MFDN: Multi-scale Frequency-domain Enhanced Dual Network for electrocardiographic detection of cardiac hypertrophy

Cardiac hypertrophy (HYP) is associated with an increased risk of heart failure and sudden death. Although the electrocardiogram (ECG) is widely available, conventional diagnostic criteria suffer from low sensitivity. Existing deep learning methods often fail to effectively capture frequency-domain abnormalities and inter-lead dependencies in hypertrophic ECG signals. To address these limitations, we propose the Multi-scale Frequency-domain Enhanced Dual Network (MFDN) for automated HYP-label diagnosis, which fully leverages both the time- and frequency-domain information in 12-lead ECGs. Specifically, the backbone network integrates a Multi-scale Frequency Enhancement (MSFE) module utilizing a trainable soft-adaptive threshold mechanism to adaptively decouple overlapping frequency components and extract precise pathological spectral representations. Furthermore, a lightweight topology-aware branch network and Prior Lead Fusion Network (PLFN) is designed to preserve the spatial topology of the 12-lead ECG. These dual-network representations are dynamically aligned and integrated through a PLFN via a channel-wise gating mechanism to guide the backbone network for precise hypertrophy detection. The model is evaluated on the PTB-XL database that comprises 2,649 subjects are annotated with HYP, using an 8:1:1 train–validation–test split. It achieves highly competitive performance compared to recent benchmarks at a low computational cost, achieving a recall of 0.7175, an AUC of 0.9070, an Accuracy of 0.8906, and an F1-score of 0.6506. This robust performance highlights the value of effectively capturing frequency-domain features for automated ECG annotation. Finally, feature importance analysis using Shapley values confirms that MFDN relies on clinically relevant pathological features.