Hybrid GTO-SA Metaheuristic for Feature Subset Selection in High-Dimensional Medical Datasets

The rapid growth of medical data in recent years has highlighted the urgent need for efficient preprocessing techniques that can enhance classification performance while reducing model complexity. High-dimensional datasets, especially in healthcare, often contain redundant or irrelevant features, which may lead to poor diagnostic accuracy, increased computational costs, and reduced interpretability of machine learning models. Feature subset selection (FS) therefore plays a crucial role in retaining the most informative attributes while eliminating non-contributory ones, ultimately improving decision-support systems in medicine [1].

To address the FS problem, researchers have increasingly turned to metaheuristic optimization algorithms, which are well known for their ability to navigate complex search spaces without requiring domain-specific assumptions [2-6]. Traditional approaches such as Particle Swarm Optimization (PSO) [7, 8], Genetic Algorithms (GA) [9, 10], Ant Colony Optimization (ACO) [11, 12], and Harmony Search [13] have shown promising results. However, they are often susceptible to premature convergence and may struggle to balance exploration and exploitation in high-dimensional spaces [14, 15].

Consequently, hybrid metaheuristic methods have emerged as a dominant research trend. These approaches integrate global exploration strategies with local refinement mechanisms, offering a more robust search process. For example, combinations such as Grey Wolf Optimization with Simulated Annealing (GWO-SA) [16], Harris Hawks Optimization with SA (HHO-SA) [17, 18], Whale Optimization Algorithm with SA (WOA-SA) [19], Salp Swarm with SA [20], and Moth-Flame Optimization hybrids [21, 22] have demonstrated superior performance compared to their standalone counterparts. Other novel designs, including the Capuchin Search Algorithm [23], Spotted Hyena Optimization [24], and Artificial Gorilla Troops Optimizer (GTO) [25], have further enriched the metaheuristic landscape.

Recent years have seen a surge in specialized hybrid algorithms tailored for biomedical and clinical datasets. Notable examples include the hybrid binary COOT algorithm with SA for microarray data [26, 27], a fuzzy joint mutual information approach combined with the Binary Cheetah Optimizer [28], and hybrid feature optimization techniques for cancer and brain tumor datasets [29-32]. Similarly, applications in cardiovascular disease prediction [33], biomedical signal analysis [34], and healthcare informatics [35-38] demonstrate the increasing adoption of hybrid FS methods in real-world medical domains. Several surveys also highlight this growing trend [39-42], confirming that hybridization is now a mainstream direction in FS research.

Despite these advancements, several challenges remain. Many recent studies emphasize algorithmic improvements but overlook clinical interpretability and practical deployment in healthcare environments [43]. Furthermore, while hybrid methods reduce the risk of stagnation, they can still incur high computational complexity, limiting their scalability on large medical datasets.

In this paper, we propose GTO-SA, a novel hybrid feature selection approach that combines the global search capability of the Gorilla Troops Optimizer [18] with the local refinement strength of Simulated Annealing [12, 19, 33]. By embedding SA within the GTO framework, our method aims to achieve a more effective exploration–exploitation trade-off, avoiding local optima and yielding compact yet discriminative feature subsets. To evaluate its performance, GTO-SA was applied to sixteen benchmark medical datasets and compared against widely used algorithms, including PSO [39, 40], ACO [22], GA [23, 37], Ant Lion Optimization [3], and the original GTO [18].

Our main contributions can be summarized as follows. First, we develop a hybrid GTO-SA algorithm that explicitly balances global exploration with local exploitation, thereby reducing the risk of premature convergence. Second, we design a wrapper-based framework that employs the K-Nearest Neighbors (KNN) classifier to effectively evaluate both classification accuracy and feature compactness. Finally, we conduct extensive experiments on a diverse set of medical datasets and demonstrate that GTO-SA consistently outperforms state-of-the-art metaheuristics in terms of accuracy, dimensionality reduction, and robustness [1-17, 28].

Feature selection is formulated in this work as a multi-objective optimization problem in which two competing goals must be addressed simultaneously: maximizing classification accuracy and minimizing the number of selected features. In the medical domain, where datasets are typically high-dimensional and noisy, retaining too many irrelevant features can degrade diagnostic performance, increase computational cost, and reduce interpretability.

To balance these objectives, we adopt a weighted formulation in which classification error and the feature ratio (the proportion of selected features relative to the total available features) are combined into a single fitness function. Following common practice in medical data analysis, classification accuracy is given greater emphasis with weight α = 0.99, while feature reduction is still encouraged with weight β = 0.01. This ensures that while predictive power remains the primary goal, compact subsets are also favored as shown in Table 1.

