Together with Zenicor Medical Systems, Modulai trained a deep learning model to predict underlying paroxysmal atrial fibrillation (AF) from normal-looking ECG recordings. The project, partly funded by Vinnova, addresses a critical gap in stroke prevention by detecting a condition that is notoriously difficult to diagnose from standard ECG readings.
Outcome
+7%
Detection performance over baseline
Challenge
Zenicor has become one of the leading Medtech companies in Europe in early diagnosis of arrhythmias and stroke prevention. Every day, thousands of people suffer a stroke due to an undiagnosed AF condition, often resulting in paralysis or death. Paroxysmal AF occurs when a rapid, erratic heart rate begins suddenly and stops on its own within seven days. Detecting AF while it is happening is straightforward from the ECG pattern, but detecting an underlying AF condition from a normal-looking recording is a much harder challenge.
Solution
Working with the client and leading cardiologists at Karolinska Institutet as domain experts, the Modulai team moved beyond standard approaches. Rather than rely on the limited labeled data alone, they made use of the large amount of unlabeled ECG data through representation learning, developing a novel method for the problem, and then applied transfer learning to adapt it to AF detection specifically. Explainability methods were used to understand how the model arrived at its predictions, which matters when the output informs a clinical decision.
Tools
Data were processed using various open-source Python libraries. Model architectures were developed in TensorFlow, and the model was wrapped in a Flask API and dockerized for deployment.
Value created
By learning from unlabeled data instead of the scarce labeled examples alone, the approach improved performance by 7% over the baseline, the difference between a model with limited practical value and one that can detect underlying AF in recordings that look normal to standard analysis. That points directly at earlier diagnosis and fewer missed cases of a condition that is a leading, preventable cause of stroke.
Learn more
