How accurate is the Withings ECG?

Short answer: three peer-reviewed studies have measured the Withings electrocardiogram against a hospital reference. Two tested the ScanWatch directly against a 12-lead hospital ECG,  while the third validated the current detection algorithm across three different Withings devices. Across all three studies, atrial fibrillation was identified with a sensitivity of 96.3% to 99.63% and a specificity of 99.85% to 100%. In an independent study, the algorithm correctly classified 100% of readable ECGs, leaving far fewer inconclusive recordings than a competing six-lead device.

These figures are strong, but what they actually measure is another matter: without medical training, sensitivity and specificity are rarely self-explanatory. Before diving into the studies themselves, here is what those key terms actually mean.

The three terms behind every accuracy figure

Sensitivity: does it catch what is there?

Sensitivity measures how often a test correctly identifies a condition in people who actually have it.

A sensitivity of 96.3% means that out of 100 people with confirmed atrial fibrillation, the test flags around 96. The remaining 4 go unnoticed: these are false negatives.

High sensitivity matters a great deal here. Missing atrial fibrillation is a high-stakes error: the condition increases stroke risk roughly fivefold and frequently produces no symptoms at all.

Specificity: does it stay quiet when nothing is wrong?

Specificity is the mirror image: how often a test correctly returns a normal result in people whose heart rhythm is genuinely healthy.

A specificity of 100% means that, within the tested group, no one with a normal rhythm was told they might have atrial fibrillation—meaning zero false alarms were raised.

This is more crucial than most people realize. A screening tool is truly effective only when both metrics remain high, ensuring false positives are prevented as carefully as false negatives.

Classifiable recordings: was the signal clear enough?

This is the term almost nobody talks about, yet it is often the most revealing.

A wrist-based ECG, like the one in Withings watches, uses dry electrodes held against the skin. If your arm moves, your finger loses contact, or the signal picks up interference, the recording may become unreadable. When that happens, a well-designed algorithm does not guess: it returns an "inconclusive" result.

A classifiable recording has a signal clean enough for the algorithm to issue a definitive answer. Sensitivity and specificity are calculated solely on these recordings. This is the only honest methodology, as guessing on an unreadable trace yields unreliable data.

This leaves an important practical question: how often does the device manage to produce a usable reading? A highly accurate device that fails to classify one out of three attempts quickly becomes frustrating. This is precisely where the comparison in the second study becomes compelling.

Study 1: Does a watch ECG agree with a hospital ECG?

262 patients across four hospital sites in Paris, compared against a 12-lead ECG. Full study: Atrial Fibrillation Detection With an Analog Smartwatch: Prospective Clinical Study and Algorithm Validation, Campo et al., JMIR Formative Research, 2022

This initial validation evaluated cardiology patients who had two ECGs recorded simultaneously: one on a ScanWatch, one on a standard 12-lead hospital machine, which is the reference (gold standard) used in cardiology worldwide.

Each recording was reviewed by three board-certified cardiologists working independently, without access to the watch algorithm's verdict. This blinded review was critical, eliminating any risk of the reference diagnosis being biased by the device.

Result: For distinguishing atrial fibrillation from a normal rhythm, the algorithm reached 96.3% sensitivity and 100% specificity.

Two key details clarify this figure:

• Scope: These values reflect the task the device was designed for: detecting atrial fibrillation versus normal sinus rhythm. Patients with other arrhythmias, along with recordings the algorithm labelled as noise, were excluded from this calculation.

• Noise rate: 6.9% of recordings (roughly 1 in 14) were labeled as noise, which the algorithm declined to classify.

The cardiologists also assessed the raw quality of the single-lead trace, and this is arguably the most striking result. On the watch recordings, P waves were visible and correctly shaped in 96.9% of cases, and QRS complexes in 99.2%. The heart rate calculated by the algorithm differed from the cardiologists' own count by an average of 0.55 beats per minute. The researchers concluded that the trace is sufficient for physician use in routine practice, and not just as a consumer app.

Study 2: How does ScanWatch compare with another home ECG device?

176 adults with congenital heart disease, compared head-to-head at Amsterdam UMC. Full study: A comparison of ECG-based home monitoring devices in adults with CHD, Pengel et al., Cardiology in the Young, 2022

This study was conducted independently of Withings. A hospital team compared three home ECG devices against a 12-lead reference, in the same patients: the Withings ScanWatch (one lead), the Eko DUO (a precordial lead) and the AliveCor Kardia 6L (six leads).

