Electrocardiogram (ECG) signals are indispensable for diagnosing cardiovascular diseases (CVDs), yet conventional 12-lead ECG acquisition is limited in out-of-hospital environments. Although wearable devices enable real-time monitoring with single-lea...
Electrocardiogram (ECG) signals are indispensable for diagnosing cardiovascular diseases (CVDs), yet conventional 12-lead ECG acquisition is limited in out-of-hospital environments. Although wearable devices enable real-time monitoring with single-lead ECGs, their diagnostic capacity remains restricted due to spatial incompleteness. To address this limitation, we propose a novel multi-stage deep learning framework that enhances single-lead ECG utility through signal generation, image-based transformation, and early event prediction.
First, we introduce a GAN-based model that, for the first time, synthesizes full 12-lead ECGs from only lead I which is, the signal commonly acquired from smartwatches. Unlike previous augmentation-focused GANs, our model directly targets diagnostic generation and achieves superior classification performance compared to real ECGs, validated across external datasets. This resembles that the generated ECG’s capability not only in practical use but also could help diagnose CVDs.
Second, we present one of the earliest deep learning frameworks capable of predicting acute myocardial infarction (AMI) up to six months before its clinical onset using large-scale, pre-event ECG records. Leveraging more than 6,600 AMI patients and 55,000 controls, our model analyzes temporally stratified ECG windows ranging from 1–3 days to 3–6 months pre-onset to identify subtle electrical and morphological signatures that precede infarction. Unlike prior studies focused exclusively on detecting MI at or near onset, our approach demonstrates that preclinical ECG alterations contain meaningful predictive information, enabling early risk stratification of asymptomatic individuals. Performance analyses show that predictive accuracy increases as the ECG approaches the AMI event, yet the model remains capable of distinguishing long-term (up to 180 days) risk patterns. Prospective study using newly diagnosed patients further confirms the generalizability of the framework, underscoring its potential utility as a proactive screening tool in both clinical and ambulatory monitoring environments.
Collectively, this dissertation pioneers a unified strategy to overcome the limitations of single-lead ECGs, integrating generative modeling and longitudinal prediction. These contributions represent a significant step toward practical, AI-enabled, real-time cardiac monitoring in ambulatory and remote settings.