Why is the Bayesian method still unemployed?

Bruno, Michael mbruno at PENNSTATEHEALTH.PSU.EDU
Thu May 31 23:15:27 UTC 2018

You've summarized it perfectly, Tom!

It's amazing how difficult it is to even get my radiology residents to fully understand this concept!  How much harder to communicate this idea to ER physicians, and how much harder again to teach this concept to our patients!

"the same test result returns a different disease probability based on pre-existing probabilities"


From: Tom Benzoni <benzonit at GMAIL.COM>
Sent: Thursday, May 31, 2018 5:54 PM
Subject: Re: [IMPROVEDX] Why is the Bayesian method not employed for diagnosis in practice

What is often neglected in understanding the application of Bayes' theorem is it establishes the PRE-test probability, not post-diagnosis.
Then the process self-learns, being modified by the output.

For example, you're screening for the presence/risk/chance of heart disease by doing a cholesterol. (Don't belabor if this works.)  Upper limit of normal = 200; 2 people have 220 totals.

A comes from a family living into their 90s, taking an occasional vitamin.  B has a grandpa with first MI before 45 and a sibling with an MI @ 43.

Which person is more likely to have the disease in question?
Bayes helps answer this; it does not make the diagnosis.

In other words, the same test result returns a different disease probability based on pre-existing factors.

This principle is obvious to practitioners but not so much to people who feel testing is absolute.

Bayes' theorem can get a whole lot more wonky, but that's an application; there are others.
I'll make other examples should that help.


On Thursday, May 31, 2018, Tibbits, Paul A. <Paul.Tibbits at va.gov<mailto:Paul.Tibbits at va.gov>> wrote:


What wiki entry? Is there a link? Tnx. PT

Sent with Good (www.good.com<http://www.good.com>)

From: Tom Benzoni
Sent: Wednesday, May 30, 2018 1:08:48 PM
Subject: [EXTERNAL] Re: [IMPROVEDX] Why is the Bayesian method not employed for diagnosis in practice

I think there may be sone fundamental misunderstandings of Bayes' theorem and its application.
The Wiki entry on the topic is quite good.

On Wednesday, May 30, 2018, Jain, Bimal P.,M.D. <BJAIN at partners.org<mailto:BJAIN at partners.org><mailto:BJAIN at partners.org<mailto:BJAIN at partners.org>>> wrote:

The reservation I have about the Bayesian method is about considering a prior probability which is a frequency in a population to be prior evidence for a disease in a given patient. This notion may make us not suspect a disease if its prior probability is low leading to a diagnostic error.
It is of interest that in the two examples given by Dr. Elias, the presence of a disease is assessed in terms of a likelihood ratio, and not a prior probability. For example, IBS being more likely than splenic artery aneurysm given the GI symptoms means that the likelihood ratio of IBS is high compared to aneurysm.

Similarly, the likelihood ratio of viral illness is high compared to strep throat given the nasal and respiratory symptoms.

What these likelihood ratios are, I do not know, but they may be high enough, that further testing may not be required.

Another example of a high likelihood ratio  for a disease is that for herpes zoster, given unilateral, blistering skin lesions in a dermatomal distribution which does not require further tests for a diagnosis.

In diagnosis, in which our goal is accurate determination of a disease in a given patient, evidence from which we diagnose a disease is represented by a likelihood ratio as it represents a change in probability ( odds ) of a disease in a given patient , locating evidence in this particular patient.

A probability, being a frequency in a population locates evidence in a population. Therefore, an inference from a probability  is made in a field such as life insurance business in which the goal is long run accuracy with tolerance for errors in some individual persons.
I have never come across a diagnosis being made purely from a probability in the absence of data with significantly high likelihood ratio in published case discussions or in practice.

There seems to be a general impression that evidence in an uncertain situation can only be represented by a probability and an inference made from it alone. This is not true, as there is flourishing school of inference from likelihood alone as seen in books by AWF Edwards, Richard Royall, Y. Pawitan and many more authors.

