Missed and Erroneous Diagnoses Common in Primary Care Visits

Dean F. Sittig Dean.F.Sittig at UTH.TMC.EDU
Mon Dec 30 16:54:24 UTC 2013

These concepts have been covered in the medical informatics literature for many years. See this seminal paper from 1959 that is often given credit for starting the field of medical informatics.
and here is another practical implementation of these concepts from one of my PhD advisors: Homer Warner

From: John Brush [mailto:jebrush at ME.COM]
Sent: Sunday, December 29, 2013 12:48 PM
Subject: Re: [IMPROVEDX] Missed and Erroneous Diagnoses Common in Primary Care Visits

Hi Bimal,
            Wow, you have given me a lot to respond to.
            Regarding your point #1, I think you are disregarding the prior probability that would be determined by the situation and context. There may be a cold wave moving in, which is the new evidence, but you have disregarded the baseline conditions. Are you in Florida or Minnesota? Is it winter or summer? These conditions will give you some general idea of the baseline probability of snow. You should have some general idea about whether you are initially skeptical, a believer, or neutral about the proposition that it will snow tomorrow.
            Regarding #3, prior probability is your starting point, to which you apply new evidence to derive a posterior probability. That posterior probability can then become the prior probability for a new bit of additional evidence. You can daisy-chain new evidence updates to constantly update your probability assessment. You can easily do this with likelihood ratios, but you need to convert probability to odds before you start multiplying by the LR's, and then convert post-test odds back to probability in the end.
            Regarding #6, I agree that it is important to factor in the strength and accuracy of new information. Likelihood ratios help you do that. They incorporate the sensitivity and specificity of new information to tell you how much weight to put on a positive test result or a negative test result.

            So here's how I would put this into practice. You start developing some idea of the possibilities and their associated probabilities at the very beginning of a patient encounter when you elicit a chief complaint. This starts you thinking about the possibilities. You start general - is it acute or chronic, localized or systemic, serious or mild, injury or illness? You ask a series of probing questions, and you start to develop in your mind a short list of the possibilities. You continue to probe with more targeted questions to get a clearer idea of the possibilities. Once you finish the history and physical, you should have a list in your mind of 3-5 possible diagnoses. But you generally wouldn't rank all 5 of them equally. You would rank them in your mind according to their relative likelihood. If the diagnostic possibilities are mutually exclusive and collectively exhaustive (it has to be one of them), the total cumulative probability of all of your possibilities has to add up to one. Thus, you can start to give a provisional probability to each possibility. This is the process of early hypothesis generation. At this point, you have developed a list of hypotheses, sorted according to relative probability, but you haven't concluded anything yet. You now have to go through the process of iterative hypothesis testing to decide whether to accept or reject each hypothesis, until you reach a final conclusion.
            Let's say that the patient has chest pain. There are typical and atypical features. After questioning and examining the patient, you think it may be anxiety or angina, but could be pericarditis or GERD, or even a dissection. You may think that anxiety and angina are likely, pericarditis and GERD are less likely, and aortic dissection is very unlikely. Knowing that the probabilities of all the possibilities add up to one, you can start to estimate prior probabilities, based on the relative probability of you 5 possibilities.  We'll assign initial probability estimates, giving anxiety and angina 0.3 each, dissection 0.04, which leaves 0.18 for both pericarditis and GERD. You could tinker with the numbers a bit, but this is a starting point. You look at the CXR and see no mediastinal widening and decide that dissection is so low that you will discard it. You decide to do a stress echo to evaluate for the possibility of angina. Your prior probability is 0.3 giving a prior odds of 0.43. A stress echo has a LR(+) of 6 and a LR(-) of 0.12. Thus, a positive stress echo changes the odds to 2.6, which is a post-test probability of 0.72, whereas a negative stress echo changes the odds to 0.05 and a post-test probability slightly less than 0.05. Let's say that the stress echo is negative and there's no rub on exam. Thus, the probability of angina and pericarditis goes way down, and the probability of anxiety or GERD has to rise dramatically.
            I'm not suggesting that we should make these strict numerical calculations on every patient. But this is the system that we all use intuitively, and the expert, through experience, has a sense of the probabilities, and how to weight the new evidence. This is the essence of expert intuition for medical diagnosis.
            This is obviously a Bayesian approach. The usual criticism of the Bayesian approach is that the estimation of the prior probability is too subjective. But you have to get your thinking started somewhere. This system of thinking helps us avoid some common pitfalls, like base-rate neglect and early closure.
            All of this is in my book! I self-published my book as a iBook so I could make it free. If you have an iPad, you can go to the iBookstore and download it. This discussion in in Chapters 3 and 4. I hope this is helpful.

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
Cell: 757-477-1990
jebrush at me.com<mailto:jebrush at me.com>

On Dec 26, 2013, at 4:55 PM, Bimal Jain <bjain at PARTNERS.ORG<mailto:bjain at PARTNERS.ORG>> wrote:

Hi John, thank you for your many insightful comments. Let me respond in the numbered points below. 1. Let us take your  example of 75 percent chance of snow tomorrow. If we absolutely wished to know if it would snow or not tomorrow due to some very important event, such as launch of a space capsule, we shall seekstrong evidence for that particular day. Suppose this evidence is a cold airfront meeting a moistureladen warm airm airfront, which has a high likelihood ratio for snow, we will be nearly certain, I suggest, it will snow tomorrow, regardless of prior chance (probability of snow tomorrow. 2. We are in a similar situation in clinical diagnosis, where we aim to diagnose a disease correctly in a given , particular patient. 3. The important point is, a prior probability is not a measure of prior evidence in a given patient. It only sets the order in which we test various suspected diseases. However, a high prior probability could be trumped by other factors, such as potentially serious consequences of a disease or ease of testing it. 4. I think the danger of considering prior probability as prior evidence is that a very low prior probability may be talen as strong evidence against a disease which may be ruled out without testing. 5. In any case, we need to investigate role of probability by looking at diagnosis in actual practice and noting errors. 6. Thus we find cardiologist reading EKGs to diagnose Acute MI from acute EKG changes and radiologists to diagnose acute pulmonary embolism from positive chest CT angiogram and deep vein thrombophlebitis from positive venous ultrasound study without knowledge of prior probabilities of these diseases. We need to look at accuracy rates of these diagnoses to see if any errors are being made. 7.. It would be helpful if we established institutional, regional and national registries for recording diagnostic errors of various diseases. They could then be studied to classify various errors and perhaps identify their causes. 8. My hunch is most diagnostic errors occur due to failure to think of and test for diseases with low prior probabilities. This could be eliminated to a great extent by teaching that a prior probability is not evidence for or against a disease in a given patient.

To unsubscribe from the IMPROVEDX:


For additional information and subscription commands, visit:

Moderator: Lorri Zipperer Lorri at ZPM1.com<mailto: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<http://list.improvediagnosis.org/scripts/wa-IMPDIAG.exe?SUBED1=IMPROVEDX&A=1%3c/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<mailto:Lorri at ZPM1.com>, Communication co-chair, Society for Improving Diagnosis in Medicine

To learn more about SIDM visit:

HTML Version:
URL: <../attachments/20131230/62a6b1a2/attachment.html>

More information about the Test mailing list