Missed and Erroneous Diagnoses Common in Primary Care Visits

David Gordon, M.D. davidc.gordon at DUKE.EDU
Tue Dec 24 00:12:10 UTC 2013

Hi Bimal,

I have to say I am somewhat confused by your comments.  As a fellow front line clinician, I completely resonate with Karen's comments that our world is surrounded by uncertainty, probability, and risk stratification. I spend a lot of time excluding disease - at least to an acceptable post-test probability-  as opposed to confirming diagnoses.

In regards to your comment: "I would have no objection to assigning a prior probability of 0.5 to every patient with a suspected disease regardless of presentation"," I have some follow-up questions:

1) Don't you think the data from the rational clinical exam series enables us to estimate the likelihood of disease within an individual? It gives us a lot of useful information about the predictive value of a clinical findings for a given disease process. So if there are 2 patients with chest pain of the same age, sex and risk factors but one presents with substernal chest pressure with bilateral arm radiation and the other with sharp, pleuritic chest pain, aren't you assigning them different individual prior probabilities for MI?

2) How do you decide which tests to order in the first place without assigning prior probabilities? If I don't assign patients with chest pain some clinical prior probability for a given disease, how do I decide who to get a chest CT on or who to keep overnight for serial cardiac markers and provocative testing?  We need to estimate clinical pretest probabilities based off both risk factors and clinical presentation to establish thresholds for pursuing diagnostic tests.

3) Most tests are far from perfect, lack specificity, and are best interpreted in the context of a well developed pre-test probability. That was one of the most interesting findings of the PIOPED study and v/q results. When just applied independently, v/q did not preform as well. It was when the results - either positive or negative- aligned with prior clinical probability that the test performed well. When there is discordance, more testing is needed. I think the same applies to most other tests as well.

Because we have to make decisions about what test to order to begin with combined with the imperfection of our tests, I would say Bayesian reasoning is a reflection of the inescapable process clinicians engage in.  The mistake is not that we engage in probabilistic reasoning, it is when we forget that most diagnoses are encased in probability - not certainty.



David Gordon, MD
Assistant Professor
Undergraduate Education Director
Division of Emergency Medicine
Duke University

The information in this electronic mail is sensitive, protected information intended only for the addressee(s). Any other person, including anyone who believes he/she might have received it due to an addressing error, is requested to notify the sender immediately by return electronic mail, and to delete it without further reading or retention. The information is not to be forwarded to or shared unless in compliance with Duke Medicine policies on confidentiality and/or with the approval of the sender.
From: Karen Cosby [kcosby40 at GMAIL.COM]
Sent: Monday, December 23, 2013 5:51 PM
Subject: Re: [IMPROVEDX] Missed and Erroneous Diagnoses Common in Primary Care Visits

As I've listened to this discussion I'm struck that there is an assumption that we make specific diagnoses with  absolute standards and therefore, the diagnosis must always be accurate.  In primary care, and emergency medicine, we rarely have tools refined well enough to make many diagnoses with a high degree of certainty or specificity.  Mostly we make estimates of likelihood of disease based on epidemiology (we worry about influenza during flu season, we consider carbon monoxide poisoning at the start of cold weather), and assess patient risk profiles (a older male with poorly controlled hypertension and and long smoking history makes me worry a lot more about aneurysm; a young male with no past medical history not).  We look for indications for intervention, thresholds for action.  But we are often not so concerned about making an exact and highly refined specific diagnosis than we are about ruling out things we can do something about.  We and the public are highly invested in making diagnoses and being accurate, but in fact, most diagnoses are at best estimates of disease likelihood and estimates of risk.  There is always a broader differential, even if it's a possibility of an atypical presentation and/or uncommon condition.  Also, we tend to have this discussion without much consideration for the pragmatic aspects of medical care.  When I first trained in medicine, the standard of diagnosis for pneumonia was a culture from a quality sputum sample.  We've since learned that most sick patients can't wait for treatment (and specific diagnosis) to depend upon culture results.  We may not have the exact diagnosis, but are good enough to treat the majority of patients.  I think part of our struggle is to determine what end point we want to perfect: just diagnosis, or are we really just talking about optimizing outcomes.  Front line clinicians really have to function with assessments of probability.  When we review cases of diagnostic error in retrospect, we often feel secure and convinced of the one final diagnosis.  But that same case could have had many other explanations for the same presentation.  We fool ourselves into thinking that the case as presented has the final complete and perfect answer, when in fact, I've seen cases critiqued and judged harshly only to have the patient to return with another final diagnosis.  Maybe a diagnosis should just be considered a placeholder that serves us well until it is replaced with another, better, more refined label.  Just like theories in science, a theory is assumed until it is disproven, but is itself never actually proven.

