Computer STEMI Interpretation

Several years ago I told a group of paramedics that while not relaying on the computer interpretation in the case of STEMI, they should consider it as a fellow paramedic offering his opinion. I told them that the studies showed a 60% sensitivity and a 90% specificity in identifying STEMI. In other words, the computer would only identify 60% of the STEMIs, but if it identified a STEMI, it was going to be right about 90% of the time. I based this on several older studies I had read.

I can tell you as a paramedic when I suspect a patient may be having a STEMI and the computer agrees and the tracing meets criteria, I feel much better about calling a STEMI alert. I confess once when I had a patient with an ECG suggesting a small elevation in I and AVL. I did 9 ECGs en route to the hospital, and finally on the 5th one, the computer came around and agreed with my suspicion that the patient was possibly having a STEMI, giving me the maniless to call in the alert (I know I was being a wimp). Though the computer reversed course again on the 6th and 9th ECG, the patient was in fact having a lateral STEMI. He coded in the ED, was revived and successfully had his 100% occlusion in his circumflex artery stented.

Conversely when the computer calls it a STEMI, but I disagree with the reading, I recheck my lead placement and do another. Some make me doubt my reading ability, others make me want to put a dunce cap on the monitor and put in the corner of the ambulance facing the wall. You weren’t paying attention in paramedic school obviously, bad monitor!

A new study in the April-June 2013 Prehospital Emergency Care, Prehospital Electrocardiographic Computer Identification of ST-segment Elevation Myocardial Infarction, reports a 58% sensitivity and a 100% specificity for computer interpretation.

They reviewed the transmitted 12-leads of a 100 consecutive STEMIs and found 58 of them had the interpretation “Acute MI Suspected” on the 12-lead. Then to determine specificity, they analyzed 100 random 12-leads from non STEMI patients and found none of them incorrectly said “Acute MI Suspected.”

I must ask: Does this high specificity match anyone’s experience? Or even come close?

How often do you do a 12-lead and it says ***Acute MI Suspected*** or “Consider Acute MI” depending on the monitor you use, and you can look at the 12-lead and say No. Not even close.

In my hospital I receive an email each time a 12-lead is transmitted. I enjoy reading the strip. Our medical director and I play a STEMI game where we immediately email each other with our vote STEMI or no STEMI assuming the patient always to be symptomatic. I can tell you the number of 12-leads transmitted that say “Consider Acute MI” is way more than the number that actually are STEMIs.

Consider these:

The computer seems particularly inaccurate in wide complex and tachycardic strips.

I have had to go back to the medics and say the studies are wrong. Don’t put as much stock in the computer interpretation as I suggested you should. It is not close to 90% accurate.

An informal tally using the same criteria as the 2012 study for sensitivity shows the Life Pack 15 is about 60%. It properly identifies 60% of STEMIs, but misses 40%.

But then when I analyze specificity, I admittedly use a different criteria, which will yield a markedly different, but more telling conclusion. I ask, what percentage of the time when the computer reads “Meets ST Elevation Criteria” and “Consider Acute Infarct,” is it actually a STEMI? And the answer in this small sample (50 cases) is 32%. Not 100%.

Note: The published study cited above used the Life Pack 12, my figures are for the LifePack 15, which uses a different software.

The computer seems to consistently misread tachycardic rhythms and wide complex rhythms.

These, in my mind, are the proper questions we should be asking if we are going to incorporate the machine reading into our decisions:

What percent of the time is it a STEMI and the machine missed it? Seems to run about 60%

When the machine does call it a STEMI, what percentage of the time is it right? That answer is clearly not 100%, or even close.

Obviously, this should be confirmed in formal studies, which can be easily done at any PCI facility.

Does this mean I would be in favor of removing the interpretation from the machines? No, I think they can be valuable. I am just saying we shouldn’t be misleading others and ourselves by claiming the machines are better than they are.

When making policy that incorporates the computer reading into the decision making, we need to understand its limitations.

We need to make certain that paramedics are not relaying on the computer alone to call, and that if the machine does read STEMI, the medic makes agrees with the interpretation.

And we need to continue to work to improve basic paramedic interpretation, as well as proper placement and technique of ECG aquisition.

This will help limit false activations for services that cannot transmit.

Note: In our state we use the following to help determine diversion to a PCI center:

1. Active chest pain or equivalent symptoms (nausea, SOB)

2. 12-Lead ECG of good quality showing STEMI

a. ST-elevation
i. > or = 2mm in 2 contiguous leads (V1-V4), and/or
ii. > or = 1 mm in 2 contiguous leads (limb, lateral)

b. QRS duration < or = 12 seconds

c. ***ACUTE MI*** or equivalent prints on 12-lead ECG; paramedic agrees

16 Comments

  • Dina says:

    Only proving yet once again that 1) it’s the humans who ultimately make the interpretations and decisisons; and 2) we are to treat the patient and not the machine!

    • medicscribe says:

      Hi Dina. I agree with you on number 1, but on number 2, I think STEMI is one area where the monitor can provide information I can’t get elsewhere. The issue is not the 12-lead, but the machine’s interpretation of the 12-lead. Thanks again for commenting, Peter

  • Christopher says:

    [aside: I loathe “treat the patient, not the monitor”. You do your patient a disservice whenever you do this. Treating the patient includes every piece of information at your disposal; ignoring the monitor because you don’t understand it is hardly a good idea.]

