This study aims to create an ensemble ML-based model using the Boruta IMRMR technique to select features and the SSA algorithm to optimize those features. The research’s key contribution is summed up as follows.

• To develop an ensemble ML model for efficient Cancer diagnosis.

For the current work, 130 reports are initially identified. Several records are excluded from the study with different steps. Twenty duplicate records are removed before the screening process. The remaining 110 numbers of records are considered for the screening phase. In this phase, ten irrelevant records are excluded. From the screening phase, 100 records are processed for the retrieval phase, out of which 7 records are excluded as those records could not be retrieved. Hence, the remaining 93 full-text records are considered. From the considered full texts, 17 records are excluded as sufficient data for analysis are unavailable. In addition, 20 records are excluded as those dealing with diseases other than cancer, and 7 records are found to be irrelevant to the study. Finally, the remaining 49 numbers of articles are considered for the current work. From the considered records, 20 articles are used in the literature survey part, and the remaining 29 numbers of records are considered in the rest of the manuscript. The Figure in Appendix I shows the current work’s Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA).

Ibrahim et al. (2017) introduced a novel hybrid approach, the Salp Swarm Algorithm in feature selection (SSA-FS), on the real datasets obtained from Iraqi hospitals for breast, bladder, and colon cancers. Hegazy et al. (2018a) introduced a novel method, the chaotic salp swarm algorithm (CSSA), to enhance the SSA on 28 datasets. Hegazy et al. (2018b) introduced a novel method, the improved salp swarm algorithm (ISSA), which is consolidated with the KNN classifier for feature selection on 23 UCI-ML datasets and claimed to achieve enhanced accuracies on Breast Cancer, Lung Cancer, and BeastEW disease datasets respectively. Using a neighborhood entropy-based uncertainty measures model, Sun et al. (Sun et al., 2019) successfully applied machine learning (ML) methods, including k-nearest neighbor (KNN), C4.5, and Support Vector Machine (SVM), to the classification of the colon, diffuse large B-cell lymphoma (DLBCL), leukemia, lung, and small round blue cell tumor (SRBCT). Ghoniem (2020) introduced a novel bio-inspired liver cancer diagnosis model considering a deep learning (DL) approach, i.e., Convolutional Neural Network (CNN) along with SegNet and UNet, and the optimization technique, i.e., Artificial Bee Colony optimization (ABC) on Radiopaedia and LiTS datasets. Shukla et al. (2020b) introduced an adaptive inertia weight teaching-learning model considering machine learning approaches, i.e., Support Vector Machine (SVM), Extreme Learning Machine (ELM), and Naïve Bayes (NB) on Breast Cancer, Colon Cancer, DLBCL, Leukaemia, SRBCT, Lung Cancer.

Meenachi and Ramakrishnan (2020) introduced differential evolution and global optimal feature selection for cancer data classification model considering the machine learning (ML) approach, i.e., Decision Tree (DT) and optimization technique, i.e., Ant colony optimization (ACO) on five datasets, i.e., DLBCL, Breast Cancer, Leukemia, SRBCT, Gisette datasets. Nouri-Moghaddam et al. (2021) introduced a new hybrid solution based on a multi-filter and adaptive chaotic multi-objective forest optimization algorithm (AC-MOFOA) considering Forest optimization algorithm (FOA), Extreme learning machine (ELM), multi-objective optimization (MOO), and five filter methods, i.e., IG, mRMR, RelifF, CFS, and Fisher-score, on nine datasets, i.e., SRBCT, Tumors_9, Leukaemia3, colon_prostate, Lung, GCM, Breast, Rsctc_5, Rsctc_6.

