac cath. His accuracy rate of acute MI of 85 percent in these
patients is close to the rate of 90 percent in my paper. If he were to
analyze his data he would find the majority of patients with acute MI to
have intermediate or high prior probability.
I refer the Central Limit Theorem with regard to distribution of prior
probability which is a continuous variable.
The main problem with a probabilistic approach is that it takes probability
as evidence in a given individual patient while it is true only in groups
of patients. There is no proof that it improves diagnosis in actual
practice. Its use appears to have become a dogma which is hindering efforts
to reduce diagnostic errors. It is only by looking at diagnosis in actual
practice such as in studies of H. Singh and J. Ely and analyzing results
without putting on probabilistic glasses that we shall make progress.
I mention three examples from history of science of dogmatic beliefs
hindering progress which was made only when phenomena as they occur were
analyzed.
1. Since the time of Plato, the belief in planetary orbits being
circular due to perfection of a circle as a geometrical figure. All
contrary observations were explained away by drawing circles(epicycles)
within circles. It was only two thousand years later that Kepler determined
the orbit of Mars to be an ellipse when he actually observed and analyzed
its movement.
2. Since the time of Aristotle, every movement was believed to require
a mover. Contrary observations such as flight of an arrow were explained
away in an absurd manner. Again, about two thousand years later, the true
law of motion, that it is change in motion and not motion itself that
requires a force was discovered when Galileo observed and analyzed actual
motion of rolling balls.
3. And nearer to our age, there was a widespread belief in Absolute
Time since Newton declared it to exist in the 17th century. It was only in
early 20 the century this belief was overthrown by Einstein by his
insightful analysis of actual time in terms of clocks and trains.
Bimal
Bimal P Jain MD
Pulmonary-Critical Care
North Shore Medical Center
Lynn MA 01904
From: John Brush [mailto:jebrush at ME.COM]
Sent: Saturday, October 03, 2015 8:58 AM
To: IMPROVEDX at LIST.IMPROVEDIAGNOSIS.ORG
Subject: Re: [IMPROVEDX] Fwd: [IMPROVEDX] IOM report is released -
Diagnosis in actual practice
I’m afraid that I can’t agree with Dr. Jain’s argument. I think his
argument is circular, difficulty to follow, and selectively self-serving.
We have an adage in medicine: “Common things are common.”
Otherwise, every diagnostic exercise would become a wild goose chase,
leading us to look into every remote possibility every time. Having said
that, I can also say that if we collect cases over time, uncommon things
become common. Someone somewhere will eventually win the lottery. Uncommon
diagnoses do occur eventually. But the exceptions should not define the
rules.
The STEMI case that Dr. Jain presents proves my point. I am in
interventional cardiologist who frequently takes patients with suspected
STEMI to the cath lab for intervention. I have been getting direct feedback
on these cases for about 25 years. I can tell you that there is a false
positive rate of about 15% among STEMI alerts that are taken to the cath
lab (numerous reports in the literature confirm that estimate). We allow
that false positive rate because we make a subjective calculation of
expected value. Even if a patient has a relatively low initial prior
probability of STEMI, like Dr. Jain’s example, we don’t want to miss a
serious diagnosis like a STEMI. The EKG findings change the probability
estimate and make a STEMI quite plausible in such a patient. In a patient
like Dr. Jain’s example, we know that there is about a 50-50 chance of
finding an occluded artery, which is certainly high enough to activate the
cath lab. And sure enough, over time, 50% is about the frequency that we
find in such patients.
Dr. Jain references central limit theorem. That theorem applies
to probability for a continuous variable, and states that for any
distribution, the sample means of repeated samples will become a normal
distribution. I’m not sure I follow his argument that it applies to a
probability distribution of categorial variables. A diagnostic category is
a countable variable. Kolmogorov’s principles, however, do apply. The
probabilities of all of the possibilities do add up to one, if they are all
independent. General knowledge of these probability principles can help us
organize our thinking.
When we see a patient with chest pain in the ED, we start to
narrow the sample space by asking questions and making observations. For
example, we can eliminate the possibility of a stab wound very quickly by
noticing that there is no knife in the chest. Through early hypothesis
generation, we narrow the range of possibilities to the point were we can
start the process of iterative hypothesis testing. We have at our disposal
many possible tests that we can perform. We can send a troponin, do a CT
scan for dissection, do a stress echo, go directly to the cath lab, etc. We
can’t do all of these tests at the same time, and we probably don’t want to
do every test on every patient. So how do we decide what test to do first?
We do a little mental calculation of the subjective probabilities, which
gives us an idea of the expected value of each test. We don’t want to miss
a diagnosis with serious consequences, like MI or dissection, so an EKG and
CXR are done on virtually everyone, regardless of the prior probability.
But we narrow the sample space as we hone in on the correct diagnosis. We
don’t want to narrow the search prematurely, and we use a differential
diagnosis to help us guard against jumping to conclusions. All of this is
guided by some notion of relative probabilities.
Moderator: David Meyers, Board Member, Society to Improve Diagnosis in Medicine
HTML Version:
URL: <../attachments/20151015/e9ab1595/attachment.html>
ATTACHMENT:
Name: graycol.gif
Type: image/gif
Size: 105 bytes
Desc: not available
URL: <../attachments/20151015/e9ab1595/attachment.gif>
More information about the Test
mailing list