    The cardiac monitor is only as good as the data it receives. Poor data quality? Poor interpretations. (I mean how does a human even begin to interpret ECG #3?)

    If the cardiac monitor does not have good data…you won’t have good data either! You’ll make poor interpretations too.

    Atrial flutter, paced rhythms, LVH, etc…all of these fool paramedics probably at least as often as it fools the computerized interpretation (perhaps less with high data quality).

    Most studies involving the computerized algorithms feature high quality tracings without artifact or significant wander. With high data quality you should trust (but verify) the computerized interpretation.

    Perhaps the most legitimate criticism of most of the current computerized algorithms is their heuristics which allow them to state STEMI in the face of profound tachycardia.

    Something that is useful with a Marquette algorithm (LP12, Zoll E/M-series): If you see “ectopic atrial _____” it is pretty sensitive for lead swaps; the lead reversal message only happens with classic LA-RA swaps

    • medicscribe says:

      Thanks for the comments, Christopher. I know poor data quality can affect the reading. I guess my issue is that given that poor data quality seems to be a general problem, the problem is exacerbated by systems building STEMI responses based on an algorithm that often does not work because of poor data quality. I worry that systems reading the literature will have false confidence in their system and patients may be harmed because of it. Peter

  • Curt says:

    I’ve seen HeartStart Mrx’s consistently call ventricular pacemakers STEMI’s.

    • medicscribe says:

      I have not seen this consitently with the Lifepacks, but I have seen it. One of our false calls was a clear ventricular pacer, Thanks for the comment.

  • Richard Pilbery says:

    Please correct me if I am wrong, but I thought specificity is a measure of how good the test is at correctly identifying people who are well. I think what you are describing (“if it identified a STEMI, it was going to be right about 90% of the time”) is actually the positive predictive value, i.e. the chance that a positive test result will be correct.

    Sensitivity, specificity, positive and negative predictive values have left me with quite a headache. I recommend this article from the British Medical Journal as a cure: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC200804/pdf/bmj32700716.pdf

    • medicscribe says:

      Hi Richard. I have a hard time grasping the specificity concept. I thought I was using a similar formula as the authors of the study. I will check out the article you recommend. Thanks, Peter

  • DaveB says:

    “Treat the patient, not the machine”. I am sorry, but that is really a meaningless comment. Skillfully interpreting a patient’s 12 lead IS part of treating the patient. Usually, that comment is the result of throwing our hands up because we can not interpret what we see. It happens to all of us sometimes, but more often than not it is a sign that more education is needed, not a sign to disregard the “machine”.

    Even if we are to trust the computerized interpretation given high data quality, it does not address the issue of poor sensitivity.

    Those that have been improving their 12 lead interpretation skills realize that it is more about morphologic criteria and less about measured intervals.

    I believe a great number of STEMIs caught by the monitor would and should be caught by the provider as well. But what happens to all of the subtle ECGs with hyperacute T waves, or inferior ST elevation of 0.75 mm with T wave inversion in aVL?

    Yes, it is nice when we see a STEMI and the computer agrees. It is not nice when we miss >40% of them because the computer did not catch them. That is what we are here to do.

    While it may be an added convenience (debatable), we should not rely on the computerized interpretation at all. If we can not operate efficiently without it, an educational gap has been identified.

  • Brandon O says:

    Like the folks said: quality matters, tachycardias give it a hard time, and in particular, all the old data on the Marquette algorithm is becoming increasingly less relevant since the newfangled models of ALL the major prehospital monitors now use different algorithms — which, as a rule, seem to be less specific, although perhaps more sensitive.

    I agree with using it as an “expert opinion,” but would ignore it almost entirely when there’s lots of artifact, tachyarrhythmias, or other weirdness.

  • deezy says:

    Food for thought:

    We use the physio control monitors. Lifepak 12 and 15’s.

    Our protocols state that if the monitor’s algorhythm states that it suspects a STEMI on two consecutive 12 lead ECGs, a STEMI alert must be activated.
    Shame, right?

    Or…the typical STEMI alert criteria with the ST elevation in more than 2 contiguous leads

    • medicscribe says:

      Thanks for the comment, deezy. I like at least that your system requires two consecutive readings. I know of many cases where I have gotten one that was clearly due to artifact or a bump in the road or sudden arm movement. Still I am troubled by the use of the computer interpretation in systems, including my own, without the people making the guidelines perhaps being fully aware for the real world track record of the computer. Peter

  • Christopher says:

    Peter,

    I think if data quality is a systemic issue which confounds computerized interpretation…it likely means paramedic interpretation will also be confounded systemically. With good quality ECG’s I see a large amount of agreement between paramedics and the computer. With lesser quality ECG’s I see larger amounts of disagreement. Ultimately you need a clinician’s interpretation of the 12-Lead, which could be the paramedic in the field or transmission to a receiving facility otherwise.

    A decent middle of the road for systems where paramedic interpretation is considered a non-starter is to simply require the paramedic to concur with the machine interpretation; or switch to a transmission protocol.

  • For those of you who think that computer interpretations are so smart should look at this.

    http://tooldtowork.com/2013/04/im-not-convinced/

    Tell me (over at my blog) what you think this is and then I’ll tell you what the computer told the ER staff it was.

1 Trackback

Leave a Reply to Brandon O Cancel reply

Your email address will not be published. Required fields are marked *