Does AI help doctors read post-arrest ECGs? Specificity improved, but wrong advice pulled decisions with it
Fifteen physicians made 2,070 interpretations of 69 ECGs recorded after cardiac arrest. Algorithm support raised AUC from 0.882 to 0.912 and specificity from 84.64% to 90.43%, while 97% of changed decisions followed the AI—including cases where it was wrong.
The short version
- What happened
- Fifteen physicians interpreted 69 real-world post-arrest ECGs twice, without and then with Queen of Hearts output, for 2,070 interpretations.
- Who it affects
- Doctors may gain a useful second opinion on difficult post-arrest ECGs, particularly by reducing false-positive occlusion calls.
- What remains unproven
- The study used 69 retrospective ECGs and 15 physicians, not a prospective clinical deployment.
Keep in mind
The study used 69 retrospective ECGs and 15 physicians, not a prospective clinical deployment.
Unaided interpretation always preceded assisted interpretation, so a second-look effect cannot be fully excluded.
The overall error-rate change and adjusted odds of correctness were not statistically significant.
No patient outcome, treatment timing, procedure rate or survival effect was measured.
Source analysis · 1 linked source. Primary analysed source date: 4 November 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The direct finding: discrimination and specificity improved, but errors followed the algorithm
A multi-reader study in Scientific Reports found that showing doctors an AI assessment modestly improved their ability to distinguish occlusive myocardial infarction on electrocardiograms recorded after cardiac arrest. Fifteen physicians interpreted 69 anonymized ECGs after return of spontaneous circulation, first without and then with output from the Queen of Hearts algorithm. Across all readers and cases, the area under the receiver-operating-characteristic curve rose from 0.882 to 0.912, an absolute increase of 0.030 with a 95% confidence interval of 0.008 to 0.058 and a p value of 0.018. Specificity increased from 84.64% to 90.43%, while sensitivity moved only from 77.97% to 78.55%.
The human-factor result is at least as important. Readers changed 66 decisions after seeing the algorithm and moved in its direction in 64, or 97%, of them. When the AI was correct, assistance reduced the doctors' error rate by 5.03 percentage points. When it was wrong, assistance increased their error rate by 10.56 points. The study therefore supports AI as influential decision support, not as a uniformly protective second reader. A tool that improves average discrimination can still concentrate harm when clinicians defer to an incorrect output.[1]
Why post-arrest ECG interpretation is difficult
An occlusive myocardial infarction occurs when a coronary artery blockage deprives heart muscle of blood and may require urgent angiography and reperfusion. After cardiac arrest, the ECG can be distorted by global oxygen deprivation, resuscitation effects, metabolic disturbance and conduction abnormalities. Traditional ST-elevation criteria can miss occlusions without classic elevation and can also produce false alarms. Clinicians must decide quickly while balancing the risks of delayed treatment against unnecessary invasive procedures in a critically ill patient.
Queen of Hearts is designed to identify an occlusion pattern rather than rely solely on the conventional STEMI label. The study asks whether its probability or classification helps doctors interpret this unusually noisy setting. That is narrower than asking whether the model improves survival or should control catheter-laboratory activation. The reference diagnosis came from clinical information and angiographic assessment in the retrospective cases, while the experimental outcome was a reader's interpretation of a static ECG. Treatment timing, neurological prognosis, bleeding risk, transfer logistics and other bedside factors were not part of the decision task.[1]
The distinction that matters
What changed when doctors saw the AI output
The average result improved, but the direction of change depended strongly on whether the algorithm itself was correct.
Unaided reading
15 doctors interpret 69 ECGs
Baseline AUC 0.882
AI shown
Readers reconsider the same cases
AUC 0.912 and specificity 90.43%
Changed decisions
66 answers moved
64 followed the AI
AI correct
Support guides reconsideration
Error rate fell 5.03 points
AI wrong
Support pulls toward the error
Error rate rose 10.56 points
The denominator: 69 ECGs, 15 readers and 2,070 interpretations
The investigators selected 69 real-world ECGs recorded after successful resuscitation. Fifteen physicians from five professional groups reviewed every case at two timepoints, creating 69 × 15 × 2, or 2,070, interpretations. The repeated-reader, repeated-case design allows the analysis to separate variation between doctors from variation between ECGs more carefully than a simple accuracy calculation. It also exposes whether support changes confidence and agreement. The paper reports that both rose after AI output was displayed, consistent with decisions converging on the algorithm.