The patients had been chosen among the most difficult cases. All 176 had a congenital heart defect; 24% a severe form, 84% had prior heart surgery, and 24% had right bundle branch block, a pre-existing conduction abnormality that distorts the trace and complicates automated reading.

Result: on the ECGs it classified, the Withings algorithm was 100% correct. The Kardia 6L algorithm scored 97%.

The critical gap highlighted by researchers, however, lay in the inconclusive rates. The Withings algorithm returned an inconclusive result on only 5% of recordings. One of the competing devices, on 31%, nearly a third of all attempts. The difference is statistically significant (p < 0.001), and vital in practice: a device that successfully provides an answer 95 times out of 100 is not the same experience as one that gives up every third attempt.

Conclusion: The authors noted that while all tested devices were accurate at detecting atrial fibrillation, the Withings algorithm generated the fewest uninterpretable results.

The study also highlights a more negative result: the ScanWatch is the weakest of the three at measuring the QTc interval, a precise timing measurement on the trace: acceptable in 51% of cases, against 70% and 74% for the two competing devices. It is worth saying plainly, because it marks the boundary of the figure. The ECG on a Withings watch is validated specifically for detecting atrial fibrillation, not for measuring QT intervals, and it is not marketed as such.

Study 3: The current algorithm on a large-scale dataset

4,646 recordings from 1,441 participants, plus a public reference database. Full study: Design and validation of Withings ECG Software 2, a tiny neural network based algorithm for detection of atrial fibrillation, Edouard & Campo, Computers in Biology and Medicine, 2025

While the first two studies tested the original algorithm, this paper validates its successor, Withings ECG Software 2, which powers the ScanWatch 2, ScanWatch Nova and ScanWatch Nova Brilliant watches.

The engineering constraint is worth understanding, because it explains the design. The algorithm is a neural network of only 3,633 parameters, minuscule by current standards, where model parameters are routinely counted in billions. It is deliberately compact to run directly on the watch, rather than sending heart data to a server and waiting for a verdict. It functions by locating the QRS complexes (the sharp spikes of each heartbeat) and passing that compressed pattern to the network rather than the raw trace.

It was tested on 4,646 thirty-second recordings from 1,441 participants, pooled from three clinical studies conducted on three different Withings devices.

Result: 99.63% sensitivity (95% CI: 99.15–99.84) and 99.85% specificity (95% CI: 99.61–99.94).

Two things give these figures weight. 

First, the narrow confidence intervals*—a direct benefit of a large sample size—mean the trial results closely reflect real-world performance. While testing an entire population is impossible, statistical precision bridges that gap. Second, the algorithm was benchmarked against the MIT-BIH Arrhythmia Database, a public, independent dataset uninfluenced by any manufacturer. Across 2,624 segments, it achieved 99.87% sensitivity and 100% specificity.

Note on transparency: This study was authored by Withings researchers using data from Withings-funded trials. This is standard practice for algorithm validation papers, as manufacturers collect and hold the primary data; however, it reinforces why independent validations (like Study 2) are essential.

Why you may see a slightly different figure. Product pages and the instructions for use quote 99.7% sensitivity and 99.8% specificity, measured in a clinical study of 626 subjects. The figures above come from a larger pooled dataset: 4,646 recordings from 1,441 participants across three clinical studies. Both are calculated on classifiable recordings, and both describe the ECG feature covered here.

*A confidence interval is a measure of how precise an estimate is. Saying that the measured mean is 50 with a 95% CI of [40;60] means you can be 95% confident that the true value (in the whole population) lies between 40 and 60. The narrower the CI, the more precisely that value can be located. This matters in medicine because not everything can be controlled in the life sciences: these intervals are needed to give an idea of how reliable a measurement is. It depends on the nature of the measurement, the experimental parameters (such as sample size) and the context.

The three studies at a glance

Three different research teams, three different populations, two generations of algorithms, and one consistent picture: when a Withings watch ECG reaches a verdict on atrial fibrillation, that verdict agrees in the vast majority of cases with a cardiologist reading a 12-lead ECG—and it yields a verdict more frequently than the competing device.

What these figures do not mean

Any discussion of accuracy must outline its operational boundaries:

An ECG trace is not a diagnosis. These studies measure the agreement between an algorithm and a cardiologist's reading. Only a qualified physician can confirm and diagnose atrial fibrillation and prescribe treatment. If you receive an atrial fibrillation detection, or if you feel unwell, the next step is a prompt appointment with a healthcare professional.

