The early, and accurate diagnosis of the gastrointestinal diseases is crucial for the effective clinical management as well as the treatment planning. The endoscopic imaging plays the central role in the diagnosis; however, the manual interpretation remains time consuming, and subject to the inter observer variability. To address these challenges, the study proposes the hybrid attention guided deep learning framework for the robust as well as the interpretable gastrointestinal disease classification using the endoscopic images. The framework integrates the dual convolutional neural network backbones, namely the EfficientNetB3, and ResNet50 architectures, to extract the complementary texture as well as the structural features. The channel wise, and the spatial attention mechanisms are incorporated through the squeeze, and excitation blocks together with the convolutional block attention modules to enhance the discriminative feature learning. In addition, the genetic algorithm is employed for the feature selection in order to reduce the redundancy, and improve the generalization capability. The model is evaluated on the KVASIR dataset comprising 8,000 images distributed across the eight gastrointestinal disease classes. The experimental results demonstrate the average training accuracy of 99.2 %, validation accuracy of 96.7 %, as well as the test accuracy of 96.1 %. The Gradient weighted Class Activation Mapping (Grad-CAM) technique is utilized to provide the visual interpretability by highlighting the clinically relevant regions influencing the model predictions. The results indicate that the proposed framework achieves the competitive classification performance while maintaining the interpretability, thereby making the frame-work suitable for the computer aided gastrointestinal disease screening applications.