The following are this paper’s primary contributions:

This paper presents the Improved Animated Oat Optimization (IAOO-PSO) algorithm, which integrates the Animated Oat Optimization (AOO) with the Particle Swarm Optimization (PSO) algorithm and the Oppositional-Based Learning (OBL) technique. Through increased exploration capabilities, these improvements overcome the drawbacks of current FS techniques, boosting feature subset selection and overall performance.

The remainder of the paper is structured as follows: Recent studies on two major subjects are reviewed in Section 2: Section 2.1 discusses the creation of CAD systems for DED, while Section 2.2 discusses the use of MH algorithms for FS in eye disease datasets. The suggested approach and methods used are described in full in Section 3. First performance tests utilizing the CEC2022 benchmark functions are shown in Section 4. Moreover, the comprehensive experimental results of the classification model on the DED dataset are shown in Section 5. Section 7 concludes with a summary of the paper’s conclusions.

In detecting and diagnosing DED, CAD systems are making significant strides. The enhancement of the precision, effectiveness, and applicability of these systems has been the main focus of recent research. The methods, constraints, and results of the most recent CAD systems created for DED classification are thoroughly reviewed in this section. Therefore, DED therapy depends heavily on early identification. Normal DED detection using tear break-up time, meibomian pictures, and other methods is labor-intensive, time-consuming, and inefficient. Artificial intelligence (AI) has been evolving, producing several notable advancements in the medical domain [29]. The critical value of this approach for detecting ocular eye diseases is rising [30]. The diagnosis and prediction of eye diseases, including high myopia [31], diabetic retinopathy [32], and cataracts [33], have made extensive use of it.

Yabusaki et al. [34] developed an ML model to diagnose DED that automatically defines 11 features for Interfering Fringe Color (IFC) images, such as brightness, color saturation, color ratio, and local complexity. This methodology demonstrated better inter-rater agreement (Kappa coefficient = 0.820) with eye-care practitioners and higher F1-scores (0.815, 0.095) when compared to a Bag of Visual Words (BVW) ML model. According to the researchers’ results, ML is a resource-efficient and successful method for detecting DED that reduces diagnostic biases. In a different study, Cartes et al. [35] used ML methods to classify DED patients based on their mean osmolarity records (20 with DED and 20 normal). Because the logistic classifier algorithm achieved an accuracy of 85%, the results indicate that tear osmolarity in DED can be effectively discriminated. According to the research, ML is a successful method for detecting DED utilizing variables related to tear osmolarity [35].

Yang et al. [36] studied using ML models to predict DED in patients who had undergone a vitrectomy. Using a dataset of 217 cases with features such as OSDI, TBUT, osmolarity, interocular difference, ocular surface staining corneal spot number, conjunctival spot number, length and width of lid margin, and demographic data, they used supervised ML models, specifically Logistic Regression (LR) and Artificial Neural Network (ANN). An Area Under the Curve (AUC) of 0.741 for LR and 0.786 for ANN is revealed by the Receiver Operating Characteristic (ROC) curve analysis. The study concludes that ML models are effective in detecting DED, with ANN outperforming LR in this regard [36]. Most of the research has been published in the last few years, particularly after 2014. Table 1 displays their relevant references, datasets used, objectives, algorithms used, and results.

The success of MH algorithms in addressing issues such as large dimensionality, complexity, and noise has led to their increased prominence in the literature for FS in medical datasets. This section examines the main contributions and uses of these algorithms, highlighting how they can enhance diagnostic precision and prediction performance.

An approach for detecting optic discs has been presented that combines the Ant Colony Optimization algorithm with anisotropic diffusion [50]. Another automated method for the detection of optic discs has been described using the firefly algorithm [51]. Various swarm intelligence-based methods have also been used to detect the optic disc [52]. Using their method, the authors have demonstrated that the firefly algorithm produces the best results. To efficiently detect optic discs, the gravitational law has also been applied [53], but there is still room for improvement in performance. The bat optimization approach [51] has also been used recently for optical disc detection, and the method has improved the disc segmentation results in a relatively short calculation time.

Nam et al. [54] proposed an Explainable Artificial Intelligence EAI-based ML prediction system for the detection of DED and the identification of risk and protective factors. Of the 4391 cases collected for the study, 80% were used for training from the KNHANES dataset. Several packages, such as rpart, psych, glmnet, survey, WeightedROC, ggplot2, weights, and qgraph, were used together with the R programming language (version 3.6.1). To minimize overfitting and improve robust performance, ML models were used for FS and prediction model simplification.

