component analysis, Parallel processing.

Disease diagnosis is essential for effective patient management and improved clinical outcomes. Clinicians traditionally rely on manual interpretation of laboratory tests and diagnostic imaging to detect diseases. However, this approach is time-consuming and often incapable of capturing the full complexity of high-dimensional biomedical data [1]. Artificial Intelligence (AI) is a branch of computer science that focuses on creating systems capable of performing tasks that typically require human intelligence. Machine Learning (ML) trains algorithms to learn from data and improve their performance without explicit programming [2]. By applying optimization techniques to tune model parameters, AI-driven frameworks automatically extract intricate features from clinical and imaging data, thereby enhancing the predictive accuracy and consistency for diseases [3]. However, the field of AI continues to evolve from rigid rule-based automation to systems that imitate human reasoning, perception, and foresight, allowing machines to take autonomous actions in complex environments. These advancements demonstrate that AI is not merely a tool for efficiency but also a transformative enabler that augments human capabilities across daily life and industry. Hyperparameter tuning is essential in ML because it directly affects model accuracy, generalization, and stability by determining how the learning algorithm behaves during training [4]. Proper tuning helps handle underfitting and overfitting by optimizing parameters or model complexity [5]. Without tuning, models often fail to reach their full predictive potential.

Traditional methods, such as grid search and random search, exhaustively evaluate all possible combinations, making them computationally expensive and inefficient in high-dimensional spaces —a problem known as the curse of dimensionality [6]. These limitations impose more adaptive methods, such as Genetic Algorithms (GAs), that can intelligently search for better hyperparameter combinations. The GA is increasingly used for hyperparameter tuning in ML due to its ability to explore large, complex, and nonlinear search spaces more efficiently than traditional methods, such as grid or random search. GAs mimic the process of natural selection by encoding hyperparameters as chromosomes, evaluating their performance (Fitness), and iteratively improving them through genetic operators such as selection, crossover, and mutation [7]. This evolutionary process enables the GA to avoid local minima and converge toward globally optimal solutions, making it particularly useful for high-dimensional and nonconvex tuning problems. Unlike traditional techniques, which are static and uninformed, the GA is adaptive and uses past evaluations to guide future searches, reducing redundant trials and improving convergence speed. While the GA is effective for exploring large search spaces, it often suffers from slow convergence and high computational cost, primarily when complex models are evaluated over many generations [7]. Additionally, its performance is sensitive to the tuning of GA-specific parameters such as population size, crossover, and mutation rates, which can lead to premature convergence or suboptimal results if not well-balanced [7]. This work aims to propose a predictive model for disease prediction via an optimized ML model with an improved GA with parallel fitness evaluation called the Multiprocessing Interface Genetic Algorithm (MIGA) to increase disease prediction and tuning efficiency.

Recent works have shown the effectiveness of the GA in optimizing ML models by addressing the limitations of traditional hyperparameter tuning methods. Rodrigues et al. [8] proposed a hybrid model combining a ResNet-50v2 CNN with a GA for classifying Acute Lymphoblastic Leukemia (ALL) from microscopy images. The GA was used to fine-tune the Convolutional Neural Network (CNN) hyperparameters, resulting in a classification accuracy of 98.46%, which outperformed both the random search and Bayesian optimization approaches. Kaur et al. [9] developed a GA-aided hyperparameter optimization framework combined with an ensemble learning model to predict respiratory diseases via clinical data. Their methodology also incorporated SHAP-based explainable AI to interpret model predictions. The GA-optimized AdaBoost classifier achieved the highest accuracy among all the models evaluated, showing the efficacy of GAs in tuning complex classifiers and reducing human involvement in parameter selection. In the energy sector, Liu et al. [10] applied a GA to enhance the performance of a Random Forest (RF) model for predicting grid faults. By selecting relevant meteorological features and tuning the model hyperparameters with a GA, the proposed method improved the prediction accuracy by 14.77% over that of standard RF models, illustrating the adaptability of GAs in real-time fault prediction under environmental variability.

