ECG Waveform Classifier Lab

What PTB-XL Is, and Why It Changed ECG Machine Learning

Training a reliable ECG classifier requires two things that used to be scarce together: a large volume of real recordings, and trustworthy labels attached to them. PTB-XL, released by researchers building on the long-running Physikalisch-Technische Bundesanstalt ECG database (Wagner et al.), addressed both at once. It provides over 21,000 clinical 12-lead ECG recordings, each ten seconds long, drawn from a genuinely diverse patient population, and each one carries diagnostic labels assigned by expert cardiologists across dozens of rhythm and morphology categories, from normal sinus rhythm through various conduction abnormalities, hypertrophy patterns and myocardial infarction subtypes, with many records receiving more than one label at once, since real ECGs frequently show overlapping findings. That combination of scale and annotation quality is precisely what made it practical to train deep learning models directly on the raw waveform signal, feeding a neural network the actual voltage-over-time trace rather than a small hand-picked set of measurements like QRS duration or PR interval alone. Earlier ECG classification approaches often relied on exactly those hand-engineered features because there wasn't enough richly labeled raw-signal data to learn more nuanced patterns directly. Datasets at PTB-XL's scale let models instead learn which subtle shapes, slopes and combinations of features in the raw trace correlate with which diagnostic categories, sometimes surfacing patterns that are hard to fully capture in a short list of interval measurements, while still, importantly, remaining broadly consistent with the same standardized clinical criteria cardiologists have used for decades. This simulation demonstrates the standardized-criteria side of that picture in an intentionally simple, rule-based way.

QRS Duration: Reading the Ventricles' Electrical Wiring

The QRS complex is the sharp spike in an ECG trace that represents ventricular depolarization, the electrical signal spreading through the heart's two large pumping chambers to trigger a contraction. In a healthy heart, that signal travels down a fast, dedicated conduction pathway, the bundle branches, so the whole ventricles activate together almost instantly; a normal QRS complex typically lasts only about 80 to 100 milliseconds. If one of those bundle branches is blocked or delayed, the electrical wave can no longer spread through both ventricles simultaneously via the fast pathway; it has to detour through slower, ordinary heart-muscle tissue to reach the delayed side, and that detour takes measurably longer. Standard ECG interpretation criteria set the cut-off at 120 milliseconds: a QRS complex at or beyond that duration is the accepted threshold for suspecting a bundle branch block. This simulation uses that same 120-millisecond figure, drawn directly from standard ECG interpretation criteria, not an invented number, and widens the drawn QRS complex proportionally as you move the slider so the visual widening and the numeric threshold move together.

PR Interval: Timing the Handoff Through the AV Node

Before the ventricles can depolarize, the electrical signal has to first spread across the atria, producing the small, rounded P wave, and then pass through the atrioventricular (AV) node, a natural checkpoint that briefly delays the signal to give the atria time to finish contracting and fill the ventricles before they squeeze. The PR interval measures the total time from the start of the P wave to the start of the QRS complex, capturing both atrial depolarization and that AV-node delay together. A normal PR interval falls roughly between 120 and 200 milliseconds. When the AV node's conduction slows down further, whether from age-related fibrosis, certain medications, increased vagal tone, or other causes, the PR interval stretches out. A PR interval beyond 200 milliseconds is the standard threshold for first-degree AV block, the mildest category of AV conduction delay, in which every atrial signal still eventually reaches the ventricles but takes longer to get there. This simulation applies that same 200-millisecond threshold and visibly lengthens the gap between the P wave and QRS complex as you raise the PR-interval slider.

ST-Segment Elevation: The Signature Cardiologists Watch Most Urgently

After the QRS complex, the trace briefly settles onto the ST segment before rising into the broader, rounded T wave that represents the ventricles resetting (repolarizing) for the next beat. In a healthy heart, that ST segment sits essentially flat, at the same baseline level as the rest of the trace between beats. When a portion of heart muscle is acutely deprived of blood flow, most seriously during a full-thickness heart attack, the injured tissue's electrical behavior changes in a way that lifts the ST segment above the baseline on the ECG, producing the ST-segment elevation pattern that gives ST-elevation myocardial infarction, or STEMI, its name. Standard diagnostic criteria measure this elevation at the J-point, where the QRS complex ends and the ST segment begins, and generally require at least 1 millimeter (0.1 millivolt) of elevation in at least two anatomically contiguous limb leads, with a higher threshold of roughly 2 millimeters used in the precordial leads V2 and V3, where a slightly elevated baseline is more common even in healthy people. This simulation uses the 1-millimeter limb-lead figure as an illustrative threshold and lifts the ST segment of the drawn waveform correspondingly, while explicitly noting that a genuine STEMI reading also requires that the elevation appear consistently across multiple contiguous leads on a real 12-lead ECG, something one simplified single-lead-style trace cannot fully demonstrate.

From Individual Thresholds to a Combined Classification

Real ECG interpretation, whether performed by a cardiologist or an algorithm trained on data like PTB-XL, rarely comes down to a single isolated measurement. A trace can show a wide QRS complex and a long PR interval at the same time, or an elevated ST segment sitting inside an otherwise near-normal rhythm; the four features in this simulation are not mutually exclusive, and the classification readout is built to reflect that, flagging every real threshold that is crossed at once rather than forcing an artificial single answer. Heart rate itself is treated a little differently here: rates below 60 beats per minute (bradycardia) or above 100 beats per minute (tachycardia) are noted as rhythm context, since they are genuinely meaningful observations, but they are not folded into the four headline classification categories the way QRS duration, PR interval and ST elevation are, since a fast or slow rate alone does not, by itself, indicate a bundle branch block, AV block or STEMI pattern the way those three specific interval and segment abnormalities do.

Augmenting Cardiologists, Not Replacing Them

It is worth being direct about what datasets like PTB-XL have and have not enabled. They have made it realistic to build automated systems that flag likely-abnormal ECGs quickly, help triage which traces in a busy emergency department or telehealth queue deserve the fastest expert review, and occasionally surface subtle patterns a rushed or fatigued reader might miss. Research in this space is consistently framed around augmenting clinical workflows rather than replacing the clinician: automated ECG classification models are trained against, validated against, and generally deployed alongside continued cardiologist over-reading, and real-world clinical practice still requires a qualified professional to confirm any diagnostically significant finding before it drives patient care. That same spirit carries over to this simulation, on a much smaller and purely educational scale. It exists to make the logic behind ECG feature thresholds tangible and explorable, not to function as a diagnostic instrument. Nothing about crossing a threshold on these sliders, or in a real automated classifier's output, is meant to stand in for the judgment of a qualified clinician looking at an actual patient's full clinical picture.

Frequently asked questions

Try it live

Everything above runs in your browser — open ECG Waveform Classifier Lab and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab.