Feature selection is considered a complex problem to address in the statistical framework and becomes much more demanding as the number of variables increases. Thus, the methodologies for selecting and identifying important features in the random forest (RF) have been in constant evolution to better the interpretability of RF models. Among these, the minimal depth has been used as an indicator of importance "MINDepth". It considers that a variable is most important for predicting and interpreting when the minimum distance between the root node of the tree and the root of the subtree is shorter and lower than the minimal mean depth of all features. However, MINDepth does not get penalized when the feature is selected in a tree with "bad" performance. In this study, we introduce a "Deep Quality Score (DQS)" which is calculated using the error prediction.