The study found that several important features, such as age, rhinitis, lipid-lowering medicine, sex, cataract surgery, current depression, rhinoplasty, stress and omega-3 fatty acid intake, were associated with DED [54]. Moreover, DED is a serious postoperative consequence and an important factor to take into account while developing a laser surgical strategy. EAI is a useful tool for expert-level knowledge of the criteria utilized in the selection of refractive surgery techniques, according to Yoo et al. [55]. They used ANN ML, SVM, and XGboost models to classify patients into four categories of surgeries according to the features of DED, lifestyle, and financial constraints. Accuracy was assessed using both internal and external datasets and the 10-fold cross-validation approach. Table 2 displays an overview of the MHs of FS on DED and their relevant references, datasets used, objectives, algorithms used, and results.

Although CAD systems and MHs-based FS techniques have advanced, many current methods still have issues with generalization, efficiency, and scalability. Following the no-free-lunch (NFL) [66], which states that no single optimization algorithm excels at solving every optimization problem, this study suggests a CAD system that uses the Improved Animated Oat Optimization Algorithm (IAOO) to improve the classification of DED. Using AOO with OBL and PSO to overcome the drawbacks of the original techniques and enhance the exploration and exploitation capabilities.

AOO is a recently released MH algorithm that draws inspiration from the animated oat’s natural behavior in the surroundings [17]. The three distinct behaviors of animated oat are simulated by AOO:

Seed dispersal by natural forces like wind, water, and animals.

The PSO [68] requires fewer parameters to adjust and is straightforward to use. The field of swarm intelligence studies computer systems that are motivated by the collective intelligence that emerges from the collaboration of homogeneous agents in the environment, such as fish schools and bird flocks. PSO is modeled after the structured behavior of schools of fish and flocks of birds [69].

PSO is a population-based optimization tool in which an initial population of random particles is used to guide the algorithm’s search for optima through generational updates. Let’s say the colony is made up of M particles and the search space is D-dimensional. With i=1,2,...,M, each particle has a position Xi=(xi1,xi2,...,xiD) and a velocity Vi=(vi1,vi2,...,viD). The best position discovered by any particle in the swarm is known as the global best (gbest), and when a particle travels, it recalls its best position prior, known as the personal best (pbest).

By monitoring their own best, the swarm will progress towards the top spot globally as the number of iterations increases. Pg=(pg1,pg2,...,pgD) represents the gbest of the entire swarm, while Pi=(pi1,pi2,...,piD represents the pbest of the ith particle. Three basic phases are iteratively carried out by the PSO algorithm until a stopping criterion is met. The following are [70]:

Calculate each particle’s fitness within the population by applying the optimized objective function.

(15)νi,d=ω.νi,d+c1.r1.(pi,d−xi,d)+c2.r2.(pg,d−xi,d) (16)xi,d=xi,d+νi,dwhere the learning variables c1 and c2 determine the relative influence of the cognitive and social components, respectively, and x is the inertia weight.

Conventional MH algorithms begin the search with an initial solution consisting of a set of randomly generated numbers. The algorithm’s convergence rate is unstable and typically slow. OBL is implemented to get over these issues, and both randomly produced and reverse solutions are taken into consideration.

In 2005, the field of machine learning was exposed to the main opponent concept [18]. Various OBL techniques have since been effectively added to MH algorithms to boost performance. A thorough analysis of the application and development of OBL techniques may be found in [71]. The interval [LBi,UBi] defines xi for every point X=(x1,x2,...,xD) in the D dimensional space. A distinct opposite point, xi¯, is defined as follows: (17)xi¯=LBi+UBi−xi

SVM technique, commonly used for regression and classification applications, were used as a classifier in the classification phase. Vapnik was the first to develop SVM in the late 20th century [72]. Since then, it has been used in several fields, including bioinformatics and biometrics. The fundamental idea is to use hyperplanes to divide the data into distinct classes. When the data is linearly separable, SVM performs exceptionally well; when the data is not linearly separable, it struggles. Kernel functions allow for linear separation by mapping the data to a higher-dimensional region. The performance of SVM depends on choosing the right kernel function and parameters [73]. Two important parameters are C and γ. C balances the trade-off between smooth decision boundaries and accurate training data classification, while γ shows the impact of specific training instances. Eq. (18) defines the SVM approach used for handling binary classification problems. (18)f(x)=[y,ϕ(x)]+z

