Deep neural networks (DNNs) have achieved remarkable results across various tasks, yet their decision-making processes remain opaque. In this paper, we introduced an Extended Garson Algorithm (EGA) framework designed to enhance interpretability by evaluating the importance of features in deep belief network-based autoencoders (DBNA). The EGA framework, applicable to any autoencoder with multiple hidden layers, is demonstrated through its effectiveness on classification and regression problems, comparing results with the Wald chi-square (χ2) and considering it a baseline. Our results show that adding more hidden layers usually improves the importance assigned to features, especially in the auto-insurance fraud detection (AIF) dataset with the Logistic Regression (LR) model and the credit card fraud detection (CCF) dataset with the Decision Tree (DT) model. However, some datasets, like credit card churn prediction (CCP), showed decreased feature importance when more hidden layers or features were added. Adding more features generally improved feature importance, but results varied by dataset. Increasing hidden layers typically improves model performance, as seen in reducing average Symmetric Mean Absolute Percentage Error (SMAPE) for regression datasets. Some changes were not statistically significant, indicating that further analysis is needed. Overall, the EGA framework offers a valuable way to understand better and interpret the complex decisions made by deep neural networks.