Enhanced Skill Optimization Algorithm and Stacked Long Short-Term Memory with Sech Activation Function for Gastrointestinal Disease

Farah Mohammad and Muna Al-Razgan [16] presented a cosine support vector machine (CSVM) along with CNN-based Inception V3 and DenseNet201 for feature extraction to identify stomach disease. This approach combined feature effusion and concatenation using parallel maximum covariance to perform feature fusion. The CSVM effectively handled the high-dimensional feature vectors extracted by CNN, ensuring accurate classification. However, CSVM struggled with poor scalability on large datasets due to computing pairwise cosine similarity, which affected model efficiency and reduced classification performance.

Sharma et al [17] developed an LPNet–CNN with discrete wavelet transformation (DWT)-based pooling in place of conventional max and average pooling to automatically classify polyps and non-polyps. The DWT enabled multi-resolution analysis, allowing LPNet-CNN to capture both global and local features effectively, thereby improving classification accuracy. However, LPNet was sensitive to noise and texture variations because wavelet decomposition amplified high-frequency components, leading to overlap with noise and misclassification.

Naz et al. [18] implemented a hybrid feature model that fine-tuned the ResNet18 model and evaluated it using XcepNet23 and local binary pattern (LBP) features combined with the CNN method. The LBP excelled at capturing features, making the hybrid approach more robust to variations in medical images. ResNet18 and XceptNet23 captured high-level features, while LBP focused on low-level spatial patterns. Nevertheless, ResNet18, having relatively shallow depth, struggled to capture highly complex features, leading to inaccurate performance.

Intissar Dhrari Hajsalem et al. [19] presented an encoder-decoder network (EDN) to detect GI disease. Initially, the Region of Interest was detected by utilizing the encoder and decoder network to select appropriate and significant information. Pre-trained methods such as ResNet50, VGG16, InceptionV3, VGG19, and support vector machine (SVM) were employed to extract features and accurately classify GI diseases. The EDN consumed more time but achieved better results by eliminating inappropriate information and effectively determining pertinent features. However, EDN lost fine-grained spatial information because the encoding process involved successive downsampling and pooling operations that minimized feature map resolution.

Samra Siddiqui et al. [20] established SNet to classify GI disorders using endoscopic images. Initially, endoscopic images were preprocessed using resizing and augmentation to ensure uniform image dimensions and enhance image quality. The Minimum Redundancy Maximum Relevance was used to select features, thereby enhancing model performance. However, SNet had limited ability to capture complex and subtle features because its shallow feature extraction layers could not sufficiently model complex variations and textures.

Hyun-Cheol Park et al. [21] suggested a Star-Generative Adversarial Network (Star-GAN) aimed at resolving challenges in learning mappings across multiple domains by integrating the multi-translation with InceptionNet-V3 method. The Star-GAN handled multiple domains with a signal generator and discriminator, significantly reducing complexity compared with traditional GAN. The InceptionNet-V3 extracted features and captured intricate patterns in the data, ensuring improved classification performance. The Star-GAN was efficient in capturing complex dependencies but lacked model diversity, as InceptionNet-V3 performance depended on the quality of extracted features. However, Star-GAN generated unrealistic images due to heavy reliance on limited data, which restrained the model’s ability to learn diverse and accurate features.

Henrietta Adjei Pokuaa et al. [22] introduced a CN to recognize GI conditions. Although the CN achieved low performance on complex dataset, it was enhanced by modeling complex dependencies between features, which benefited the classification of GI images and enabled the extraction of richer information from the GI disease image. Nevertheless, CN was sensitive to small variations because it encoded spatial relationships and posed information about features; hence, minor distortions misaligned vectors, leading to incorrect feature routing among capsules and reduced accuracy.

From the overall analysis, existing techniques had limitations such as poor scalability, sensitivity to noise and texture variations, difficulty in capturing highly complex features, and failure to select relevant features, leading to insufficient representation and reduced accuracy. Hence, this research proposes ESOA with Stacked LSTM-SAF to classify GI diseases effectively. ESOA selects the most appropriate features, minimizing the impact of irrelevant data and ensuring more informative feature representation. Moreover, the Stacked LSTM captures highly intricate temporal and spatial patterns, enhancing the model’s ability to identify subtle regions. This integration improves robustness against noise and texture variations while maintaining scalability. As a result, the proposed method achieves more accurate and reliable classification compared with existing methods.

