How AI Reads an ECG: Inside Automated Rhythm Interpretation

Every clinician has seen it: the computer-generated interpretation printed across the top of a 12-lead ECG. "Normal sinus rhythm." "Possible anterior infarct — age undetermined." Those few lines are the output of decades of work in automated ECG analysis — and the technology behind them has changed more in the last ten years than in the previous fifty.

So how does a machine actually read an ECG? And how much should you trust it?

The First Generation: Rule-Based Algorithms

The earliest computerised ECG interpreters, dating back to the 1950s and 60s, were rule-based expert systems. Engineers and cardiologists translated the same criteria taught in medical school into explicit if-then logic: measure the QRS width, measure the PR interval, check the R-wave amplitude in specific leads, and flag whatever crosses a threshold.

This is essentially how the classic Minnesota Code and its descendants work. The algorithm:

• Detects the beats — finds each QRS complex and segments the P, QRS, and T components.
• Measures intervals and amplitudes — PR, QRS, QT, ST-segment deviation, axis.
• Applies diagnostic criteria — compares those measurements to published thresholds (for example, ST elevation ≥1 mm in two contiguous leads).

Rule-based systems are transparent and predictable — you can trace exactly why the machine said what it said. But they are also brittle. They struggle with noisy tracings, overlapping abnormalities, and the enormous biological variability between patients. This is why the printed interpretation has always carried a warning to confirm with a physician.

The Second Generation: Deep Learning

The modern approach flips the logic. Instead of a human writing the rules, a neural network learns them from data. A convolutional neural network (CNN) is trained on hundreds of thousands — sometimes millions — of ECGs, each labelled with a known diagnosis. The model gradually learns which patterns in the raw voltage signal correspond to which conditions.

The results have been striking. In a landmark 2019 study published in Nature Medicine, a deep neural network trained on more than 90,000 single-lead recordings detected and classified arrhythmias at a level comparable to board-certified cardiologists. Other models have learned to do things no human can do by eye at all — such as estimating a patient's likelihood of low left-ventricular ejection fraction, or flagging a risk of future atrial fibrillation, from an ECG that looks completely normal to a clinician.

Deep learning shines where rule-based systems fail:

• It tolerates noise and artifact far better.
• It captures subtle, distributed patterns that resist simple thresholds.
• It improves as more labelled data becomes available.

Reading a Photo of a Strip

Both approaches above assume access to the raw digital signal from the ECG machine. But a huge amount of real-world ECG interpretation happens from a picture — a photo of a printed strip, a screenshot of a monitor, a scanned rhythm from a textbook.

Interpreting an image adds a computer-vision step. The system first has to reconstruct the waveform from pixels — separating the trace from the background grid, correcting for angle and lighting, and converting the curve back into a signal it can measure. Only then can rhythm analysis begin. Modern multimodal models — including large language models that accept images — can now describe a rhythm strip, estimate the rate, and explain the reasoning in plain language, which makes them especially useful as teaching tools.

Where AI Still Falls Short

Automated interpretation is a powerful assistant, not an oracle. Every clinician using it should understand its limits.

• Single-lead vs. 12-lead. Wearable and handheld devices usually capture one lead. That's excellent for detecting atrial fibrillation but cannot localise ischaemia the way a full 12-lead can.
• Artifact and edge cases. Motion, poor electrode contact, and pacing spikes still fool automated systems — sometimes confidently.
• Automation bias. The most studied risk isn't the machine — it's the human. Clinicians who over-trust a confident but wrong interpretation can be led astray. The computer read is a second opinion, never the final word.
• Correlation without context. An algorithm sees the tracing, not the patient. Chest pain, vital signs, and history still determine what the ECG means.

How to Use AI ECG Tools Well

• Treat the interpretation as a prompt to look more carefully, not as a conclusion.
• Always correlate with the clinical picture and, when it matters, a repeat or serial ECG.
• Know what data your tool was validated on and what leads it uses.
• Use AI to learn — have it explain its reasoning, then check that reasoning against first principles.

Practicing in CardioLens

The CardioLens AI Scanner is built exactly for that partnership. Snap a photo of any ECG strip and Claude AI identifies the rhythm, calculates the rate, measures intervals, and — most importantly for learning — explains the clinical teaching points behind its read. It's designed as an educational assistant: a fast, transparent second look that helps you build your own pattern recognition, not a substitute for it.