xgboost - Feature selection using xgboost and SHAP algorithm, does the base model score matters?

Question

so Im currently doing feature importance using xgboost and SHAP algorithm. The approach that Im currently doing right now is

1. Split the data into training and testing
2. Predict the model using testing dataset and evaluate
3. Using SHAP to conduct feature selection

my question is, does the prediction score effect the result of the feature importance? also does better prediction score means more accurate in the feaute importance? Thank you