<p>Voice pathology detection is important in the early diagnosis as well as treatment of vocal disorders; if left untreated, it can severely impact the ability of a person to communicate. Traditional diagnostic methods can take a lot of time, be subjective and reliant on specialized clinical equipment. In order to overcome these problems, this study introduces a dual-stage multimodal framework for automatic voice pathology detection utilizing the Saarbruecken Vocal Database (SVD). The deep spectral features are extracted from voice samples utilizing a Multi-Stream ResNet-152 combined with a Sparse Transformer (MSRSpT-152). The extracted features from multiple sources are fused using a Multi-Objective Water Strider Algorithm (MObjWSA) method, which confirms the retention of the most informative characteristics while maintaining balance across evaluation metrics. In the first stage, the Convolutional Neural Network alongside a Hierarchical Self-Attention (Conv-HSA) is utilized to refine the extracted spectral representations and improve discriminative feature learning. During the second stage, classification is carried out utilizing the Multi-Objective Gradient-Boosted Ensemble (MObjGB-Ensemble). This ensemble integrates XGBoost, LightGBM and Random Forest (RF) models, all trained on the fused features. The ensemble approach addresses data imbalance by optimizing multiple performance metrics simultaneously. Experimental outcomes show that the proposed model obtains an accuracy of 99.02% along with a precision of 98.74%, a sensitivity of 99.08%, a specificity of 98.94% and an F1-score of 98.91%.</p>