Missed and Erroneous Diagnoses Common in Primary Care Visits
John Brush
jebrush at ME.COM
Wed Jan 1 14:06:47 UTC 2014
Thanks, Bimal, for a very interesting discussion.
You seem to be tenaciously holding on to the frequentist notion of probability. Many statisticians agree with you and argue that this is the only way to think about probability. But I think that we, as clinicians, need to use both the frequentist notion and the personal notion of probability to properly apply our experience to the care of individual patients.
We can use the frequentist notion to look back at our experience, and learn from the variability that we observe. But we also have to go to work each day and take care of unique patents one at a time. We have to place these patients into diagnostic categories, and our challenge is to do that correctly and in a timely fashion.
I agree that a Bayesian approach is not always necessary. A broken arm is a broken arm. And if you always strictly used a pure Bayesian approach, you would never diagnose a rare disease. I also agree that some clinical findings are such strong evidence that they dominate any probabilistic assessment. Some findings are pathognomonic. But very few tests have a positive likelihood ratio of greater than 10. We need a Bayesian approach to guide us when the diagnosis is obscure, the signs and symptoms are ambiguous, and the available tests are imperfect.
Try this: The next time you see a patient who you think has a pulmonary embolus, ask yourself, “am I absolutely certain of the diagnosis?” If your answer is no, then ask yourself “on a scale of 1 to 10, how certain am I of the diagnosis?” I think you would agree that it is reasonable to pin yourself down with that question, to determine your level of certainty. With the personal or Bayesian notion of probability, you are doing essentially the same thing. Rather than a scale of 1 to 10, you are using a scale of 0 to 1. The Bayesian approach that I suggested in my last email gives you a way to arrive at that number in a logical and systematic way that takes advantage of all of the information that you have.
One other thought: Imagine that you are a cardiologist like me and many patients are sent to you to determine if they have CAD. You have noticed over the years, given your referral patterns, that you ultimately diagnose CAD in about half of the patients who are referred to you. Lets say that you see 10 patients every day. Over the years, you observe that on many days, 5 of the 10 patients have CAD. Other days, it’s 4 or 6. On rare days it's 1 or 9, and almost never will it be 0 or 10. What you can do is create a binomial distribution that shows how patients add up over time. So single patients with a binary outcome can be turned into a frequency distribution. See the graph below (if it gets through the listserv).
ATTACHMENT:
Name: pastedGraphic.pdf
Type: application/pdf
Size: 22774 bytes
Desc: not available
URL: <../attachments/20140101/6f4c0812/attachment.pdf>
People have been arguing about the meaning of probability for over 300 years. Most people have only a cursory understanding of the meaning of probability. And most of our cognitive errors (base rate neglect, availability, representativeness, regression to the mean, etc) are really just miscalculations of probability estimates.
The reality is that we practice a messy, uncertain world where we have to use inductive reasoning and we are forced to use probability. I think we should spend more time in medical education helping people understand these concepts. I don’t claim to fully understand probability. But I’ve wrestled with it a lot and I think that discussions like this one are really useful and important.
Happy New Year to all!
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 Dec 31, 2013, at 10:49 AM, Bimal Jain <bjain at PARTNERS.ORG> wrote:
Hi John, I congratulate you on your excellent book which I enjoyed immensely. I found it to be very well written and highly informative. It should be read by every practising physician.
I would like to make a few cautionary remarks about role of probability in clinical diagnosis.
1. A probability, whether prior or posterior symbolises a distribution or frequency in a series of similar patients while clinical diagnosis seeks to determine a disease correctly in a given, individual patient.
2. Therefore, while probabilistic considerations help move the diagnostic process forward, a probability cannot be considered evidence from which a disease is diagnosed in a given patient.
3. Thus, it is entirely reasonable to test for a disease with high prior probability first, as he is drawn from a series of patients in most of whom the disease is present. Therefore, the chance of this disease being found in our patient is high.
4.What would be inappropriate, I suggest, would be to equate high prior probability with high prior evidence and use it as such for diagnosis.
5. This occurs, I suggest in patients with low prior prob. when a disease is ruled out (declared absent) without testing. This has been reported in several cases of acute MI being missed in healthy young women with atypical chest pain.
6. Evidence in a given patient, I believe, is best (and perhaps only) measured by a likelihood ratio. It is customary to diagnose a disease definitively only if a test result with LR of 10 or higher is observed, regardless of prior prob. ( diagnosis of acute MI from acute Q wave and ST elevation EKG changes, LR 13).
7. It is well known from experience, any given disease occurs in different patients with clinical presentations and therefore prior prob. which vary over a wide range. Our goal as clinicians is to diagnose a disease correctly in a given, individual patient regardless of prior prob.
8. In practice, therefore, the correct approach, when confronted by a patient with symptoms, I suggest, is to look upon a presentation as a problem to be resolved and not as evidence for a certain disease.
9. The presentation functions as a clue which makes us suspect several diseases. The order in which we test them may be determinedby prior prob. as discussed earlier.
10.Some test results, at least, such as acute EKG changes (LR 13), positive chest CT angiogram (LR 21), positive venous ultrasound (LR 19) lead to definitive diagnosis of their repective diseases, acute MI, pulmonary embolism, DVT, regardless of prior prob.
11. This suggest, diagnosis of some diseases at least is not performed in a Bayesian manner in actual practice.
12. I believe, it is important to record diagnostic errors in registries as I suggested earlier, so we can study these cases and learn why errors were made. Was a disease not suspected because of its low prior prob. or was an inappropriate test emplyed to diagnose or rule out a disease?
13. Till such registries are formed, we can present and discuss such cases in our forum hwere.
14. I think we need more observational and experimental studies of diagnosis in actual practice to ascertain the method which minimises error.
15. In my view, a strictly Bayesian appproach is not such a method.
Bimal P Jain MD
Pulm.-Crit.Care
Northshore Med. Center (Union)
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>
</p>
More information about the Test
mailing list