Eq. (19) computes y, Eq. (20) determines z, and ϕ(x) displays the mapping of sample x to the high-dimensional space. (19)g(y,ϵ)=12||y||2+C∑i=1Nξi (20)yi([y,ϕ(xi)]+z)≥1−ξi,ξi≥0where the slack variable ξi is used to deal with misclassifications. Eq. (21) illustrates how the Radial Basis Function (RBF) kernel nonlinearly translates input vectors to the appropriate feature space. (21)k(xi,xj)=[ϕ(xi),ϕ(xj)]

This section presents the IAOO-PSO, a suggested algorithm that improves on the original AOO by combining the OBL technique and the PSO algorithm. The OBL technique is integrated into the initialization phase to increase the probability of escaping local optima by enhancing population diversity and guiding the search process closer to the global optimum. To produce more useful candidate solutions during exploitation, we also use the PSO method to accelerate convergence. This improves convergence toward the optimal solution while expanding the search in potential areas. The main drawbacks of the original AOO, such as its vulnerability to early convergence, inadequate exploitation in intricate environments, and a poor balance between exploration and exploitation, were intended to be addressed by these changes. To guarantee robustness and convergence stability, effective optimization necessitates resolving these problems, particularly the propensity to become stuck in local minima. The following describes the primary stages of the proposed IAOO-PSO algorithm:

The primary objective of developing an automated diagnostic system using machine learning (ML) techniques is to construct a predictive model that not only fits the training data effectively but also generalizes well to unseen test data—demonstrating high testing accuracy. To improve diagnostic performance, particularly disease classification accuracy, various data mining algorithms are used during the feature preprocessing stage. This preprocessing generally falls into two main categories: feature extraction and FS.

FS aims to identify a subset of features from the original dataset that are most relevant to the target labels, using either statistical or search-based techniques. This process plays a vital role in enhancing model training, controlling overfitting, and promoting generalization. Binary variants of MH algorithms offer an effective approach for exploring the feature space and identifying optimal feature subsets, even in the presence of large-scale feature dimensions. Integrating MHs into FS helps improve both model performance and classification accuracy.

Within FS and ML frameworks, the fitness function in MHs plays a central role by evaluating the quality of each candidate feature subset. Recent research has enhanced these systems by using hybrid MH approaches to reduce search space complexity while improving classification accuracy [74]. In this study, the integration of a novel DED classification method is demonstrated using a publicly available dataset of DEDs. The classification pipeline consists of four main stages: data preprocessing, feature selection using the proposed IAOO-PSO algorithm, classification, and evaluation. The overall framework of the proposed IAOO-PSO-based SVM classification model is illustrated in Fig. 1.

In this subsection, the IAOO-PSO algorithm is compared with AOO and six other MH techniques. For 12 benchmark functions with a problem dimension of 10, the mean (μ) and STD (σ) of the fitness scores achieved by IAOO-PSO and the competing algorithms are presented in Table 6.

As evidence of its improved performance, IAOO-PSO obtains the mean best fitness value of 1.21×104 for the unimodal test function (F1). PSO comes second, followed by INFO at 2.45×104. With a mean of 2.66×104, AOO is in fourth place, and BWO is in fifth place with 1.92×105. With the greatest mean value of 4.61×1010, WOA performs the worst. By consistently outperforming other algorithms and achieving the lowest mean fitness value, which demonstrates efficient convergence toward the global optimum, these results demonstrate IAOO-PSO’s excellent exploitation capabilities.

The proposed IAOO-PSO consistently outperforms all competing algorithms for the multimodal functions (F2–F5), placing first in each of the four test functions. The mean best value for function F2 is 3.85×102, which is the lowest for IAOO-PSO. INFO (1.20×103) and AOO (1.56×103) come in third and fourth, respectively, while BWO comes in fifth with 1.58×103. WOA and SCA perform less effectively, with mean best fitness values of 3.92×103 and 4.04×103, respectively. These results demonstrate how well IAOO-PSO can explore intricate multimodal environments.