Yan et al. (2021) introduced a novel feature selection model considering a machine learning (ML) approach, i.e., K-nearest neighbor (KNN) and a wrapper feature selection algorithm FS_SSA based on Salp swarm, on five datasets, i.e., ALL-AML-4, Colon Cancer, Lymphoma, MLL, SRBCT datasets. Sarala et al. (Arun Prabha et al., 2021) introduced a decision-based Salp Swarm Optimization (DT-SWO) algorithm considering machine learning (ML) approaches, i.e., Decision Tree (DT), Support Vector Machine (SVM), Naïve Bayes (NB), Kernel Support Vector Machine (KSVM), and optimization technique, i.e., Salp Swarm Optimization (SWO) on four datasets, i.e., DLBCL, Leukemia, Lung Cancer and colon datasets. Alomari et al. (2021) introduced a hybrid filter-wrapper approach considering robust Minimum Redundancy Maximum Relevancy (rMRMR) as a filter approach, Modified Gray Wolf Optimization (MGWO) as a wrapper approach, and ML approaches including Random Forest (RF), Elastic Networks (EN) and Decision Tree (DT) on nine datasets. Balakrishnan et al. (2021) introduced an improved salp swarm algorithm (iSSA) based on the levy flight for feature selection model and Support Vector Machine (SVM) classifier on six datasets, i.e., Oral Squamous Cell Carcinoma (OSCC), Ovarian cancer, Breast Cancer, CNS, Colon Cancer, Leukemia datasets. Hameed et al. (2021) included the binary particle swarm optimization (BPSO), the genetic algorithm (GA), and the cuckoo search algorithm (CS) for selecting the features. ML approaches, including SVM, NB, KNN, and RF, are applied to twelve datasets.

Gene selection for microarray data categorization was developed by Rostami et al. (2022) using a multi-objective graph theoretic-based approach model that considers the idea of community detection with node centrality (CDNC). On six datasets, Aziz (2022) presented a metaheuristics model that was inspired by nature. The model utilized ML approaches such as Support Vector Machine (SVM), Naïve Bayes (NB), Artificial Neural Network (ANN), cuckoo search (CS), genetic algorithm (GA), and artificial bee colony (ABC). Alromema et al. (2023) used logistic regression, Support Vector Machine, K-Nearest Neighbours, Neural Networks, Naive Bayes, Decision Tree, and eXtreme Gradient Boosting to gene expression datasets. On 17 microarray expression datasets, including CNS, Colon, Leukemia_3C, Leukemia_4C, Leukaemia, Hungtington Disease, DLBCL, Lymphoma66 ö4026_3c, Lymphoma, Prostate, SRBCT, Lung Cancer, Breast Cancer, Sarcoma, Mycloma, and Ovarian, Ke et al. (2022) suggested a population initialization method based on ranking criteria (PIRC) using NB, C4.5, genetic algorithm (GA), and ant colony optimization (ACO). In their study on the Wisconsin Breast Cancer Dataset (WBCD), Rustagi et al. (2024) presented a method for breast cancer detection that relies on Salp Swarm and Grey Wolf Optimisation. They took SVM and KNN classifiers into account. Ünalan et al. (2024) highlighted ensemble learning methodologies’ efficacy in classifying breast cancer. It was revealed that performance improved over standalone classifiers. Classifiers AdaBoost, GBM, and RGF gave an impressive accuracy of 99.5%. However, ensembles of this kind surpassed the respective individual algorithms LGBM for accuracy and gave an F1 score of 99.2% alongside an accuracy of 98.9%. Incorporating stratified shuffle split and k-fold cross-validation raises the question of the strict evaluation technique in obtaining credible and clinically relevant classification outputs. Table 1 shows the analytical study of the literature mentioned above. Batool and Byun (2024) showed that ensemble learning applied to the task of breast cancer classification was doing well by enhancing the general predictive accuracy through a set of multiple models. Several studies on the WBCD dataset showed that the ensemble method, for example, voting classifier involving ETC, LightGBM, RC, and LDA models, performed better than individual models. The proposed model succeeded in surpassing all known state-of-the-art classifiers utilized in the detection and diagnosis cases of breast cancer with an average accuracy of 97.6% and an F1 value of 98.1%. Mahesh et al. (2022) proposes an early prediction of breast cancer through a blended ensemble learning approach with SVM, KNN, DT, RF, and LR as base classifiers. The model’s performance is checked using a breast cancer dataset, which has yielded considerable improvements in accuracy at 98.14%. Accurate, recall, precision, and F1-score metrics validate the ensemble model’s effectiveness over individual classifiers.