Furthermore, Guido et al. [11] introduced a multiobjective GA approach for tuning Support Vector Machine (SVM) hyperparameters in imbalanced datasets. Their model, which incorporates decision trees to accelerate GA evaluations, significantly reduces the computational time while improving balanced accuracy metrics such as the G-mean, which are vital for healthcare-related classification tasks. Building on constrained optimization techniques, El-Hassani et al. [12] proposed a novel framework named MLPRGA+5, aimed at configuring Multilayer Perceptron (MLP) networks through real-coded GAs. The work implemented advanced genetic operators such as tournament selection with elitism, Simulated Binary Crossover (SBX), and Polynomial Mutation (PM). The model’s efficiency was validated on four UCI datasets, where MLPRGA+5 achieved higher accuracy and reduced network complexity compared to traditional methods, confirming the robustness of the evolutionary design in practical neural network tuning. El-Hassani et al. [12] introduced an ensemble model (EGACNN) that stacks a CNN with a GA to optimize hyperparameters such as the dropout rate, batch size, and learning rate. Their method achieved 99.91% accuracy on MNIST, outperforming the classical CNN and other ensemble models, thereby validating the robustness of GAs in fine-tuning deep neural networks for image classification tasks. Ranga et al. [13] optimized an MLP using a GA to predict Chronic Kidney Disease (CKD) and achieved better accuracy scores of 98.34% and 98.54% for the training and testing processes, respectively.

The study was able to propose a modified GA for optimizing the MLP. The optimized model was used to predict diseases, and the framework was transformed into an app that incorporates the proposed hyperparameter tuning technique to achieve enhanced performance. The study evaluated the performance of the proposed model on three different datasets: breast cancer, CKD, and Parkinson's. Table 1 presents the hyperparameters optimized via the GA during the training of the MLPC. The search space for the hidden layer sizes included various configurations ranging from (50), (100), and (150) to two-layer architectures such as (50, 50) and (100, 100). The activation functions considered in the search process were ReLU, tanh, and logistic, allowing the model to explore different nonlinear transformation capabilities. The initial learning rates were tuned across three values: 0.001, 0.01, and 0.1, enabling the GA to adjust the convergence behaviour of the neural network. Finally, the solvers ‘adam’ and ‘SGD’ were explored to determine the most effective optimization technique for minimizing loss during training.

Table 2 outlines the configuration settings employed for the proposed GA in tuning the hyperparameters of the MLP. The population size was set to 10 individuals, and the algorithm evolved over 10 generations. A mutation rate of 10% was applied, where one hyperparameter in a chromosome was randomly altered. The selection strategy used was elitist selection, retaining the top 50% of individuals based on fitness scores. A uniform crossover approach was adopted to recombine the parent chromosomes. The fitness function was defined as the accuracy score on the test dataset, and evaluations were executed in parallel via ThreadPoolExecutor to accelerate computation. Each chromosome encodes a complete model configuration represented by the tuple [Hidden layer size, activation function, learning rate, solver]. The proposed technique was explicitly applied to optimize an MLP with a maximum iteration limit of 500.

Table 3 presents the performance metrics of the MIGA by generation when applied to the Parkinson dataset. Across 10 generations, the algorithm consistently demonstrated strong optimization capability, with the best accuracy reaching 0.9487 from generation 2 onwards. The minimum and maximum accuracy values for each generation reflect the diversity in model performance within the population, with early fluctuations gradually stabilizing in later generations. The results indicate that the algorithm effectively converges toward high-performing hyperparameter combinations, maintaining an accuracy of 0.9487 for the majority of generations. Table 4 summarizes the performance of the MIGA across 10 generations on the Breast Cancer dataset. The results show strong classification performance, with the best accuracy, reaching a peak of 0.9912 as early as generation 2 and remaining consistently high throughout subsequent generations. The minimum and maximum accuracies recorded per generation demonstrate moderate variation in early generations, followed by stabilization as the algorithm converges. The sustained best accuracy across multiple generations indicates the robustness and efficiency of the MIGA in evolving optimal hyperparameter configurations for MLP-based classification on the breast cancer dataset.

Table 5 presents the accuracy performance of the MIGA over 10 generations when it is applied to the CKD dataset. The algorithm consistently achieves exceptional accuracy, with the best performance reaching 1.0000 (100%) from the first generation and maintaining this level throughout all subsequent generations. The minimum and maximum accuracy values indicate only instability in the population, with minimum values ranging from 0.9625 to 0.9875. This high and stable performance across generations highlights the effectiveness of the MIGA in discovering optimal MLP hyperparameters for CKD classification tasks, suggesting strong model generalizability and convergence.