That convergence can be useful when the model is right, but agreement is not an independent validation signal when everyone sees the same recommendation. Because the unaided review always came first, the design cannot fully separate the effect of AI from a second look at a familiar tracing. A randomized crossover—with some cases presented in the opposite order, adequate washout and different but matched ECG sets—would better isolate support from repetition. The sample also remains modest: 69 cases can estimate reader behaviour in this controlled task, not performance across all arrest causes, hospitals, devices and patient populations.[1]
The average improvement did not reach every clinically important outcome
The primary AUC change was statistically significant, and the specificity gain suggests fewer false-positive occlusion calls. Yet the absolute sensitivity increase was only 0.58 percentage points. The reported change in overall error rate was a 2.32-point reduction, with a confidence interval spanning from a 0.9-point increase to a 5.6-point reduction and a p value of 0.156. In other words, this study did not establish a statistically significant fall in the overall proportion of wrong decisions. An adjusted model estimated 1.29 times the odds of a correct answer with support, but its 95% confidence interval of 0.98 to 1.69 also included no effect.
A clinically important subgroup remained difficult. Sensitivity did not improve for high-risk occlusive infarctions that lacked conventional ST elevation—the cases in which an expanded occlusion framework is supposed to add particular value. Subgroup denominators are smaller and estimates less precise, so that result should not be treated as proof the model cannot help. It does mean the broad AUC headline cannot substitute for reporting performance where conventional criteria fail. Deployment decisions should prioritise missed occlusions, false catheter-lab activations and time to reperfusion, not only a composite discrimination score.[1]
Automation bias is visible rather than theoretical
The 64 of 66 alignment figure shows that the model output was highly persuasive. It does not prove irrational reliance: clinicians may have changed only when the AI gave a compelling reason. But the asymmetric error changes show the risk. Correct advice delivered a five-point benefit, while incorrect advice produced a larger ten-and-a-half-point harm in reader error. A hospital evaluating this system needs to know how often those wrong recommendations occur in its population and whether the interface helps clinicians detect them. Average model accuracy alone cannot answer that question.
Safer design could hide the prediction until a clinician records an initial assessment, display the ECG features supporting the result, flag low-confidence or out-of-distribution cases and require escalation rather than automatic action. Training should include examples where the model is deliberately wrong, so users practise disagreement. Monitoring should track overrides, concordance, false negatives and downstream care, with a way to suspend the tool if performance drifts. None of those mitigations is validated by this paper. The experiment demonstrates influence under controlled conditions; it does not determine the best interface or governance system.[1]
What changes for emergency care—and what evidence would change this assessment
For emergency physicians and cardiologists, the study gives a reason to test AI-assisted ECG interpretation prospectively, especially as a second opinion that may reduce false-positive occlusion calls. It does not justify replacing expert review or automatically activating an invasive pathway. Patients should understand that the evaluated outcome was interpretation accuracy, not survival, neurological recovery, procedure complications or length of stay. The retrospective records were anonymized, ethics review waived consent for the case data, and participating physicians gave consent; the authors declared no competing interests.
Confidence would rise with a preregistered, multi-centre randomized crossover and then a silent prospective study across different ECG machines, arrest rhythms and care systems. A clinical trial should measure time to angiography, missed occlusions, unnecessary procedures, survival and neurological outcomes, while reporting performance when the AI is wrong. Independent calibration and subgroup results are essential. The assessment would weaken if gains disappear in prospective use or automation bias increases harmful actions. This is consequential peer-reviewed research but not Breaking: it is a controlled reader study, not a live clinical alert or proven change in patient outcomes.[1]