In this section, two datasets—Kvasir V1 and V2—are considered for GI tract classification. The dataset details, including image counts and categories of GI diseases, are described below.

After collecting the images, the preprocessing stage is performed using AHE and the Bilateral filter [25] to enhance image quality and minimize noise. A detailed description of these methods is provided below.

AHE: addresses this issue by using a unique histogram for each pixel in the image. This histogram is calculated based on the intensity values within a surrounding window, known as the contextual region. Compared with HE, AHE provides better performance as it effectively addresses contrast enhancement. It effectively controls the level of contrast enhancement in the output image, especially when the original image has low contrast, where most values are concentrated in the middle of the grayscale image.

Bilateral filter: This method filters the image based on both range and pixel values. It is a local, non-linear, and non-iterative technique that considers gray-level similarities and neighboring pixels to enhance image contrast by preserving fine information and local structures. This approach effectively addresses issues of image noise and varying brightness by combining AHE and bilateral filtering to enhance contrast. The process involves applying AHE to each color channel of the input image, followed by bilateral filtering on the equalized channels. The primary goal of the bilateral filter is to preserve edges while filtering the image. The filtered channels are then recombined into a unified RGB image, and normalization procedures are applied to enhance overall contrast. The bilateral filter is mathematically expressed in Eq. (1), where the image intensity at a pixel position P is defined I_P, where G_σ denoted as a 2D Gaussian kernel. The normalization factor, G_P ensures that the sum of pixel weights equals 1.0. The parameters s & r control the extent of filtering applied to the image, as represented in Eqs (2)–(4) for the HE process, which modifies image intensities to improve contrast: (1) BFIP=1Wp∑qεsGσsP−QGσsIp−IqIq BF{\left[ I \right]_P} = {1 \over {{W_p}}}\sum\limits_{q\varepsilon s} {{G_{{\sigma _s}}}\left( {\left| {\left| {P - Q} \right|} \right|} \right)} G{\sigma _s}\left( {\left| {{I_p} - {I_q}} \right|} \right){I_q} (2) WP=∑qεsGσsP−QGσsIp−IqIq {W_P} = \sum\limits_{q\varepsilon s} {{G_{{\sigma _s}}}\left( {\left| {\left| {P - Q} \right|} \right|} \right){G_{{\sigma _s}}}\left( {\left| {{I_p} - {I_q}} \right|} \right){I_q}} (3) Ii,j=floorL−1∑n=0fi,jPn {I_i},j = {\rm{floor}}\left( {L - 1} \right)\sum\limits_{n = 0}^{{f_i},j} {{P_n}} (4) Pn=InTPn=0,1,.....,L−1 {P_n} = {{{I_n}} \over {{T_P}}}n = 0,1,.....,L - 1

Here, L denotes the count of possible intensity values, I_n represents the number of pixels with a intensity, and n and T_P indicate the total number of pixels. These two strategies—AHE and the Bilateral filter—are utilized to enhance image quality and remove noise in GI images. The preprocessed images are then fed into the feature extraction to extract the high-level feature. Figure 2 represents the sample original and preprocessed images.

Figure 2:

Sample original and preprocessed image.

The preprocessed input is fed into the feature extraction using DenseNet-201 and EfficientNet-B3 to extract high-level features from GI images. Deep CNN models are recognized for their robust classification performance in GI images. Their layers have direct connections to the subsequent layers, which enhance the learning rate while minimizing information loss. Compared with other CNN models, DenseNet-201 network requires fewer parameters. The model ensures high information retention with minimal loss from the first to the last layer and incorporates a gradient function to reduce risk of overfitting.

DenseNet-201: The input size is set to 224 × 224 × 3 with 201 structured layers organized into four dense blocks. A global average pooling layer produces a 1920-dimensional feature vector for feature extraction. The GI images ζ_z are used as input to DenseNet-201 [26], which consists of M layers and transformation filters denoted by the non-linear function S_m (.). The transformation filter s_m (.) is a concatenated function that includes convolution, batch normalization, pooling, and rectified linear unit (ReLU). In a classical CNN model, the output of m^th layer serves as the input to (m + 1)^th layer, mathematically represented in Eqs (5) and (6): (5) Bm=SmBm−1 {B_m} = {S_m}\left( {{B_m} - 1} \right) (6) bn=Smb0,b1,.......,bn−1 {b_n} = {S_m}\left( {\left[ {{b_0},{b_1},.......,{b_{n - 1}}} \right]} \right)

Here, each layer is directly connected to every other layer, and m^th layer carries information evaluated from all network layers b₀, b₁,......., b_n−1. The computed feature map of layer 0,..... n − 1 is defined in Eq. (6).

