This was sent to me by a former resident from a community hospital:
A middle-aged woman complained of chest pain and was seen in triage.
She had a ECG recorded.
The computer interpreted the ECG (GE Marquette 12 SL) as:
"Sinus Bradycardia. Normal ECG."
It was not seen by a physician.
Here it is:
The ECG was not shown to a physician.
Later, the patient collapsed in triage and became unresponsive. She was in full arrest with ventricular fibrillation.
Fortunately, she was resuscitated.
She had a 100% LAD occlusion.
She survived neurologicially intact. I don't know what her subsequent cardiac function was, but that is not the point of this post.
But it could have been a disaster.
Comment
This paper has received some press recently:
It purports to show that you don't need to read the ECG if the computer says "normal". That it saves you from pesky interruptions. Many on Twitter seem to agree.
222 of 855 (26%) were interpreted as normal. The confidence intervals were 97-100%. Would you like to be 97% sure you're not missing an emergency?
Neural Networks are coming:
We compared a new deep neural network (DNN) machine learning artificial intelligence algorithm from Cardiologs technologies(DNN) to Veritas conventional algorithm in the interpretation of 1473 ED ECGs:
Veritas: Of 364 ECGs (24% of total) that it interpreted as "normal," is missed 5 emergencies.
DNN: Of 493 ECGs (33% of total) that it interpreted as "normal," it missed 2 emergencies.
As you can see, it had both fewer false positives and fewer false negatives. That is powerful, and it will only get better as it learns from more and more ECGs.
So things will get better as technology advances.
By the way, this deep neural network algorithm does calculate the 3-variable "Smith score" and displays it for the physician. It did not use machine learning for this task; it was programmed.
Learning Points:
1. It is hazardous to trust the computer, even when it interprets "Normal"
2. I never trust a computer interpretation. I only look at it after I have read the ECG, and only in order to see if it found something that I overlooked.
3. If the computer says "normal", it takes very little time to confirm this with visualization (at least for me, and anyone can get good at it with practice)
4. Learn to recognize LAD occlusion. It is not always obvious. And the computer will not help you.
Here are some previous posts in which the computer interpreted the ECG as "normal."
A middle-aged woman complained of chest pain and was seen in triage.
She had a ECG recorded.
The computer interpreted the ECG (GE Marquette 12 SL) as:
"Sinus Bradycardia. Normal ECG."
It was not seen by a physician.
Here it is:
 |
| Computer interpretation: "Sinus bradycardia. Normal ECG" GE Marquette 12 SL algorithm |
--This ECG does, at first glance, appear normal
--It is sinus rhythm.
--The QRS is normal
--There is possibly an abnormal Q-wave in lead III, and non-specific T-wave in aVL, but the computer did not see it.
--All the intervals are normal, including the PR, QRS, and QTc at 455 ms
--There is ST elevation in V2-V4, but it is within normal limits.
That is, it appears to be normal ST Elevation and the computer thinks it is normal STE.
There is one thing that is clearly abnormal, but computers are not programmed to detect it:
When there is ST Elevation in V2-V4, it is only normal when there is:
1) high R-wave voltage and
2) a short QTc
3) relatively high QRS voltage in V2
This ECG has a relatively long QTc, low R-wave amplitude, and low QRS amplitude
That is why the formulas work.
STE60V3 = 2.5 mm, QTc = 455 ms, RAV4 = 6 mm, QRSV2 = 11
3-Variable formula = 27.9 (very high, diagnostic of LAD occlusion)
4-Variable formula = 23.05 (very high)
To my knowledge, no computer algorithm has yet been programmed with the formulas
Years later, I sent it to the Queen of Hearts:
The Queen of Hearts PM Cardio App is now available in the European Union (CE approved) the App Store and on Google Play. For Americans, you need to wait for the FDA. But in the meantime:
YOU HAVE THE OPPORTUNITY TO GET EARLY ACCESS TO THE PM Cardio AI BOT!! (THE PM CARDIO OMI AI APP)
If you want this bot to help you make the early diagnosis of OMI and save your patient and his/her myocardium, you can sign up to get an early beta version of the bot here. It is not yet available, but this is your way to get on the list.
____________________
The ECG was not shown to a physician.
Later, the patient collapsed in triage and became unresponsive. She was in full arrest with ventricular fibrillation.
Fortunately, she was resuscitated.
She had a 100% LAD occlusion.
She survived neurologicially intact. I don't know what her subsequent cardiac function was, but that is not the point of this post.
But it could have been a disaster.
Comment
This paper has received some press recently:
Safety of Computer Interpretation of Normal Triage Electrocardiograms
The algorithm used was also the GE Marquette 12 SL.It purports to show that you don't need to read the ECG if the computer says "normal". That it saves you from pesky interruptions. Many on Twitter seem to agree.
222 of 855 (26%) were interpreted as normal. The confidence intervals were 97-100%. Would you like to be 97% sure you're not missing an emergency?
Neural Networks are coming:
We compared a new deep neural network (DNN) machine learning artificial intelligence algorithm from Cardiologs technologies(DNN) to Veritas conventional algorithm in the interpretation of 1473 ED ECGs:
Veritas: Of 364 ECGs (24% of total) that it interpreted as "normal," is missed 5 emergencies.
DNN: Of 493 ECGs (33% of total) that it interpreted as "normal," it missed 2 emergencies.
As you can see, it had both fewer false positives and fewer false negatives. That is powerful, and it will only get better as it learns from more and more ECGs.
