<p>Artificial Neural Networks (ANNs), particularly Multilayer Perceptrons (MLPs), have been widely employed in medical data classification because of their ability to model the complex and nonlinear relationships inherent in medical datasets. However, their performance relies heavily on the optimal configuration of both the training parameters (weights/biases) and network architecture. Traditional gradient-based methods, such as backpropagation, often encounter challenges, including slow convergence and susceptibility to local minima. Meta-Heuristic Algorithms (MHAs) have emerged as powerful optimizers that can mitigate these limitations; however, most existing MHA-based approaches optimize the training parameters and architecture components in isolation, ignoring their critical interdependence. To bridge this gap, this study proposes XCOA-MRFO, a novel hybrid framework for the simultaneous optimization of the MLP architecture and training parameters. By employing a dual-level strategy, Manta Ray Foraging Optimization (MRFO) dynamically explores discrete architectural space, whereas the Extended Coyote Optimization Algorithm (XCOA) efficiently exploits the continuous training parameter space. The framework relies on a specialized adaptive composite fitness function designed to evaluate candidate solutions based on the predictive precision and structural efficiency. Evaluated on five benchmark University of California, Irvine (UCI) medical datasets (Breast Cancer, Diabetes, Liver, Parkinsons, and Vertebral), XCOA-MRFO demonstrated superior performance, achieving mean accuracies of 98.4%, 81.0%, 79.8%, 91.9%, and 88.6%, respectively. A comparative analysis of eight MHAs and traditional classifiers confirmed that the proposed simultaneous optimization approach significantly enhances the convergence speed, solution robustness, and classification accuracy in complex medical diagnostics, offering a robust, efficient, and highly dependable solution for medical data classification tasks.</p>