Malaria remains a critical public health burden in sub-Saharan Africa, contributing significantly to illness and mortality, particularly among children and pregnant women. Rapid and accurate diagnosis is essential for timely treatment and effective disease management. Convolutional Neural Networks (CNNs) have shown substantial promise in automating malaria detection from microscopic blood smear images, but their performance heavily depends on optimal hyperparameter tuning, a task that is computationally intensive and highly sensitive to initial conditions. To address this challenge, this study proposes an enhanced Adaptive Marine Predators Algorithm (AMPA) for efficient hyperparameter optimization. The proposed method introduces a dynamic step-size adjustment strategy, which adaptively modifies the search behavior in response to real-time validation loss trends during training. This mechanism improves convergence stability and helps the optimizer focus on promising regions of the search space. Furthermore, a multi-objective fitness function is employed to jointly optimize classification accuracy, generalization capability, and computational efficiency. The effectiveness of the proposed approach is demonstrated using the publicly available Kaggle Malaria Cell Images Dataset, which consists of over 27,000 annotated images of parasitized and uninfected red blood cells. Empirical results show that the adaptive MPA consistently outperforms conventional optimization strategies, yielding CNN configurations with superior detection accuracy and faster convergence. These findings highlight the potential of intelligent, nature-inspired optimization algorithms in improving the deployment of deep learning-based diagnostic systems in real-world, resource-constrained healthcare settings, and contribute to the broader goal of enhancing malaria control through automated, scalable diagnostic tools.