Table 1. List of metrics

The optimization process is based on the Gorilla Troops Optimizer (GTO), a recently introduced metaheuristic inspired by the social intelligence of gorilla groups [18]. GTO alternates between two main behavioral phases. In the exploration phase, gorillas spread out to search new regions of the solution space, thereby maintaining diversity. In the exploitation phase, individuals are influenced by the “silverback,” representing the best solution identified so far, allowing the population to converge around promising regions as presented in Table 2.

Table 2. List of components

This dual mechanism has been shown to be effective in global exploration but, as with many population-based algorithms, GTO is prone to premature convergence when applied to highly complex, high-dimensional medical datasets [2, 22, 39].

To mitigate this limitation, Simulated Annealing (SA) is embedded within the GTO framework as a local refinement mechanism. SA is a stochastic optimization technique that probabilistically accepts not only better solutions but occasionally worse ones, especially in the early stages of the search.

This behavior, governed by a temperature parameter that gradually decreases according to a cooling schedule, allows the algorithm to escape local minima and perform more thorough exploitation of the search space [12, 14, 33]. In this work, the initial temperature is set to 100, with a cooling rate of 0.95, which provides a balance between exploration flexibility and convergence stability. The hybridization of GTO with SA therefore allows the algorithm to benefit from the global search ability of GTO and the local refinement strength of SA, achieving a more effective exploration–exploitation balance.

The overall framework of the proposed GTO-SA Algorithms 1-3 begins with the random initialization of a population of candidate feature subsets. Each solution is evaluated using the weighted fitness function, and the best-performing subset is designated as the silverback. At each iteration, GTO operators update the positions of the gorillas, simulating exploration and exploitation. After this step, SA is applied to each updated solution, locally refining it through probabilistic acceptance of moves.

The refined solutions are then re-evaluated, and the silverback is updated if a superior subset is discovered. This iterative process continues until the maximum number of iterations is reached, at which point the final silverback represents the optimal feature subset. In this study, the algorithm is configured with a population size of 20, a maximum of 100 iterations, and an exploration probability of 0.5 to balance the tendency toward diversification and intensification.

The computational complexity of GTO-SA is dominated by two operations: the evaluation of candidate subsets using the KNN classifier and the local refinements performed by SA. For a dataset with FFF features, a population of NNN solutions, and MMM iterations. While this appears computationally demanding for very large feature spaces, the progressive reduction in dimensionality achieved during the optimization significantly reduces runtime as the algorithm converges. This makes GTO-SA suitable for medium- and large-scale medical datasets, where both accuracy and interpretability are crucial [1, 4, 6, 43].

GTO-SA consistently achieved the highest classification accuracy on the majority of datasets, with an average improvement of 3–7% compared with baseline methods such as PSO [39], ALO [3], and SCA [23]. The performance advantage was especially notable on high-dimensional datasets such as the Colon Cancer and Leukemia datasets, where GTO-SA reduced dimensionality substantially while preserving predictive power.

In terms of feature reduction, GTO-SA demonstrated clear superiority, eliminating more than 50% of features on average across datasets. For example, in the Cardiovascular dataset, the proposed method selected only 12 out of 40 features while maintaining 95% accuracy, outperforming PSO and ALO, which required nearly twice as many features to achieve comparable results. These findings highlight the strength of GTO-SA in producing compact and discriminative subsets, a property of particular importance in medical applications where interpretability and cost-effectiveness are critical.

Across different datasets, GTO-SA exhibits smooth and steady improvement in fitness values over iterations, while competing algorithms often stagnate prematurely. A more detailed comparison of the classification accuracy and feature reduction rates, which present bar charts across all tested datasets. GTO-SA consistently ranks first or among the top two methods, outperforming not only classical algorithms but also hybrid approaches such as WOA-SA [25] and HHO-SA [14, 21]. The visual results make it clear that GTO-SA achieves an effective balance: it delivers superior accuracy while simultaneously reducing the number of selected features.

To assess the reliability of these results, statistical significance testing was performed. A Wilcoxon signed-rank test was conducted on the classification accuracies across 20 independent runs, and the results confirmed that GTO-SA achieved statistically significant improvements (p < 0.05) over PSO, ALO, and SCA in 13 out of 16 datasets. Against WOA-SA and HHO-SA, the proposed algorithm was significantly better in 10 datasets and competitive in the remaining six. These outcomes are consistent with the results shown in Table III, which presents average rankings of algorithms across datasets, where GTO-SA consistently holds the top position.

In terms of runtime, reported in Table 4, GTO-SA required slightly more computation than classical algorithms due to the embedded SA refinement. However, this increase was modest, and in many cases, GTO-SA converged more quickly to optimal solutions than its competitors, offsetting the additional cost. Given the critical importance of diagnostic accuracy and interpretability in medical decision support, this runtime trade-off is considered acceptable.