On Mon, Dec 23, 2013 at 1:23 PM, Bimal Jain <bjain at partners.org<mailto:bjain at partners.org>> wrote:
Your point about prior probability is well taken. The point I am trying to make is  that with any presentation with any estimated prior probability, we have no evidence for presence or absence of a suspected disease in a given, individual patient. I would have no objection to assigning a prior probability of 0.5 to every patient with a suspected disease regardless of presentation. This would lead to a post test probability of disease of 90 percent in every patient in whom a test result with likelihood ratio of 10 is observed. This post test test probability will actually correspond to our experience as I discuss below.
It is well known from experience any given disease , acute myocardial infarction for example occurs in different patients with varying presentations and therefore varying prior probabilities. As a presentation is constituted by combination of a number of independent factors such as symptoms, age, sex, risk factors,etc. we can expect the prior probability to be distributed normally in patients with disease encountered by us. (In PIOPED study on diagnosis of pulmonary embolsism, 68 percent patients with pulmonary enbolsm had mid range prior probabilities JAMA 263: 1990, 2753-2759) The average prior probability in these patients will then be close to 0.5. Observation of a test result with likelihood ratio of 10 in these patients will lead to an average post test probability of 90 percent indicating the test result diagnoses diease correctly in 90 percent patients. This high accuracy was actually observed in a large series of patients in whom acute Q wave and ST elevatio EKG changes with likelihood ratio of 13 diagnosed acute MI correctly in 90 percent patients (Rude et al Am J Card 52: 1983, 936-942). The problem with the standard Bayesian approach in which a prior probability is estimated from presentation in a given patient is that it leads to a diagnosis sometimes which is wrong from a clinical standpoint. This is seen in the following case, discussed in a clinical problem solvig exercise(Pauker et al NEJM 326:1992, 688-691). A 40 year old healthy woman without any cardiac risk factor presents with highly atypical chest pain and is found to hav acute Q wave, ST elevation EKG changes.Her prior probability estimaterd to be 7 percent of MI is combined with LR of EKG changes of 13 to yield a post test probability of 50 percent. The Bayesian diagnosis of MI being indeterminate is in sharp contrast to near certain diagnosis of MI from EKG changes alone by discussing physician.

To unsubscribe from the IMPROVEDX:


Moderator: Lorri Zipperer Lorri at ZPM1.com, Communication co-chair, Society for Improving Diagnosis in Medicine

To unsubscribe from the IMPROVEDX list, click the following link:<br>
<a href="http://list.improvediagnosis.org/scripts/wa-IMPDIAG.exe?SUBED1=IMPROVEDX&A=1" target="_blank">http://list.improvediagnosis.org/scripts/wa-IMPDIAG.exe?SUBED1=IMPROVEDX&A=1</a>


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: Lorri Zipperer Lorri at ZPM1.com, Communication co-chair, Society for Improving Diagnosis in Medicine

To learn more about SIDM visit:

HTML Version:
URL: <../attachments/20131224/54e4bca7/attachment.html>

More information about the Test mailing list