Friday, March 10, 2017

How unreliable are computer algorithms in the Diagnosis of STEMI?

A 61 year-old with chest pain arrived to the ED by ambulance with resolving chest pain.  Here is his ED ECG:
The computer interpretation was "normal"
What do you think?

This ECG is NOT normal.  The T-waves in V2-V4 are very large in proportion to the QRS.   They are suspicious for hyperacute T-waves. There is low QRS amplitude.  

Can you trust a computer interpretation of "normal?" 
---Most of the time, but obviously not all the time.

  1. Hughes KE et al. Safety of Computer Interpretation of Normal Triage Electrocardiograms. Acad Emerg Med 2017; 24(1): 120 – 24. PMID: 27519772
This recent study, discussed on Salim Rezaie's fine site REBEL EM, implies you can trust the computer interpretation of "normal." (

What to do?

Recording serial ECGs would be useful.  The chest pain is resolving, so if these are resolving hyperacute T-waves, then followup ECGs should show their size diminishing.

However, when I saw this patient, I knew that he had come by ambulance, so I knew there must be a prehospital ECG recorded somewhere and went to look for it.

I found it.

Here it is:
Need I say more?
Yes, I'll say more: see the computer interpretation:
"Normal variant ST elevation, consider early (repolarization)"
Why the medics did not see it, I'm not sure. 

The cath lab was activated, as it should be with transient STEMI.

See this case of transient STEMI:

Spontaneous Reperfusion and Re-occlusion - My Bad Thinking Contributes to a Death.

Angiogram: There was 80% obstruction of the LAD with a large thrombus and TIMI-2 flow; the thrombus was suctioned out.

All serial troponins were undetectable!

Had we not seen those subtle hyperacute T-waves, and then sought out the prehospital ECG, the patient would have ruled out, with one of three outcomes:

1. Subsequent thrombus propagation with re-occlusion, recurrent chest pain and diagnosis made (although perhaps late and perhaps only after an adverse outcome)

2. Undergone stress test with uncertain results.

3. Been discharged to home with potential disastrous outcome.

Learning Points

1.  Computer algorithms that make the diagnosis of "normal" are usually correct, but is usually good enough?

2. Computer algorithms are completely unreliable at diagnosing STEMI, with both poor sensitivity and poor specificity.  Here are two recent articles confirming this:

    a. The Comparison of Physician to Computer Interpreted Electrocardiograms on ST-elevation Myocardial Infarction Door-to-balloon Times.
    b. Electrocardiographic diagnosis of ST segment elevation myocardial infarction: An evaluation of three automated interpretation algorithms

3.  Always look for the prehospital ECG

4.  Hyperacute T-waves occur not only shortly after onset of chest pain (as the ST segment is about to rise, or "on the way up"), but also shortly after reperfusion (as the ST segment is resolving after reperfusion or "on the way down").

5.  With very brief occlusions, troponins may all be negative.  What would high sensitivity troponins have shown?  We don't know.


  1. I don't think Salon's post regarding this paper implied computers are good surrogates of doctors. As I understood, it advocated for considering the significant disruptions caused by EKGs handed to attendings, without a perceivable change in patient outcomes. I would also add the effects of such disruptions on other patients care as well as false reassurance of No STEMI signature by cursory look to other staff (junior practitioners, RNs). I think it is much more likely that I would pick up the hyperacute To waves on this EKG if I was not supposed to read it in 10 seconds.

    1. You are absolutely right. However, papers like this get misinterpreted, and the misinterpretation gets propagated from person to person. 2 years from now, the paper will be summarized as: "if the computer says normal, the doctor doesn't need to read it." We must guard against that.

  2. Having been a student of (and author on) computerized ECG interpretations for the past 30 years — I long ago learned that the first priority for providers to master is appreciation of what the computer is good at, and what it is not good at. Computerized ECG interpretations have never been shown to have high enough sensitivity for assessing subtle infarctions to be used as a “stand-alone” tool ( ). This superb case by Dr. Smith illustrates this concept to perfection — as not only are T waves in V2-thru-V4 disproportionately tall and peaked — but there is ST segment straightening that is clearly abnormal in lead V5 and V6 (as well as having inappropriately peaked T waves in those leads) — plus subtle-but-real ST-T wave abnormalities in each of the inferior leads — that in a patient brought by ambulance for chest pain says, “I am an acute or very recent STEMI until you prove otherwise!” Anyone who has studied computerized ECG interpretations will not be surprised by the fact that these clearly abnormal ECG findings were totally missed by the computer report. Except for “true expert interpreters” (ie, clinicians who have read many, many thousands of acute and non-acute tracings over time — NO provider (in my opinion) should ever look at the computerized report BEFORE they have completed their own unbiased ECG interpretation. Following this simple advice would greatly reduce the chance of overlooking the subtle-but-real series of abnormalities that are clearly present on this initial ECG. THANKS so much to Dr. Smith for posting this case!


Recommended Resources