Computer Interpretations of EKGs: Accuracy and Clinical Use
Computer-interpreted EKGs are moderately accurate but require mandatory physician overread—they should never be used as standalone diagnostic tools, as critical errors occur frequently enough to impact patient morbidity and mortality. 1, 2
The Core Problem with Computer Interpretation
While computerized ECG systems are now ubiquitous (processing over 100 million ECGs yearly in the US), the AHA/ACC/Heart Rhythm Society guidelines explicitly acknowledge that different automated systems produce significant differences in measurements and diagnostic statements that can alter clinical interpretation 1, 2. This variability alone makes reliance on computer interpretation dangerous.
Performance Data: Where Computers Succeed and Fail
Areas of Reasonable Accuracy:
- Normal ECGs: Computer interpretations of truly normal ECGs are highly reliable, with accuracy rates of 96.7-99.5% 3, 4
- Sinus rhythm: Sensitivity 96.6%, specificity 97.0% 3
- Ruling out STEMI: When computers read "normal" or "otherwise normal," negative predictive value approaches 100% for STEMI 4
Critical Areas of Poor Performance:
- Right ventricular hypertrophy: Median sensitivity only 31.8% (vs. 46.6% for cardiologists) 5
- Left ventricular hypertrophy: Sensitivity 56.6% (vs. 63.9% for cardiologists), positive predictive accuracy only 74.2% 5, 3
- Inferior MI: Sensitivity 58.8% (vs. 71.7% for cardiologists) 5
- ST-T wave changes: Lowest accuracy at 83-84% 3
- Overall accuracy: Computers achieve 69.7% total accuracy vs. 76.3% for cardiologists—a clinically significant 6.6% gap 5
The Clinical Algorithm
For ED triage ECGs interpreted as "normal" or "otherwise normal":
- These can safely wait for physician review until patient evaluation without delaying STEMI diagnosis 4
- No emergent intervention needed based on computer interpretation alone
For ANY abnormal computer interpretation:
- Mandatory immediate physician overread required 2, 6, 3
- Computer errors in detecting arrhythmias, pacemakers, and myocardial infarctions are common enough to be clinically dangerous 6
- In one study, reviewing physicians modified or deleted 24% of computer statements (1792 deletions, 484 modifications, 1320 additions out of 13,375 statements) 3
Common Pitfalls to Avoid
Never trust computer interpretation for hypertrophy diagnoses—positive predictive accuracy is only 74%, meaning 1 in 4 computer-diagnosed hypertrophies are wrong 3
Beware of artifact misinterpretation—computers cannot distinguish muscle artifacts or lead reversal, which physicians can identify with clinical context 6
Don't assume newer is better—while AI models like ECG-GPT show promise with AUROCs of 0.80-0.96 for various conditions 7, these are research tools not yet validated for clinical decision-making
System variability matters—different ECG machines use different algorithms; a "normal" on one system may not be "normal" on another 1, 2
The Bottom Line for Practice
The computer is a screening tool, not a diagnostic tool. The AHA guidelines are unequivocal: "physician overreading and confirmation of computer-based ECGs is required" 2. While computers reduce physician reading time and accurately interpret most normal ECGs, their error rate in abnormal ECGs—particularly for conditions like RVH, LVH, and inferior MI—is high enough that relying on them alone will miss life-threatening diagnoses and harm patients 6, 5.
In practical terms: Use computer "normal" interpretations to prioritize workflow, but every abnormal interpretation and every ECG in a patient with concerning symptoms requires immediate physician review regardless of what the computer says.