Map composition framework for synthetic P morphology

In this work we address the data privacy concerns and the need of extensive data sets by the deep learning models. We account the need of low computational frame work along with flexible architecture and diverse requirements. We use a novel approach to generate extensive synthetic P wave morphology sets from fifteen cycles of the real ECG record (MIT-BIH Arrhythmia dataset). Our approach is based on development of Python framework of algorithms. It consists of a composition of a variability trend of discrete realistic ECG record samples. We used regression modalities (polynomial fit and spline fit) in a novel way in R2 vector space. The space thus formed is an interspersion of real and synthetic vectors. The space indexed with each real index in the range (0–14) produced a new synthetic P morphology. We introduced a robust approach by developing threshold invariance in R peak detection. We developed R–R interval extraction and subsequently P contour extraction algorithm as part of frame-work. A similarity metric is proposed in order to compare morphologies and address the realism feel of synthetic morphology. The similarity score of generated wave is 0.72. We have observed the scope for enhancement in P contour extraction algorithm and accounting P width variability. We noted that the method used for synthetic generation is linked to the end use case application for different trade-offs. Thus there is a scope for different methods and novelty. Our method produced extensive morphologies from a fifteen cycle snapshot of the real ECG record. We also developed a GAN P wave setup to estimate the compute complexity of GAN for comparison with map composition framework. The compute complexity was found to be best case O(25K) for the Keras (P wave GAN) sequential model architecture. For map composition frame work the compute complexity is O(67K), however the compute cost is significant in case of P wave GAN which is approximately (32∗106) operations, where as for map composition the cost is approximately (0.67∗106)operations. [Display omitted] •A novel generative method which views synthetic as vectors in real image vector space.•Polynomial, spline fit and direct map equations novel usage.•Synthetic ECG generation addresses data privacy and need of extensive dataset.•A simplistic, flexible and low compute framework for synthetic ECG generation.