On the Performance of Composite 1D-to-2D Projections for Signal Quality Assessment

Abstract

Signal quality assessment is essential for health monitoring applications, as good signal quality is needed to reliably inform about the medical conditions of the patient. To achieve this, machine learning algorithms such as convolutional neural networks may be applied. However, the signal needs to be transformed into a 2D representation, which can be done using time series imaging techniques such as Gramian Angular Field (GAF), Markov Transition Field (MTF), and Recurrence Plot (RP), as well as by aggregating their results, which we refer to as Projection Mix. After preprocessing the dataset, Brno University of Technology Smartphone PPG (BUTPPG), into these images, various convolutional neural networks were trained and tested using such data, while also selecting hyperparameters through heuristic searching. The results indicate that our proposal performed better than the state-of-the-art methods.