Table 6 summarizes the best-performing hyperparameter configurations obtained with the MIGA on the Parkinson's, breast cancer, and CKD datasets. For the Parkinson dataset, the optimal configuration achieved is 150 hidden layer units, ReLU activation, a learning rate of 0.1, and the Adam solver, which achieves an accuracy of 95.00%. For the Breast Cancer dataset, the proposed model achieved the highest accuracy of 99.12% with 50 hidden units, the tanh activation function, a learning rate of 0.001, and the Adam solver. Additionally, on the CKD dataset, the proposed model performed best with 50 hidden units, tanh activation, a learning rate of 0.1, and the SGD solver, resulting in 100% accuracy. These results demonstrate the adaptability and effectiveness of the MIGA in fitting the MLP configurations for the three different medical datasets.

Table 7 shows the timing logs from the hyperparameter tuning experiments. For each of the three datasets, the study recorded the total wall-clock time of the 10-generation GA search, both in the standard (Single-threaded) mode and with the proposed MIGA (Parallel) evaluator, in seconds.

The proposed MLP tuned with the proposed MIGA performed better than the state-of-the-art optimization approaches across all three datasets. A summary of the comparison of the breast cancer datasets is shown in The study developed a GUI that performs hyperparameter tuning in a unified window: The top panel includes a select file button to load any CSV file; a dropdown to select the target variable; text fields for user hidden-layer sizes, activation functions, learning rates, and solvers; and a start-tuning button. The bottom panel displays a scrollable console that logs each generation’s best configuration and then summarizes the optimal hyperparameters, MIGA runtime, training time, test accuracy, confusion matrix, and classification report. Finally, a save model button exports the trained MLP. Fig. 2 shows the interface of the proposed model on the CKD dataset. Fig. 3 shows the interface on the Parkinson’s dataset. Fig. 4 shows the interface on the breast cancer dataset.

The proposed MIGA overcomes the computational cost of the GA by parallelizing fitness evaluations across multiple CPU cores, which reduces the tuning time by approximately 60% compared to a single-threaded approach [7]. This efficiency gain enables the algorithm to maintain or even expand its population size and number of generations without requiring excessive runtime. The combination of nonlinear feature extraction via radial-basis-function kernel PCA and MIGA-tuned MLP produced classifiers that generalize exceptionally well, achieving 99.12% for breast cancer, 94.87% for PD, and 100% for CKD. These results suggest that the framework can capture complex relationships in diverse biomedical datasets. Future work will apply this methodology to additional clinical cohorts to confirm its robustness and to discover dataset-specific feature interactions. The developed graphical user interface, which guides users through data import, dimensionality reduction, hyperparameter optimization, and model evaluation without any programming, is crucial for translating these advanced AI methods into everyday clinical practice.

This study introduced a disease-prediction framework that integrates kernel PCA for nonlinear dimensionality reduction, an MLP classifier, and a modified multiprocessing-enabled GA (MIGA) for efficient hyperparameter tuning. The approach was evaluated on three different medical datasets: breast cancer, PD, and CKD. The MLP-tuned MIGA outperforms models optimized with traditional methods, achieving the best accuracies of 99.12% for breast cancer, 94.87% for PD, and 100% for CKD. The use of kernel PCA enables the extraction of nonlinear structures, thereby improving predictive performance. In contrast, the parallel fitness evaluations in the MIGA reduce computational time and accelerate convergence toward optimal hyperparameters. A user-friendly GUI further enables non-expert clinicians to apply the framework to new datasets with minimal effort. This study has proven the potential of combining advanced optimization methods with nonlinear feature extraction to increase disease prediction accuracy and usability in real-world healthcare settings.

I.I. proposed the study conception, design, analysis, and interpretation; S.B. provided administrative support and study materials; and A.Y. was a major contributor to manuscript writing. All the authors read and approved the final manuscript.

The datasets analysed during the current study are available on the University of California Irvine Repository website and the Kaggle website repository, https://www.kaggle.com/datasets/vikasukani/parkinsons-disease-data-set,https://archive.ics.uci.edu/ml/datasets/Chronic_Kidney_Disease, https://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+(Diagnostic).

The authors declare that they have no competing interests.