Masked Modelling is all you need for Electrocardiogram Signals

Gouthamaan Manimaran1, Sadasivan Puthusserypady1, Maria Helena Dominguez2, Jakob Eyvind Bardram1
1Technical University of Denmark, 2Frederiksberg Hospital


Abstract

Electrocardiogram (ECG) signal analysis is crucial for diagnosing cardiovascular diseases, yet traditional approaches often struggle with noise, missing leads, and limited training data. In this work, we present a novel supervised learning approach for ECG classification that leverages masked modeling techniques. Our method randomly masks 2-6 leads during training, forcing the model to learn robust representations from incomplete data. We implement this approach using a modified EfficientNetV2 architecture adapted for 1D signal processing, incorporating Squeeze-and-Excitation modules and both Fused and standard Mobile Bottleneck Convolution blocks to efficiently capture temporal dependencies in ECG signals. Our team ‘DTU_HealthTech' finished 7th in the unofficial phase of the PhysioNet Challenge 2025, having a Challenge score of 0.789. Experimental results in internal testing show that our masked modeling approach not only improves F1 scores by 8.2% compared to baseline methods but also maintains robust performance when tested on real-world clinical data with varying signal quality. This work demonstrates that supervised masked modeling provides an effective framework for ECG analysis, potentially reducing the need for perfect multi-lead recordings in clinical settings while maintaining diagnostic accuracy.