The performance of IAOO-PSO is strong when evaluated on hybrid functions (F6–F8). Function F6 outperforms all other methods, achieving the mean value of 1.03×106. With 2.33×107, PSO comes in second, and INFO comes in third with 9.70×107. With the mean value of 2.03×103 for function F7, IAOO-PSO once again takes the lead. INFO (2.11×103) and BWO (2.15×103) are next in line. In function F8, IAOO-PSO delivers the most favorable outcome, achieving a mean value of 1.30×103, marginally surpassing INFO, which records 2.27×103. AOO secures third place with a mean of 2.30×103, while BWO and PSO follow with 2.35×103 and 1.00×104, respectively. WOA and SCA exhibit the least effective performance, each yielding a mean of 2.23×105. These findings highlight IAOO-PSO’s robust capability in avoiding entrapment in local optima.

IAOO-PSO obtains the optimal mean value of 3.04×103 for the composition functions (F9–F12). With a mean value of 7.71×103, SCA, WOA, and MFO exhibit the worst performance. According to these results, IAOO-PSO is highly capable of balancing exploration and exploitation in the resolution of intricate, composite optimization issues.

The overall rank and the Friedman mean rank provide additional evidence of the efficacy of IAOO-PSO. IAOO-PSO regularly performs better than the other algorithms in the comparison research, with a Friedman mean rank of 1.00 and the highest overall rank of 1. Together, these results show how well IAOO-PSO performs on a wide range of benchmark functions, indicating its accuracy, resilience, and potential as a competitive optimization method.

Boxplot analysis is a valuable tool for visualizing the distributional characteristics of data. In this study, boxplots were utilized to illustrate the performance distribution of each algorithm, with results corresponding to those presented in Table 6. The whiskers in the boxplots indicate the minimum and maximum values achieved by the algorithms, while the lower and upper bounds of the boxes represent the first (Q1) and third (Q3) quartiles, respectively. A narrower box indicates a more consistent and stable performance.

Fig. 2 displays the boxplots for functions F1 through F12 with a problem dimensionality of Dim=10. The boxplots for the proposed IAOO-PSO algorithm consistently demonstrate lower minimum values and tighter interquartile ranges for most benchmark functions. These observations suggest that IAOO-PSO not only achieves superior optimization results but also maintains a high level of stability across multiple runs. The narrow spread and lower medians reflect the algorithm’s robustness and precision, further confirming its advantage over competing MH approaches.

In comparison to several other optimization algorithms, including PSO, BWO, HHO, INFO, SCA, WOA, MFO, and the original AOO, this subsection provides a thorough examination of the convergence performance of IAOO-PSO. For every method, Fig. 3 shows the convergence graphs on the CEC’2022 benchmark functions (F1 to F12). When it comes to function F1, the IAOO-PSO algorithm outperforms its rivals in terms of early exploration capabilities and rapidly converges to reduced error levels.

WOA performs the worst, converges more slowly, and becomes stuck at larger error values than the other techniques. Similarly, IAOO-PSO outperforms all other algorithms in F2, demonstrating faster convergence and achieving a considerably lower error rate. The F3 results demonstrate IAOO-PSO’s steady declining trend, suggesting effective optimization and superior convergence over the alternative techniques. The efficient approach for space exploration is demonstrated by IAOO-PSO’s quick drop in error values in function F4.

In terms of convergence speed and accuracy, both IAOO-PSO surpass all other algorithms for F5, exhibiting comparable performance. This suggests that the IAOO-PSO is suitable for this kind of issue, successfully striking a balance between exploration and exploitation. In contrast to IAOO-PSO, other algorithms such as WOA, MFO, and SCA converge more slowly and have difficulty achieving competitive error values. In comparison to all other algorithms, IAOO-PSO shows noticeably faster convergence for F6, especially in the early phases. This demonstrates the outstanding optimization capabilities of IAOO-PSO for intricate search areas.

IAOO-PSO routinely outperforms alternative approaches, such as INFO and BWO, which suffer from slower convergence rates and less-than-ideal final outputs, in terms of convergence speed and final error levels for functions F7 through F12. This thorough performance across several functions demonstrates the IAOO-PSO’s resilience in resolving various optimization issues.

As shown in Table 7, the IAOO-PSO algorithm outperforms all other compared methods in the mean fitness of benchmark functions solved.