Table 1 shows the analytical study of the considered literature in which it has been observed that the reported literature follows a two-stage feature selection process. In addition, the adopted approaches follow a hybrid approach for cancer classification. Applying the two-stage feature selection does not impact the undertaken dataset more, i.e., ALL-AML, Lymphoma, MLL, and SRBCT cancer dataset. Thus, the current research aims to apply a pipelined feature selection method consisting of three different feature selection approaches, starting from Boruta and followed by IMRMR and SSA. However, to make it more effective, Table 2 has been taken from Table 1, which is dedicated to the literature that deals with the dataset considered for the current study, such as ALL-AML, Lymphoma, MLL, and SRBCT.

The remaining parts of the paper are organized as follows. Section 2 depicts the approach used in developing the suggested model and describes the utilized dataset. The core ideas behind the proposed paradigm are discussed in Section 3. The empirical evaluation of the suggested model is presented in Section 4. In Section 5, we do a critical analysis of the proposed model. The conclusion is presented in Section 6.

Four publicly available cancer gene expression data, including ALL-AML (D1), Lymphoma (D2), MLL (D3), and SRBCT (D4), are considered for developing the current work (Zhu et al., 2007). Among the above datasets, D1, D2, and D3 have three classes, and D4 has four classes. ALL-AML dataset contains three classes labeled B-Cell, T-Cell, and AML. The Lymphoma cancer dataset includes DLBCL, FL, and CLL classes. Considering the MLL cancer dataset, it contains three classes, including ALL, AML, and MLL. Similarly, the class names for the SRBCT cancer dataset are BL, EWS, NB, and RMS. Table 3 shows the proposed model’s dataset description and the class-wise distribution in each dataset.

Boruta is a wrapper-based feature selection that aims to improve the efficiency and interpretability of machine learning models. Boruta finds and chooses important features from a dataset by combining a random forest classifier with a significance testing method. Shadow features, essentially randomized replicas of the original features, are first generated by the algorithm. After that, the original and shadow feature sets are pooled and used to train a random forest model. Based on how each feature affects the model’s performance, Boruta ranks them and compares them to their corresponding shadow characteristics to determine their relative relevance. Initially, dataset (D) contains the total number of features. F ← f 1 , f 2 , … f N , where N is the total samples present in that dataset, then the shadow of the actual features is calculated using Equation 1 to form a shadow feature set (Fshadow): f i , s h a d o w = S h u f f l e f i = f i Ρ 1 , f i Ρ 2 , f i Ρ 3 , . . , f i Ρ N (1)

Where Shuffle is a function that shuffles the values of the feature f i with a random permutation P with indices {1,2,‥,N}. After getting the shadow features, the D is extended to include the shadow features. Hence, a new feature set ( F i n t e g r a t e d ) is derived by integrating F with Fshadow, as represented in Equation 2. F i n t e g r a t e d ← F , F shadow ← f 1 , f 2 , … f N , f 1 , s h a d o w , f 2 , s h a d o w , … f N , s h a d o w (2)

To the integrated feature set, F i n t e g r a t e d the random forest importance technique is applied to calculate each feature’s importance score(I) using Equation 3. I f i = ∑ t = 1 T Δ f i t Κ t (3)

After obtaining the I for each feature, the threshold value (Ith) is calculated as the maximum score among the shadow features. This can be represented by Equation 4. Equation 5 shows the criteria based on which the Boruta feature selection algorithms identify the important features present in the dataset, thus discarding the others to modify the feature set (F) I t h = Max I f s h a d o w , ∀   f s h a d o w ∈ F s h a d o w (4) F ← I f i ∈ F >   I t h , K e e p   t h e   f e a t u r e I f i ∈ F ≤ I t h , D i s c a r d   t h e   f e a t u r e s (5)

Once Boruta determines if a feature is relevant, the algorithm iteratively confirms or rejects it until all features have been categorized. A refined collection of characteristics expected to improve model generalization and prediction accuracy is the final output, a subset of the relevant features. Boruta provides a strong method for feature selection in the ML process, making it ideal for dealing with high-dimensional datasets (Maurya et al., 2023).

Maximum relevance and minimum redundancy (MRMR) is a filter-based feature selection technique used to select a subset of features from a more extensive set of features. The objective is to select the most useful features while minimizing redundancy. This method is often used in machine learning and data analysis for better model performance, reduced overfitting, and more interpretable models. The IMRMR is the modified version of the MRMR feature section algorithm. Like MRMR, it also selects relevant features with low redundancy scores. MRMR adopts the Mutual Information (MI) to select those features that depend on the target feature. In addition to MI, IMRMR aids Pearson correlation to focus on the linear relation between the features. The following shows the workings of MRMR and IMRMR to determine the difference between these two for selecting features (Ding and Peng, 2005; Yan and Jia, 2019; Zhao et al., 2019).

