Accurate disease diagnosis is essential for effective patient management, but the manual interpretation of complex biomedical data is both time-consuming and subject to error. Artificial intelligence, particularly machine learning models, can learn complex patterns from high-dimensional clinical and imaging data, but their predictive performance heavily relies on proper hyperparameter tuning. The study proposed a framework that combines nonlinear feature extraction, classification, and efficient optimisation. Kernel principal component analysis (KPCA) with a radial basis function kernel is employed to reduce dimensionality while preserving 95% of the variance. A multilayer perceptron (MLP) is then trained to predict disease. To enhance accuracy and computational efficiency, a modified multiprocessing genetic algorithm (MIGA) is introduced to optimise MLP hyperparameters. The framework was evaluated using three datasets: the Wisconsin Diagnostic Breast Cancer dataset, the Parkinson’s dataset, and the chronic kidney disease (CKD) dataset. Results demonstrated that the MLP tuned with MIGA achieved the highest accuracy of 99.12% for breast cancer, 94.87% for Parkinson’s disease, and 100% for CKD. These outcomes performed better than the traditional tuning methods, such as grid search, random search, and Bayesian optimisation. Kernel PCA help in uncovering nonlinear relationships that improve classification, while MIGA reduces the tuning time by approximately 60% compared to the standard genetic algorithm. Although traditional genetic algorithms experience high computational costs due to sequential evaluations, MIGA parallelises this step, significantly enhancing the process and guiding the MLP to optimal performance. The framework includes a graphical user interface that enables clinicians to load data, apply dimensionality reduction, tune hyperparameters, and perform predictions without writing code, providing a practical path toward real-world clinical adoption.