A Transformer-Based Framework with Dynamic Multiscale Attention Mechanisms for 12-Lead ECG Synthesis

Cardiovascular disease (CVD) remains the leading cause of mortality worldwide, emphasizing the need for enhanced diagnostic methods. Standard 12-lead Electrocardiograms (ECGs) are the gold standard for CVD diagnosis but require specialized equipment, limiting diagnostic accessibility. Wearable smart devices enable at-home heart monitoring but are only able to record a limited lead set, restricting their diagnostic scope. This paper proposes a methodology that uses only three leads as input-I, III, and V6-to generate a complete 12-lead ECG. The key innovation proposed in this paper is the Dynamic Multi-Scale Time Series Transformer (DMTT), which synthesizes leads V1-V5 using only the three input leads. ECG synthesis is challenging due to the presence of both sharp and subtle waveform variations. DMTT integrates Dynamic Window and Multi-Scale architectures, enabling precise synthesis of ECG morphology and timing. DMTT consists of two primary components: the Dynamic Window Module (DWM) and Multi-Scale Module (MSM). The DWM implements a saliency-based attention mechanism, which is a small sub-network that infers local relevance for each time step, generating a saliency map that enables the model to adaptively pool features based on their measured importance. The MSM pools the sequence at multiple temporal scales and integrates these representations, allowing the model to capture spatiotemporal dependencies more effectively. This mechanism facilitates accurate synthesis of ECGs. The model was trained on the PTB-XL dataset and was evaluated using two accuracy metrics: Normalized Root Mean Square Error (NRMSE) and Average Absolute Error (AAE). Results demonstrate that the synthesized ECGs achieve an average AAE of 0.0777mV and NRMSE of 0.0907 across leads V1-V5 and five disease types. Qualitative analysis confirms the preservation of diagnostic features, illustrating the feasibility of synthesizing a 12-lead ECG from a subset of leads using a generative AI model.Clinical Relevance- The novel technique of synthesizing a full 12-lead ECG from a limited lead set allows wearable devices to capture a subset of leads, which serve as inputs to DMTT for generating a complete 12-lead ECG. This approach facilitates at- home heart health monitoring for a broad range of heart diseases using wearable technology. Additionally, it enhances the applicability of automated AI classification methods, enabling the detection of cardiac conditions that manifest in any of the 12 ECG leads.