what would a good split train/test be, having a class imbalance problem in the dataset and a small number of observations (<5000 obs)? Would it make sense to consider k-fold cross validation (e.g. 5)?
If training is not too expensive, cross validation with as much folds as possible (even leave-one-out) is better. Because, let's say you have only 1% of the minority class, i.e. ~50 samples in your case, a single train/test split (like 80/20) might mean only 10 minority samples, which could increase the variance of your evaluation metrics.