EfficientNet-B3: This method processes input images of size 300 × 300 using compound scaling across width, depth, and resolution to balance accuracy and efficiency. It adjusts these characteristics using predetermined factors, and the member of this group amplifies the equilibrium by providing significant generalizability. Initially, the input image size of 600 × 600 pixels is compared with that of EfficientNet-B2, which contains 10.7 million parameters. It includes three levels replaced with additional layers for the use case of GI image classification within the subsets of classes. Furthermore, these layers were modified by removing and incorporating additional dense, batch normalization, and dropout layers. A global average pooling layer produces a 1536-dimensional feature vector. Finally, a dense layer with SoftMax activation is used for multi-class classification, effectively capturing complex patterns. Both DenseNet-201 and EffcientNet-B3 [27] are fine-tuned using the Adam optimizer with a batch size of 32, a learning rate of 0.001, and the cross-entropy loss function. The features extracted from DenseNet-201 and EfficientNet-B3 are concatenated to form a joint representation, which is then fed into the feature selection using ESOA [28].

After extracting features, the ESOA is applied to select the most relevant features. The SOA is chosen for its ability to mimic human skill learning, allowing efficient exploration and exploitation for feature selection in high-dimensional data. ESOA is more efficient in navigating complex and high-dimensional search spaces due to the enhanced exploration and exploitation capabilities of the GTO. Compared with existing methods such as the osprey optimization algorithm (OOA), coati optimization algorithm (COA), bald eagle search optimizer (BES), and SOA, the ESOA improves convergence speed and solution accuracy by incorporating GTO. Algorithms such as OOA, COsA, BES, and SOA suffer from premature convergence, which limits their ability to identify global optima. By integrating GTO into SOA, exploration ability is enhanced through diverse search behavior inspired by gorilla leadership dynamics. This integration avoids local trapping, enhances feature relevance, and ensures better generalization. Thus, ESOA combines SOA’s adaptive learning with GTO’s robust exploration capability to achieve superior feature selection and higher diagnostic accuracy. The initialization process of ESOA is defined in Eqs (7) and (8): (7) X=X1⋅⋅⋅X2⋅⋅⋅XNN×m=X1,1X1,dX1,m⋅⋅⋅Xi,1Xi,dXi,m⋅⋅⋅XN,1XN,d,XN,mN×m X = {\left[ {\matrix{ {{X_1}} \cr \cdot \cr \cdot \cr \cdot \cr {{X_2}} \cr \cdot \cr \cdot \cr \cdot \cr {{X_N}} \cr } } \right]_{N \times m}} = {\left[ {\matrix{ {{X_{1,1}}{X_{1,d}}{X_{1,m}}} \cr \cdot \cr \cdot \cr \cdot \cr {{X_{i,1}}{X_{i,d}}{X_{i,m}}} \cr \cdot \cr \cdot \cr \cdot \cr {{X_{N,1}}{X_{N,d}},{X_{N,m}}} \cr } } \right]_{N \times m}}

Here, X represents the SOA candidate solution, X_i denotes the ith candidate solution, and X_i,d indicates the ith variable corresponding to the ith solution of SOA. N refers to the population count, while m denotes the dimension. (8) F=F1⋅⋅⋅F2⋅⋅⋅FNN×m=FX1⋅⋅⋅FXi⋅⋅⋅FXNN×1 F = {\left[ {\matrix{ {{F_1}} \cr \cdot \cr \cdot \cr \cdot \cr {{F_2}} \cr \cdot \cr \cdot \cr \cdot \cr {{F_N}} \cr } } \right]_{N \times m}} = {\left[ {\matrix{ {F\left( {{X_1}} \right)} \cr \cdot \cr \cdot \cr \cdot \cr {F\left( {{X_i}} \right)} \cr \cdot \cr \cdot \cr \cdot \cr {F\left( {{X_N}} \right)} \cr } } \right]_{N \times 1}}