After applying the IMRMR feature selection, the relevant features remain in the dataset. However, applying the IMRMR does not ensure that the number of selected features has a role in the diagnosis process. Some features will be relevant, but removing them does not impact the diagnosis model. The Salp Swarm Optimization Algorithm (SSA), a wrapper-based feature selection algorithm, is implemented to ensure that the optimal set of relevant features is in the selected feature set, through which the processing time of the developed model can be decreased.

The SSA has been implemented to solve optimization challenges. It mimics the group dynamics of salps, sea animals similar to jellyfish. When applied to an optimization issue, SSA seeks to identify the best possible outcome (Mirjalili et al., 2017; Castelli et al., 2022). There are more than 1.2 million known marine creature species. Most of these species have similar habits and traits, including ways of communication, speed of movement, and foraging strategies. The salp’s habitats are notoriously tough to reach, yet scientists think this behavior aids the animals in movement and feeding. Inspired by the coordinated movement of salps (gelatinous sea organisms), the SSA uses natural selection to find optimal solutions. Collective behavior in these species serves as a paradigm for the optimization difficulties SSA seeks to solve (Ibrahim et al., 2018; Thawkar, 2021; Sayed et al., 2018).

The current work employs ML classifiers such as SVM (Alfian et al., 2022), RF (Alfian et al., 2022), ELM (Ding et al., 2013), AdaBoost (Asselman et al., 2021), and XGBoost (Asselman et al., 2021) to make the initial prediction. Then, the majority voting classifier is used as an ensemble learning approach to make the final prediction. It is a simple and powerful strategy for integrating the results of numerous models into a single prediction. When many models provide varying results and a group choice must be made based on those results, majority voting is an effective tool (Naji et al., 2021; Pati et al., 2023).

The dataset is first subjected to a preprocessing stage to ensure data quality and consistency. This involves several key steps.

• Data Cleaning: Remove or correct any noisy data, missing values, or inconsistencies within the dataset.

The preprocessed dataset is then split into training and testing sets. Two different splitting ratios are considered: 80–20. This helps in evaluating the robustness of the model.

To enhance model performance, feature selection algorithms are applied.

• Boruta Algorithm: An all-relevant feature selection method that identifies relevant features by comparing original attributes with shadow attributes.

The SSA is used to optimize the feature set selected in the previous step.

i. Initiate Population

Based on the training data and the optimized feature set, train the following five models.

• Support Vector Machine (SVM)

Select the top three classifiers from the trained models based on their highest accuracy during training.

Test the ensemble classifier using the testing dataset to obtain evaluative parameters such as accuracy, precision, recall, F1-score, and ROC-AUC.

Algorithm 1

Working of Proposed BIMSSA model.

Input: Dataset D← {D ₁ , D ₂ , D ₃ , D ₄ }, Feature set F← {f ₁ , f ₂ , ……f _n }, Max Iteration, num_salps, Lb, Ub, Γ ₁ , Γ ₂ , BL: {SVM (BL ₁ ), RF (BL ₂ ), ELM (BL ₃ ), AdaBoost (BL ₄ ), XGBoost (BL ₅ ), D ′ is the feature set selected by Boruta, D ″ is the feature selected by IMRMR, D ‴ is the feature selected by SSA}

The performance analysis of the mentioned hybrid model, such as Boruta + IMRMR + SSA + SVM, Boruta + IMRMR + SSA + RF, Boruta + IMRMR + SSA + ELM, Boruta + IMRMR + SSA + AdaBoost, Boruta + IMRMR + SSA + XGBoost for different cancer gene expression datasets is summarized as follows. Figures 3–6 show the performance of the hybrid mentioned above models in contrast to different datasets.

• For the ALL-AML dataset, the Boruta (B)+ IMRMR (IM) +XGBoost with SSA shows the highest accuracy at 0.917. The other parameters, such as PRE, REC, F1-S, F 2, and SPE, are 0.929, 0.929, 0.929, and 0.929, respectively. The other parameters, such as FNR, FPR, MCC, and MCR, are 0.071, 0.100, 0.829, and 0.083, respectively.

