Deep Survival Machines (DSMs) integrate neural networks with parametric survival distributions, yet their architecture tightly couples representation learning with parameter estimation, limiting interpretability and obscuring the role of latent risk structure in individualized predictions. This tight integration makes it difficult to disentangle how covariate information is transformed into distributional parameters, thereby reducing transparency in model behavior and complicating theoretical understanding of how latent features influence survival dynamics under censoring. This study proposes a Modified Deep Survival Model that explicitly decouples nonlinear feature extraction from probabilistic survival modeling. A multilayer perceptron transforms high-dimensional covariates into compact survival-relevant representations, which are then passed to a New Exponential Power Distribution (NEPD) for parameter estimation and survival function generation. The NEPD accommodates flexible tail behavior, enabling more accurate modeling of heterogeneous time-to-event patterns than conventional parametric forms. The model is trained using censored-likelihood optimization with prior regularization to ensure stable parameter learning. On the SUPPORT benchmark dataset, the Modified DSM achieves a Concordance Index of 88.34%, outperforming the conventional DSM by 1.34%. This improvement demonstrates that separating latent structure learning from distributional estimation enhances both discriminative accuracy and structural transparency. The proposed study offers a robust alternative for clinical risk prediction requiring interpretable, individualized survival distributions