Accurate prediction of severity classes in Hemophilia B is critical for enhancing clinical decision-making and patient evaluation. Traditional methods primarily focus on analyzing the biological and structural characteristics of protein mutations. However, these approaches often encounter challenges due to the complex, high-dimensional nature of the data, leading to issues such as overfitting and the inability to capture key non-linear interactions, ultimately affecting prediction accuracy. This study introduces a robust machine learning framework aimed at addressing these limitations. The proposed framework incorporates advanced data preprocessing techniques, systematic feature selection, and an ensemble-based Bi-LSTM predictive model. By effectively managing heterogeneous data types and optimizing the handling of complex interactions, the framework improves prediction reliability and accuracy. Moreover, the model achieves better generalization, enhanced computational efficiency, and faster processing times, making it highly suitable for analyzing intricate biological datasets and providing actionable insights into disease severity classification.