The arbitrary assignment covers all possible areas in the given N × m search space, and the objective function is evaluated for each individual member in the issue variables. The OF vector is expressed in Eq. (8), where F represents the ith member values of the OF, and F denotes the set of all members as an objective vector. The population is updated in two stages: exploration and exploitation. The second phase, based on skill enhancement, involves distinct activities and efforts.

d.iii

Exploitation

In this section, each member strives to enhance its skill through continuous activity and practice. The SOA notion formulates a local search (lS) to improve exploitation, where every member explores solutions within its neighborhood to enhance objective values. The most suitable candidate solution expands the OF, and this secondary phase is expressed in Eqs (16) and (17): (16) XiP2:Xi,dP2=fx=xi,d+1−2rt×xi,d,x<0.5xi,d+ibj−rubj−ibjt, else X_i^{{P^2}}:X_{i,d}^{{P^2}} = f\left( x \right) = \left\{ {\matrix{ {{x_{i,d}} + {{1 - 2r} \over t} \times {x_{i,d}},x < 0.5} \hfill \cr {{x_{i,d}} + {{i{b_j} - r\left( {u{b_j} - i{b_j}} \right)} \over t},\;{\rm{else}}} \hfill \cr } } \right. (17) Xi=XiP2,FiP2<FiXi,else {X_i} = \left\{ {\matrix{ {X_i^{P2},F_i^{P2} < {F_i}} \cr {{X_i},{\rm{else}}} \cr } } \right.

Here, XiP2 X_i^{{P^2}} denotes the updated position of the ith candidate after the second step, xi,dp2 x_{i,d}^{p2} represents the current position of the ith candidate in the dth dimension, Fip2 F_i^{p2} indicates the recent objective value, and t denotes the number of iterations. lb_j and ub_j define the boundaries of the jth variable, r illustrates a random number, t represents the current iteration number, lb_j denotes the lower bound of the j_th variable, and ub_j denotes the upper bound of the j_th variable. Figure 3 demonstrates the workflow of the ESOA method.

Figure 3:

Workflow of the ESOA method. ESOA, enhanced skill optimization algorithm.

In this section, GI disease classification is performed using the Stacked LSTM-SAF [29], which learns and predicts patterns in sequential data and has the ability to process entire data sequences efficiently. The Stacked LSTM provides a hierarchical structure that captures long-term dependencies more effectively compared with existing methods such as the recurrent neural network (RNN), LSTM, and gated recurrent unit (GRU). Unlike GRU and single-layer LSTM models, the Stacked LSTM learns deeper temporal patterns and intricate feature representations by integrating multiple layers. Existing models such as RNN, GRU, and LSTM are limited in handling complex sequential dependencies and lack robustness and scalability to capture fine-grained variations. Hence, the Stacked LSTM offers improved accuracy and generalization for sequence modeling tasks. LSTM uses cells and gates to control the flow of information, preventing the loss of short-term memory. An LSTM layer consists of memory blocks with input, output, and forget gates, as well as one or more memory cells coupled in loops within each block. These three multiplicative units mimic the actions of resetting memory cells. The input, output, and forget gates are evaluated using Eqs (18)–(20): (18) Gi=σytWyi+St−1Wsi+βi {G_i} = \sigma \left( {{y_t}{W_{yi}} + {S_{t - 1}}{W_{si}} + {\beta _i}} \right) (19) Gf=σytWyf+St−1Wsf+βf {G_f} = \sigma \left( {{y_t}{W_{yf}} + {S_{t - 1}}{W_{sf}} + {\beta _f}} \right) (20) Go=σytWyo+St−1Wso+βo {G_o} = \sigma \left( {{y_t}{W_{yo}} + {S_{t - 1}}{W_{so}} + {\beta _o}} \right)

The input node of the memory cell is evaluated using Eqs (21)–(23), while the internal and hidden states of memory cell are evaluated by the activation function: (21) θ˜t=tanhytWyc+St−1Wsc+βc {\tilde \theta _t} = \tan h\left( {{y_t}{W_{yc}} + {S_{t - 1}}{W_{sc}} + {\beta _c}} \right) (22) θt=Gf⊙θt−1+Gi⊙θ˜t {\theta _t} = {G_f} \odot {\theta _{t - 1}} + {G_i} \odot {\tilde \theta _t} (23) St=Go⊙tan hθt {S_t} = {G_o} \odot \tan \,h\left( {{\theta _t}} \right)