This section presents a comparative performance analysis of AOO variants based SVM classifier between the original algorithm (AOO), improved AOO-based OBL strategy (IAOO), integrated AOO with PSO algorithm (AOO-PSO), and the proposed IAOO-PSO based on an SVM classifier. The main focus of the comparison is the assessment of important classification performance measures using the DED dataset, such as ACC, SE, SP, PPV, FS, and F1. The results, as displayed in Table 8, reveal a notable performance difference between the four algorithms, underscoring the advantages of the IAOO-PSO optimization strategy over the fundamental AOO technique. Across all important classification metrics, the results in Table 8 unequivocally show that the IAOO-PSO-SVM model performs better than all other variants, including the AOO-SVM, IAOO-SVM, and AOO-PSO-SVM models. The ACC of the IAOO-PSO-SVM was a remarkable 99.8% greater than the ACC of the AOO-SVM 95.16%, IAOO-SVM 96.18%, and IAOO-PSO-SVM 96.39%.

Fig. 4 shows that IAOO-PSO-SVM has significantly improved overall classification performance, with more instances properly classified.

The SE, which measures the true positive rate, was 99.44% for the IAOO-PSO-SVM model and higher than the SE of 95.15% for the AOO-SVM, 96.42% for the IAOO-SVM, and 96.35% for the AOO-PSO-SVM models. This improvement implies that the IAOO-PSO-SVM is significantly more adept at detecting DED cases that are real positives, hence reducing FN. The IAOO-PSO-SVM achieved 99.03% for SP, which is 4.03% greater than the AOO-SVM’s 95.00%. This suggests that the IAOO-PSO-SVM is more successful in accurately detecting negative situations, lowering the possibility of false positives. The IAOO-PSO-SVM achieved 99.63% for PPV, which is 4.00% higher than the AOO-SVM’s 95.63%. This suggests that the IAOO-PSO-SVM is more successful in correctly predicting positive instances among all instances predicted as positive. The IAOO-PSO-SVM model obtained a 0.0027 FS, 0.0756 greater than the AOO-SVM’s 0.0783 and a 0.25 best fitness, 0.18 greater than the AOO-SVM’s 0.43. Lastly, the IAOO-PSO-SVM model obtained a 99.53% F1, 4.15% greater than the AOO-SVM’s 95.38%, the IAOO-SVM’s 96.41%, and the AOO-PSO-SVM’s 96.32%. These results show how the IAOO-PSO algorithm has improved its optimization capabilities, giving it a more effective and efficient option for DED diagnosis.

The suggested IAOO-PSO integrated with SVM (IAOO-PSO-SVM) is compared with several other MH techniques in this section. Using the DED dataset, the comparison analysis focuses on assessing how well these algorithms perform in terms of important classification parameters such as ACC, SE, SP, PPV, FS, Fit, and F1. By examining these measures, the IAOO-PSO-SVM model’s efficacy and resilience for FS and classification tasks are demonstrated, indicating its possible benefits over conventional optimization methods.

The FS optimizers included all comparative MH algorithms. Every technique was used separately to select a subset of the DED dataset’s 26 features. An SVM classifier was then trained on the selected feature subset, and performance was assessed using common metrics. The selected feature indices and the overall quantity of selected features for every method are shown in Table 9. Table 9 shows how IAOO-PSO increased classification accuracy while consistently selecting the smallest and most relevant subset of features. This demonstrates how IAOO-PSO can effectively reduce dimensionality without sacrificing model performance, which is especially useful in medical diagnostics.

A detailed comparison of IAOO-PSO against several popular optimization techniques, such as PSO, BWO, INFO, SCA, WOA, MFO, and the original AOO, for the classification task using the DED dataset is shown in Table 10 and Fig. 5. The original AOO (95.16%) and other MH algorithms are far less accurate than the IAOO-PSO, which achieves an impressive 99.8%. INFO is the nearest competitor, with an ACC of 97.63%. Because of its great ACC, the IAOO-PSO algorithm is a very dependable tool for predictive analysis, as it can accurately identify cases within the DED dataset.

The proposed IAOO-PSO achieves a notably high SE of 99.44%, reflecting its strong performance. This is far better than the highest SE of 97.82% attained by the INFO algorithm among conventional methods, and it is noticeably higher than the original AOO, which scored 95.15%. This enhanced SE demonstrates the IAOO-PSO’s capacity to reduce the possibility of FN, which is essential in medical diagnostics because failing to detect a positive instance might have serious consequences.

