Advances in ECG wearable technology have revolutionized continuous cardiac monitoring. Nonetheless, the acquisition environment can damage all the clinical information of ECG signals. Thus, assessing ECG signal quality is crucial. Recent approaches have utilized convolutional neural network-based algorithms. Hence, in this approach, a transformation of ECG signal into images is required. In contrast to linear time-frequency transformations, phase space plots are an increasingly popular method for characterizing complex and chaotic dynamics in time series using nonlinear techniques. However, several key parameters such as the embedded dimension or the time delay ( \(\tau \) ) may vary depending on the specific application. This work aims to optimize the \(\tau \) value for assessing the quality of long-term ECG signals. A parametric analysis of the \(\tau \) parameter was developed using different training and testing sets. Eventually, significant disparities in overall performance were observed as the \(\tau \) parameter varied, showing differences of about 20% in the sensitivity and 30% in the specificity.