Transforming Normal ECG to Myocardial Infarction Ones using Invertible Conditional GANs

Recent advances in deep learning have enabled effective style transfer methods for biosignal synthesis, partic- ularly for generating pathological variations of physiological signals. This work investigates the application of Invertible Conditional Generative Adversarial Networks (IcGANs) to modify 12-lead ECG heartbeats from normal sinus rhythms into patterns typical for myocardial infarction (inferior and antero- septal). In contrast to CycleGAN, which requires multiple models for each direction of transformation, IcGANs only require the training of a single conditional GAN along with an encoder, offering a more direct and lightweight framework. We trained both IcGAN and CycleGAN models using heartbeats from the PTB-XL dataset. The quality of the generated ECGs was assessed using both qualitative and quantitative techniques, including visual inspection, GAN-train and GAN-test scores, and comparisons of ST-segment amplitudes. Our results indi- cate that IcGAN can realistically and meaningfully alter ECG signals to exhibit myocardial infarction traits while retaining their core physiological structure. Comparisons showed IcGAN to be more efficient and effective than CycleGAN under similar architectural conditions. This approach shows promise for generating rare pathological cases, adapting models across domains, and supporting generalization in clinical applications for more personalized diagnostics.