What if AI could match cardiologist-level accuracy in detecting heart disease? Our latest study, published in Heart Rhythm Journal, brings that vision closer to reality.
Willem1, Idoven’s AI-powered ECG analysis platform, has achieved 96.4% accuracy in detecting atrial fibrillation (AF) from single-lead ECGs, even when tested on entirely new data. In a large-scale validation with 8,528 patients, Willem outperformed traditional rule-based algorithms, proving its potential for real-world clinical use.
This breakthrough is shaping the future of cardiac care—delivering more accurate, scalable, and accessible AF screening.
So, how far can AI take us in the fight against heart disease? Let 's dive in.
Key findings: How well did Willem perform?
When it comes to detecting atrial fibrillation (AF), Willem has demonstrated remarkable accuracy, even in the challenging context of an imbalanced dataset where only 8.9% of recordings corresponded to AF cases.
The results speak for themselves. Willem achieved an overall accuracy of 96.4%, with a sensitivity of 84.2% and a specificity of 97.6%. The algorithm identified 78% of AF cases after expert label review, reaching a F1-score of 80.9%.
These performance metrics surpass those of traditional rule-based algorithms and many commercial devices, proving the robustness of this AI-driven platform.

How was the study conducted?
To assess Willem’s performance, we used the publicly available 2017 PhysioNet/Computing in Cardiology Challenge dataset. This dataset included 8,528 single-lead ECG recordings collected from KardiaMobile devices by AliveCor, categorized as follows: normal (59.5%), AF (8.9%), other rhythms (28.3%), and noisy recordings (3.3%).
To ensure an unbiased evaluation, Willem was tested on completely new data, meaning it had never encountered these ECGs before. Each recording was labeled using a reference algorithm and validated by cardiology specialists, ensuring the highest quality annotations.

Willem: Certified AI software for cardiac care
Willem is a CE-marked medical device software, fully compliant with EU MDR regulations, ensuring safe and reliable deployment in clinical settings. Its cloud-based design enables seamless integration into hospitals, clinics, and telemedicine platforms, supporting AF detection and cardiac monitoring at scale.
Why is atrial fibrillation so hard to detect?
Despite affecting more than 60 million people worldwide, atrial fibrillation often goes unnoticed. It can be intermittent, silent, and difficult to diagnose, leading to serious complications like strokes and heart failure if left untreated.
Portable ECG devices have improved heart monitoring accessibility, putting powerful tools in the hands of millions. However, automated interpretations often lack accuracy, making AI-driven solutions like Willem essential in bridging the gap.
Beyond AF: Expanding the scope of arrhythmia detection
While this study focused on AF detection, Willem has also demonstrated its ability to identify other arrhythmias. The platform accurately detects premature atrial contractions (PACs), premature ventricular contractions (PVCs), and first-degree AV blocks, further proving its potential for comprehensive cardiovascular disease screening and management.
Looking ahead
The study reinforces AI’s transformative potential in cardiology. By combining advanced technology with rigorous validation, Idoven is empowering healthcare professionals with the tools to enhance diagnostics, improve patient outcomes, and combat the growing burden of heart disease.
To explore the full details of the study, read the published paper in the Heart Rhythm Journal here.
[1] This study was conducted in 2024 using the Willem 1.1 version.
The Willem™ ECG Analysis Platform interpretation results are not intended to be the sole means of diagnosis. It is offered to physicians and clinicians on an advisory basis only in conjunction with the physician’s knowledge of ECG patterns, patient background, clinical history, symptoms, and other diagnostic information. Please see Instructions of use for extended information about the product.