A positive CT study for appendicitis has a likelihood ratio of 19, about the same for positive CT angiogram for pulmonary embolism. In practice, this study should allow us to diagnose appendicitis definitively with a high degree of accuracy in any patient, regardless of prior probability.
It would be of interest to know the diagnostic accuracy of this test result across patients with varying prior probabilities  similar to the diagnostic accuracy of acute Q wave and ST elevation EKG changes for acute myocardial infarction being 85 percent in patients with varying prior probabilities.


From: Elias Peter [mailto:pheski69 at GMAIL.COM<mailto:pheski69 at GMAIL.COM><mailto:pheski69 at GMAIL.COM<mailto:pheski69 at GMAIL.COM>>]
Sent: Sunday, May 27, 2018 12:12 PM
Subject: Re: [IMPROVEDX] Why is the Bayesian method not employed for diagnosis in practice

I find the proposal that Bayesian analysis is not (or should not be) part of the initial diagnostic approach puzzling. The following examples illustrate why I feel that way:

  *   The 30-year old who presents with 2 years of intermittent left upper quadrant pain and fluctuating bowel habits is more likely to have irritable bowel syndrome than splenic artery aneurysm. One would not start evaluation with imaging the LUQ.

  *   The 8-year old with 4 days of runny nose, cough and sore throat is so much more likely to have a viral illness than a strep throat that a throat culture is poor medical practice.

Peter Elias, MD

On 2018.05.27, at 8:53 AM, Bruno, Michael <mbruno at PENNSTATEHEALTH.PSU.EDU<mailto:mbruno at PENNSTATEHEALTH.PSU.EDU><mailto:mbruno at PENNSTATEHEALTH.PSU.EDU<mailto:mbruno at PENNSTATEHEALTH.PSU.EDU>>> wrote:

Thanks Dr. Bimal for starting this interesting thread, and to Stefanie Lee for sharing that excellent 2006 paper from Blackmore, et al., as well as to Dr. Oldham for his insightful comments.

I think the Blackmore paper is really touching on the topic of "signal detection theory," which is a very useful tool to understand about how useful information is actually extracted from complex data sets (like CT scans) that have extremely high levels of uncertainty built-in. We radiologists practice in a milieu of extraordinarily high undertainty, as I noted in my 2017 review in our Society's official journal, Diagnosis (DOI 10.1515/dx-2017-0006), so this is particularly relevant to us.

I've attached an excellent PDF discussion of signal-detection theory as applied to diagnostic radiology to this message for anyone who might be interested.  While radiologists and others may recognize the use of ROC curves, which are a feature of this theory, they may not be as familiar with the concepts of d', a quantitative measure of how well the alternatives can actually be discriminated from the test, and the concept of "criterion," whereby the interpreter decides how sensitive vs. specific they wish to be.  This was the thrust of the Blakemore article, and his meta-analysis suggested that radiologists by-and-large got it right.  (For David Meyers--the attached PDF is in the public domain, provided by Professor David Heeger of NYU).

Dr. Oldham also got it right, of course, saying that Baysean reasoning is not everything in diagnosis, but I believe that it takes us most of the way.  I very much appreciate Dr. Oldham's analogy that the radiologist is acting as the "expert witness" in a courtroom setting while the treating physician and patient / patient's family serves as the judge and jury, who are charged with ultimately deciding what "truth" is.

Have a terrific Memorial Day Weekend, everybody!