The automatic detection looks for atrial fibrillation, not the rest. The algorithm behind every figure above is validated for one task: distinguishing atrial fibrillation from a normal sinus rhythm. It is not designed to detect a heart attack, other arrhythmias or any other cardiac condition, and a normal result does not rule them out.

Extracting detailed clinical insight beyond AFib detection requires human expertise. Services like Cardio Check-Up (available via Withings+ in select regions) allow a cardiologist to review your recorded trace and return a report analyzing over 15 signs of arrhythmia beyond AFib within 24 hours. Because this relies on human review, it operates outside the scope of the algorithm accuracy metrics detailed above. Cardio Check-Up is included in Withings+, with four ECG reviews per year.

Single recordings are single snapshots. Atrial fibrillation is often paroxysmal (intermittent). A clean 30-second trace confirms rhythm only for that window, reflecting nothing about preceding or succeeding hours.

Clinical trial demographics differ from general populations. Study cohorts were recruited in clinical settings where AFib prevalence is naturally elevated. Accuracy measured in a cardiology ward does not transfer automatically, unchanged, to a healthy 30-year-old at home. This is often why, for most medical devices (and medicines), real-world results differ somewhat from clinical results.

Frequently asked questions

Is the Withings ECG medically validated?

Yes. The Withings ECG Monitor 2 feature complies with European standards and is medically validated as a medical device software. The CE 0123 certificate is affixed to this product. The official declaration of conformity is available at this link. The Withings ECG App is also cleared by the U.S. Food and Drug Administration (510(k) K240795) to detect atrial fibrillation from a single-lead ECG. Regulatory status and available features vary by country.

What does "single-lead ECG" mean?

A hospital ECG uses ten electrodes to build twelve views, called leads, of the heart's electrical activity. A smartwatch produces just one, equivalent to lead I of a standard ECG. One lead is enough to observe the rhythm (whether the beat is regular and whether P waves are present), which is what atrial fibrillation detection relies on. It is not enough for the complete cardiac assessment a 12-lead ECG allows. Withings does however offer a device with a 6-lead ECG: the Body Scan and BodyScan 2 connected scales.

Why does my watch say "inconclusive"?

An inconclusive reading means signal interference prevented a confident classification, prompting the algorithm to withhold a judgment rather than risk an error. To ensure a clear trace: sit quietly, rest your arm on a table, maintain steady skin contact with the metal bezel, and avoid moving or speaking for 30 seconds.

Which is more accurate: one lead or six?

Technically, more leads always bring more information. But they also bring more technical complexity and greater algorithmic demands. A balance has to be struck between that complexity and everyday reality, so that the feature is genuinely useful in preventing cardiovascular disease. For example, to detect AFib — whose episodes can vary in frequency and intensity — the priority becomes optimising battery life so that passive checks can run more regularly, increasing the chance of catching those episodes.

Where to find more information

United States: American Heart Association (AHA), American Medical Association (AMA), U.S. Food and Drug Administration (FDA)

Europe: European Society of Cardiology (ESC), European Medicines Agency (EMA)

Rest of the world: World Health Organization (WHO)

Always refer to your own country's guidance for medical sources and advice.

Sources

1. Campo D, Elie V, de Gallard T, Bartet P, Morichau-Beauchant T, Genain N, Fayol A, Fouassier D, Pasteur-Rousseau A, Puymirat E, Nahum J. Atrial Fibrillation Detection With an Analog Smartwatch: Prospective Clinical Study and Algorithm Validation. JMIR Formative Research 2022;6(11):e37280. doi:10.2196/37280 — ClinicalTrials.gov NCT04351386

2. Pengel LKD, Robbers-Visser D, Groenink M, Winter MM, Schuuring MJ, Bouma BJ, Bokma JP. A comparison of ECG-based home monitoring devices in adults with CHD. Cardiology in the Young 2023;33(7):1129–1135. doi:10.1017/S1047951122002244

3. Edouard P, Campo D. Design and validation of Withings ECG Software 2, a tiny neural network based algorithm for detection of atrial fibrillation. Computers in Biology and Medicine 2025;185:109407. doi:10.1016/j.compbiomed.2024.109407

The ECG feature analysed in these studies is available on ScanWatch, ScanWatch 2, ScanWatch Nova and ScanWatch Nova Brilliant. Availability and regulatory status of health features vary by country. Other Withings products also offer the ECG feature. Browse our website to find out more.