I assume you have read this paper, published in 2021: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7831114/
The authors use only lead II for their classification algorithm. Lead II is typically used for a monitor display, when only a single lead is shown. I suggest you start with lead II only.
You ask "is each lead's scalogram regarded as a class instance, or is it pivotal to concatenate the 12 leads to accurately represent an ECG". I have not worked with classificaiton algorithms. Therefore please excuse me if I say things that are obvious to you, but not obvious to me. I may also say things that you already know. I am not trained in reading ECGs.
The different leads carry different information. Certain pathologies are more apparent in certain leads. Some pathologies display unique combinations of features in particular leads with a particular temporal relationship. For example, in torsades des pointes, there is a distinctive temporal progression of oscillations from one lead to another lead. Another example is ST segment elevation myocardial infarction (STEMI). The infarct may be classified according to which leads show the greatest ST segment elevation, and the diagnosis of STEMI is made with greater confidence when the electrically oppostite leads show reciprocal ST depression (see here). The feature extraction and learning algorithm should account for these sorts of phenomena.
The different leads may be thought of as projections of the instantaneous dipole moment of the heart through the body tissues onto the surface of the body. This is an overfsimplification, since the charge distribution in the heart at each instant is more complicated than a simple dipole. The electrical properties of the tissues, which affect how the dipole moment prohjects to the body surface, are complicated and change with time, due to breathing, pulse-related blood volume changes, muscle contraction, etc. I mention these facts just to make the point that the different leads of the ECG are distinct.