From: James Oldham <james.oldham at HEALTH.NSW.GOV.AU<mailto:james.oldham at HEALTH.NSW.GOV.AU><mailto:james.oldham at HEALTH.NSW.GOV.AU<mailto:james.oldham at HEALTH.NSW.GOV.AU>>>
Sent: Sunday, May 27, 2018 1:49 AM
Subject: Re: [IMPROVEDX] Why is the Bayesian method not employed for diagnosis in practice

Dear all

There is a difference between the treating doctor who makes a probabilistic diagnosis and evidence providers such as radiologists who comment on phenomena the observe and interpret it - taking Bayesian prior probability calculations explicitly into their offerings. I think the testing physician and their patients and family as a collective Judge relying on evidence provided by expert witnesses. So this final stage will be informed by Bayesian logic but not ruled by it and certainly not trying to make a prediction- just best information to hand

James Oldham
Child Psychiatrist

from  James Oldham - sent from my phone. (Please accept apologies odd sentence construction and creative additions courtesy of Apple autocorrect)

On 27 May 2018, at 13:25, Stefanie Lee <stefanieylee at GMAIL.COM<mailto:stefanieylee at GMAIL.COM><mailto:stefanieylee at GMAIL.COM<mailto:stefanieylee at GMAIL.COM>>> wrote:

An interesting paper from Blackmore and Terasawa, JACR 2006 on the complexity of CT interpretation and how clinical probability may be factored into the decision to call a test result positive or negative:

"In this paper, we present an example of how health utility assessment can be used to guide the optimum interpretation of an imaging test."

"Radiologists have the ability to alter the sensitivity and specificity of their interpretations. For objective positivity criteria, different thresholds can be chosen to consider test results positive. For example, with appendicitis, one important factor in CT interpretation is the size of the appendix. Using a lower size threshold (eg, 6 mm) to consider an appendix abnormal will result in higher sensitivity for appendicitis, at the expense of lower specificity. Using a higher size measurement to consider the appendix abnormal (eg, 8 mm) would result in lower sensitivity but higher specificity. For subjective criteria such as periappendiceal fat stranding, the process is the same but less explicit. Individual radiologists can alter how much increase in the density of the fat is necessary to be considered abnormal “stranding.”"

"In general, when a disease is rare and there are substantial costs to a false-positive diagnosis, interpreting a test with high specificity will maximize patient benefit. In contrast, when a disease is common and there are substantial costs for a false-negative diagnosis, interpretation at high sensitivity will maximize utility."

"Our prior meta-analysis indicated that radiologists interpret CT scans with approximately equal sensitivity and specificity. The current analysis indicates that this is an appropriate threshold at the intermediate probability of disease at which CT is commonly used today. If CT is to be used in populations at higher or lower probabilities of disease, then different imaging thresholds will be appropriate."

On 15 May 2018 at 13:18, Jain, Bimal P.,M.D. <BJAIN at partners.org<mailto:BJAIN at partners.org><mailto:BJAIN at partners.org<mailto:BJAIN at partners.org>>> wrote:

In this attached paper, I discuss that the prescribed Bayesian method is not employed for diagnosis in practice because probability of a diseases is considered evidence for it in a given patient in this method, which is incorrect as a probability is a frequency in a population. This leads to all sorts of errors in practice which I discuss.

The correct method of diagnosis, which is employed in practice, consists of hypothesis generation and verification in which evidence is assessed by a likelihood ratio which locates it in the given patient of interest.

Please review and comment on this paper.



Bimal P Jain MD
Northshore Medical Center
Lynn MA 01904.
The information in this e-mail is intended only for the person to whom it is
addressed. If you believe this e-mail was sent to you in error and the e-mail
contains patient information, please contact the Partners Compliance HelpLine at
http://www.partners.org/complianceline . If the e-mail was sent to you in error
but does not contain patient information, please contact the sender and properly
dispose of the e-mail.



To unsubscribe from IMPROVEDX: click the following link:

Visit the searchable archives or adjust your subscription at: http://list.improvediagnosis.org/scripts/wa-IMPDIAG.exe?INDEX
Moderator:David Meyers, Board Member, Society for Improving Diagnosis in Medicine

To learn more about SIDM visit:

Moderator: David Meyers, Board Member, Society to Improve Diagnosis in Medicine

HTML Version:
URL: <../attachments/20180531/3b1151bd/attachment.html>

More information about the Test mailing list