Here, t denotes the time step, and y_t indicates the input to the Stacked LSTM model. In the Stacked LSTM model, multiple hidden layers are employed instead of a single layer, allowing parameters to be distributed across the model to accelerate processing and better reprsent raw data using the Sech activation function. The Sech activation function is chosen because it provides smoother gradient flow, which helps prevent vanishing and exploding gradients compared with the ReLU and Sigmoid. Unlike ReLU, which suffers from the “dying neuron” problem caused by zero gradients for negative inputs, Sech remains symmetric and continuously differentiable around zero, ensuring stable updates. In contrast to the Sigmoid, which squashes values between 0 and 1—leading to saturation—Sech manages a wider dynamic range, thereby minimizing the risk of gradient information. Overall, Sech improves model stability and convergence speed. The mathematical formulation of the Sech activation function is expressed in Eqs (24) and (25): (24) fai=ai sechai f{\left( a \right)_i} = {a_i}\;{{\rm sech}} \left( {{a_i}} \right) (25) a=b+Wtx a = b + {W^t}x

Here, a_i represents the function expressed in Eq. (24), w indicates the input weight, x represents the input, and b denotes the bias. This activation function is shown in a ordination system, where x lies in the range (−10, 10). The Sech is symmetric around zero and bounded between 0 and 1, similar to the sigmoid, which helps maintain stability during gradient-based optimization. It is defined for all integers and is fully differentiable and odd-symmetric with respect to the origin, which is essential for optimization and achieving better convergence. The hyperparameters of the LSTM-SAF are optimized using grid search to achieve optimal training performane. This setup employs a learning rate of 0.001, the Adam optimizer, Sech activation function, a batch size of 64, a weight decay of 1e⁻⁴, and a maximum of 50 epochs with early stopping (patience = 8) to effectively learn patterns from the data. Algorithm 1 presents the pseudocode of the proposed method to ensure reproducibility. Figure 4 depicts the structure of the Stacked LSTM, and Figure 5 shows sample images of actual and predicted labels for (A) Kvasir-V1 (B) Kvasir-V2.

Figure 4:

Structure of the stacked LSTM. Stacked LSTM, stacked long short-term memory.

Figure 5:

Sample images of actual and predicted labels (A) Kvasir-V1 (B) Kvasir-V2 dataset.

In this section, the proposed method, involving both feature selection and classification processes, is evaluated using several performance metrics, including Accuracy, Precision, F1-score, and Recall. Table 1 presents the performance analysis of different feature selection methods. When compared with existing methods such as the OOA, COA, BES, and SOA, the proposed ESOA achieves superior accuracies of 99.60%, 99.88%, and 99.74% on the Kvasir-V1, V2, and HyperKvasir dataset, respectively. This is due to the integration of the GTO into SOA, which enhances the exploration and exploitation to escape local optima. The adaptive search mechanism enables more accurate feature selection, which directly improves classification performance. The feature selection process is independently evaluated by measuring both the objective function and fitness value, ensuring that the selection effectiveness is validated prior to being used in the classification stage. The parameter settings for the ESOA include a population size of 30, a maximum of 100 iterations, and an inertia weight of 1.00.

Table 2 indicates the performance analysis of different classification methods across three datasets. When compared with existing methods such as the RNN, GRU, and LSTM, Stacked LSTM achieves superior accuracies of 99.60%, 99.88%, and 99.74% on the Kvasir-V1, V2, and HyperKvasir dataset, respectively. This improvement is attributed to its multiple-layer structure, which captures intricate temporal dependencies and hierarchical feature representations. The Stacked LSTM enables the model to learn both long-term and short-term patterns more effectively than LSTM, thereby enhancing generalization over varying input sequences and enhance robustness.

Table 3 provides the performance evaluation of different activation functions. Compared with existing methods, the proposed method yields better performance because Sech provides smoother gradients and prevents saturation issues. This helps maintain stable training and minimizes vanishing gradient problems across multiple layers. Consequently, the model captures deep temporal dependencies with minimal information loss, enhancing overall classification accuracy. The t-test is employed to analyze whether the difference between the means of two groups is statistically significant, taking into account sample size and variance. The use of p-values obtained from the t-test quantifies the likelihood that observed performance differences occurred by chance, ensuring a reliable interpretation of results. Additionally, the confidence interval (CI) provides a range of values within which the true performance metric is likely to lie, serving as a measure of estimation reliability. The CI helps assess the consistency and robustness of the proposed method.