So things will get better as technology advances.
By the way, this deep neural network algorithm does calculate the 3-variable "Smith score" and displays it for the physician. It did not use machine learning for this task; it was programmed.
Learning Points:
1. It is hazardous to trust the computer, even when it interprets "Normal"
2. I never trust a computer interpretation. I only look at it after I have read the ECG, and only in order to see if it found something that I overlooked.
3. If the computer says "normal", it takes very little time to confirm this with visualization (at least for me, and anyone can get good at it with practice)
4. Learn to recognize LAD occlusion. It is not always obvious. And the computer will not help you.
Here are some previous posts in which the computer interpreted the ECG as "normal."
Chest Pain Diagnosed as Gastroesophageal Reflux
Another case of arrest:
Chest pain relieved by Maalox and viscous lidocaine
Another case of arrest:
Another case of arrest:
It is easy to be led astray by the computer....
Highly illustrative case! I’ll add a few points and another perspective to Dr. Smith’s superb comments. In general — computerized interpretations excel at measurements (ie, rate; axis; intervals). They are usually accurate when they say an ECG is normal — BUT — there are enough occasions when computerized interpretations overlook subtle but extremely important findings of ischemia/early infarction. The tracing in this case is but one such example. Thus, sensitivity of the computer report for acute changes is just NOT good enough to be relied upon for interpretation of the ECG in a patient with chest pain. In my experience — computerized interpretations are also unreliable in the interpretation of non-sinus rhythms of any complexity. And importantly — computerized interpretations are still inferior to the ability of “Gestalt analysis” that an astute and experienced electrocardiographer develops over many years of high-volume interpretation. By that I mean that the ECG in this case just LOOKS ABNORMAL to me, even before I dissect the specific abnormalities present in the various leads. In my study comparing skilled electrocardiographers to computerized interpretations — the computer literally tripled my speed for interpretation of less complex tracings once I became familiar with the specifics of the interpretation program that our hospital was using. That’s because I no longer had to measure rate, axis or intervals — and if the ECG said “normal tracing”, it literally took me no more than 3 seconds to overview and sign off on the tracing if I agreed the ECG was normal. But this Gestalt ability came to me only after well over 20 years of high-volume interpretation. For the vast majority of non-expert interpreters — one should always cover up the computerized interpretation until AFTER you have completed your own interpretation. That way you will not be biased if the computer says “normal” and you disagree. And that way, the computerized report can HELP by sometimes providing another opinion on some aspect of the tracing that you may not have considered …
ReplyDeleteAs to this ECG — leads III and aVF are clearly abnormal, because the QTc is relatively long for the rate; the ST segment in these leads is flat (it should be gently upsloping); and the T waves are clearly disproportionately tall (as well as “fatter”-than-they-should-be at their peak) given the small QRS amplitude. In the context of these definite ST-T wave abnormalities in these 2 inferior leads (plus the fragmented Q in lead III) — it is likely that the T wave in the other inferior lead ( = lead II) is also a bit “fatter”-than-it-should-be by the concept of “patterns of leads” (ie, Always look EXTRA CLOSE in neighboring leads for subtle abnormalities). Any doubt about whether these inferior ST-T wave changes are real should be removed by a glance at lead aVL — which I interpret as consistent with a reciprocal change to the ST-T wave in lead III.
In the chest leads — the most abnormal lead is lead V3. As per Dr. Smith, in view of minimal r wave amplitude — the 2-3mm of J-point ST elevation is more-than-expected in V3. But even more than this, is the STRAIGHTENING of the ST segment in each of the 3 QRS complexes we see in lead V3. This is real! Applying the concept of “neighboring leads” — the T waves in leads V4, V5, and probably also V6 are all “fatter”-than-they-should-be ( = hyperacute!). In the context of what we see in leads V3-V6 — we can now surmise that the coved ST segment in lead V2 is abnormal in shape (it should be concave up). Finally, given these abnormalities in leads V2-V5 — lead V1 shows uncharacteristic ST elevation, as well as ST coving. This ST-T wave in lead V1 would not necessarily be abnormal if it was an isolated finding — but in context of the other 9 leads showing clear abnormalities — I believe lead V1 is also abnormal. In summary, it should be no surprise that this patient with new chest pain arrested not long after this tracing was done ….
It seems very probable that deep learning will be better in reading ecgs than any human. How soon?
ReplyDeleteIt takes time: you need to teach it with many many cases and you have to have both outcomes and the digital file
Deletehi Steve
ReplyDeleteexcellent. some also say look for big fat T waves , esp bigger than the associated R wave. (which i think is the lazy, quick, less efficient way of what you do with the formula.
i'll tell you, i never ever look at the top, the machine's read; which is probably not the wisest thing, as you described above. but often i find it wrong, wordy, unhelpful.
thank you, agian.
tom
Hi dr Smith!
ReplyDeleteVery instructive case! At your facility, and based on this ECG alone, would you have gone for immediate angio or would you have treated aggressively with repeat ECG and go for angio only if persistent pain and/or persistent/evolving ECG changes?
Jacob
I would only NOT activate the cath lab if a high quality echo showed no wall motion abnormality. If serial ECGs did not prove occluison, then I would get that and if I couldn't get that immediately, activate the lab.
DeleteGreat comment's Ken.The T in aVF did riddle me.
ReplyDelete@ AKS — Thanks — :)
Delete