The analysis demonstrates that the B + IM + SSA + AdaBoost model, resulting in an accuracy of 0.919, uses the highest obtained Lymphoma dataset. In order to enhance the accuracy even further, it is possible to include an ensemble machine learning approach. The present study uses the majority voting approach to combine predictions from various models and make a final choice based on the majority of votes. This strategy may improve the accuracy by using the advantages of many models while reducing the limitations of each model. By using majority voting, the ultimate model may amalgamate these advantages, resulting in enhanced overall accuracy and more resilient categorization outcomes. The hybrid models mentioned demonstrate remarkable accuracy across several cancer gene expression datasets, reaching a maximum accuracy of 0.919. A majority voting ensemble classifier may help overcome the problems with the provided hybrid models, namely, their high computational cost and the possibility of overfitting caused by combining several algorithms. By integrating several models’ strengths, this method streamlines decision-making, leads to more accurate forecasts with less hyperparameter tweaking, and reduces the danger of overfitting. To further enhance accuracy, the current work integrates the prediction of the best three classifiers using the majority voting technique. This improves overall precision, minimizes mistakes, and results in more dependable and transferable results in cancer categorization assignments.

The performance analysis of the proposed BIMSSA model is shown in Table 5. Table 5 quantifies the ACC, PRE, REC, F-1S, F-2, and SPE measures with Clopper–Pearson confidence interval (CI) (Puza and O’neill, 2006).

• For the ALL-AML dataset, the proposed BIMSSA model shows the accuracy as 0.967 with a CI of 0.887–0.99. The other parameters, such as PRE, REC, F1-S, F 2, and SPE, are 0.967, 0.974, 0.974, 0.974, 0.974, and 0.955, respectively. The other parameters, such as FNR, FPR, MCC, and MCR, are 0.026, 0.046, 0.929, and 0.033, respectively.

The proposed BIMSSA does include the Boruta, IMRMR, and SSA as the feature selection algorithm. In addition, the BIMSSA includes SVM, RF, ELM, AdaBoost, and XGBoost as the base classifiers, including majority voting as the ensemble classifier. The model is evaluated over 4 different high dimensional datasets for evaluation. The computational complexity of the proposed BIMSSA model becomes O (T⋅F⋅log(F)) + O(P⋅F⋅I) + O(N²⋅F), with N << F. Where, N is the number of sample present in the dataset, F is the number of features, T is the number of trees in the RF classifier, I is the number of iterations, and P is the number of population for SSA.

Although the proposed model is efficient, it still retains several limitations that must be mitigated. First is the increased computational complexity based on the iterative structure of the Boruta algorithm, the added overhead of SSA, and the computation-intensive nature of classifiers used, such as SVM and methods based on trees. Then, hybrid approaches for models increase the complexity of their implementation and call for careful coordination between feature selection and classification phases for optimal performance. The pipeline’s repetitive nature might make it seem to extend training time, making it clumsy for situations requiring deployment and results in a very fast manner.

• For Lymphoma dataset the proposed BIMSSA shows an accuracy of 96.2%, BIMSSA again shows strong performance in classifying Lymphoma, indicating the model’s versatility.

• BIMSSA achieves an accuracy of 95.1%, showcasing its strong performance in this dataset as well.

• BIMSSA with an accuracy of 97.1%, performs exceptionally well on the SRBCT dataset, indicating its robustness and effectiveness across different cancer types.

• BIMSSA achieves an accuracy of 96.7%, which demonstrates strong performance and reliability in classifying this cancer type.

BIMSSA proves to be a very efficient and dependable model for categorizing cancer, surpassing the accuracy of other models from the existing literature in all tested datasets except ALL-AML. Though the accuracy is low as compared to Alomari et al. (2021) in case of ALL-AML dataset, still it shows an accuracy of 96.7% that can be considered as a noteworthy performance. The model’s constant and superior performance in classifying many forms of cancer, together with its excellent generalization, sets it out as an exceptional model in the field of cancer classification. Therefore, BIMSSA showcases the flexibility and applicability needed for various cancer datasets. Figure 9 shows the comparison of BIMSSA with other existing models in contrast to diverse datasets.