Figure 6 presents the confusion matrices of the Stacked LSTM-SAF for (A) Kvasir-V1, (B) Kvasir-V2, and C) HyperKvasir datasets. Each matrix shows the number of correctly and incorrectly classified instances across eight GI disease classes. The proposed model demonstrates high classification performance, with minimal misclassifications observed in a few instances such as normal-z-line and ulcerative-colitis. These results indicate the model’s strong ability to differentiate subtle variations in GI images across all datasets.

Figure 6:

Evaluation of confusion matrix for Stacked LSTM-SAF (A) Kvasir-V1, (B) Kvasir-V2, (C) HyperKvasir. Stacked LSTM-SAF, stacked long short-term memory with Sech activation function.

Figure 7 depicts the receiver operating characteristics (ROC) curve analysis of the Stacked LSTM-SAF (A) Kvasir-V1, (B) Kvasir-V2, and (C) HyperKvasir datasets. All classes achieve high Area Under the Curve, demonstrating the model’s robustness and strong generalization performance across dataset. Overall, these results confirm the high effectiveness of the proposed model in accurately classifying GI diseases.

Figure 7:

Evaluation of RoC curve for Stacked LSTM-SAF: (A) Kvasir-V1, (B) Kvasir-V2, (C) HyperKvasir. ROC, receiver operating characteristics; Stacked LSTM-SAF, stacked long short-term memory with Sech activation function.

Figure 8 represents the performance analysis in terms of the standard deviation of the proposed method for (A) Kvasir-V1, (B) Kvasir-V2, and (C) HyperKvasir datasets. The proposed method attains high accuracy with low variance, indicating consistent performance across all dataset. The small error bars reflect low standard deviation, which demonstrates model stability and robust classification performance.

Figure 8:

Analysis of standard deviation for the proposed method: (A) Kvasir-V1, (B) Kvasir-V2, (C) HyperKvasir.

Table 4 illustrates the computational complexity of different methods across the three datasets. Compared with existing methods, the Stacked LSTM-SAF achieves shorter training times of 2.54 s, 2.73 s, and 5.80 s. The Sech function contributes to this improvement, as its symmetric and smooth nature prevents gradient explosion and stabilizes backpropagation, resulting in rapid convergence. Moreover, the Stacked LSTM minimizes redundant iterations during training, further achieves rapid training time.

Table 5 demonstrates the cross-dataset validation results, where the model is trained on Kvasir-V1 and tested on Kvasir-V2 evaluate its generalization capability. Compared with existing methods, the Stacked LSTM-SAF achieves higher performance across all evaluation metrics, confirming its robustness against dataset variations. This result highlights the model’s strong ability to generalize effectively to unseen data, ensuring reliability for practical medical applications.

Table 6 describes the comparative analysis of the proposed method on the Kvasir-V1 and V2 datasets. When compared with existing methods in [17, 19, 20,21,22], the proposed ESOA with Stacked LSTM-SAF achieves higher accuracies of 99.60% and 99.88% on the Kvasir-V1 and V2 dataset, respectively. This improvement is attributed to ESOA ensuring optimal feature selection, which minimizes inappropriate data and enhances the classification. The Stacked LSTM-SAF effectively learns both shallow and deep temporal dependencies, with stable learning via SAF improving convergence, robustness, and overall classification performance.

The strengths of the proposed ESOA with Stacked LSTM-SAF and the limitations of existing techniques are discussed in this section. Existing techniques face several drawbacks: (1). CSVM [16] struggled with poor scalability in large dataset. (2). LPNet [17] was sensitive to noise and texture variations because wavelet decomposition amplified high-frequency components, leading to noise overlap and misclassification. (3). ResNet18 [18] had relatively shallow depth, which limited its ability to capture highly complex features, resulting in inaccurate performance. (4). EDN [19] lost fine-grained spatial information due to encoding process. (5). SNet [20] had limited capacity to capture complex and subtle features due to shallow feature extraction layers, leading to suboptimal performance. The proposed ESOA with Stacked LSTM-SAF overcomes these limitations. ESOA ensures optimal feature selection by removing redundant features, enhancing data relevance and reducing training time. The Stacked LSTM-SAF captures both short-term and long-term dependencies, maintains stable training, avoids overfitting, and improves convergence efficiency. As a result, the proposed method achieves higher accuracy, reliability, and robustness compared with existing methods.