Additionally, the suggested IAOO-PSO performs exceptionally well in terms of SP, attaining a SP of 99.03%, surpassing both the original AOO’s 95.00% and other compared algorithms such as SCA (96.13%) and MFO (95.33%). This high SP level increases the model’s dependability and credibility by enabling the IAOO-PSO algorithm to correctly identify patients who do not have the ailment.

The effectiveness of the IAOO-PSO algorithm is further supported by the F1-score, which balances PPV and SE, which has a score of 99.53%. The INFO score of 97.80% and the original AOO’s F1 of 95.38% are both below this. Both FP and FNs are decreased by the IAOO-PSO’s strong F1, which shows that it successfully strikes a compromise between SE and PPV. Accurate classification is crucial in medical diagnostics, where this statistic is especially pertinent.

Consequently, the results presented in Table 10 and Fig. 6 show that the suggested IAOO-PSO method provides significant enhancements over the original AOO and other MH algorithms.

Our method is superior in selecting the most relevant features while enhancing the overall classification performance, as this comparison shows. The results of the comparisons of several performance measures are shown in Table 10. The suggested IAOO-PSO once more demonstrates superiority when examining the average number of selected features. With an average of 27, it selected the fewest features compared to the other approaches.

The efficacy of IAOO-PSO in FS and classification tasks related to the DED dataset is demonstrated by its superior performance across all assessed parameters, including ACC, SE, SP, PPV, F1, and FS. Table 10 and Fig. 6 demonstrate the advanced optimization capabilities and potential of the IAOO-PSO as a powerful tool for medical data classification tasks, validating the improvements made to the system. This combination of low FS and high accuracy demonstrates the superiority of the IAOO-PSO-SVM algorithm over the other compared algorithms.

The convergence behavior of the proposed IAOO-PSO algorithm is compared against several MH algorithms using their final fitness values after optimization. A lower fitness value indicates better optimization performance for a minimization problem.

The IAOO-PSO algorithm achieved the lowest fitness value of 0.25, significantly outperforming the other algorithms. Table 10 and Fig. 7 indicate that IAOO-PSO has a superior convergence capability, rapidly and effectively finding the optimal solution.

AOO and MFO had higher fit values (0.43 and 0.4028, respectively), meaning their solutions were less optimal compared to IAOO-PSO. The results demonstrate that the IAOO-PSO algorithm outperforms all other compared MH algorithms in DED classification. This reflects the effectiveness of IAOO-PSO in balancing exploration and exploitation, likely due to its hybrid design.

To provide a comprehensive comparison of the classification performance and robustness of each optimization-based model, a boxplot analysis was conducted using accuracy values collected over 30 independent runs for each algorithm. Boxplots are particularly effective for visualizing the distribution, central tendency (median), and variability (interquartile range) of performance metrics, as well as for identifying potential outliers.

Table 10 and Fig. 8 reveal that the IAOO-PSO algorithm outperforms all other methods with a median ACC of 99.80%. The corresponding boxplot for IAOO-PSO displays a very narrow interquartile range and minimal dispersion, indicating that the algorithm consistently delivers high ACC with low variance across runs. This suggests that IAOO-PSO is not only accurate but also reliable and robust in handling classification tasks.

In contrast, other algorithms such as AOO (95.16%), PSO (95.65%), BWO (95.21%), and MFO (95.62%) exhibit lower median accuracies along with wider boxplots, reflecting higher variability in performance. These results suggest less consistency and a greater likelihood of producing suboptimal outcomes in certain runs. Similarly, SCA (96.22%) and WOA (96.10%) demonstrate modest improvements in median ACC, but still fall short in terms of robustness when compared to IAOO-PSO.

Among the competing algorithms, INFO (97.63%) shows relatively better performance, with higher median accuracies and moderately narrow interquartile ranges. However, their performance still varies more significantly than that of IAOO-PSO and does not reach the same peak accuracy level.

Overall, the boxplot analysis underscores the superior performance of IAOO-PSO, which not only achieves the highest classification accuracy but also demonstrates remarkable stability and reliability across multiple runs. This highlights its potential suitability for critical applications, such as medical diagnostics, where both accuracy and consistency are essential.