Predicting Di-2-Ethylhexyl Phthalate Toxicity: Hybrid Integrated Harris Hawks Optimization With Support Vector Machines

Phthalic acid esters (PAEs) are organic pollutants and synthetic compounds and have adverse effects on human health. In this study, we investigated whether Di-2-Ethylhexyl phthalate (DEHP), one of many PAEs, has adverse effects on rats. Adult male Sprague-Dawley rats were treated daily by oral gavage with vehicle (corn oil) or DEHP at a dose of 3000 mg/kg/day for 15 days. The results showed that DEHP caused hepatotoxicity in rats. When compared with the control group, relative liver weights, and serum alanine, aminotransferase levels significantly increased after DEHP exposure. Hepatocyte swelling and degeneration were also found in DEHP-exposed rats. This study proposes an effective intelligence framework for the prediction of DEHP poisoning. The framework is designed by integrating an enhanced Harris hawks optimization (HHO) with a support vector machine (SVM), which is called SGLHHO-SVM. The core characteristic of the developed methodology is the SGLHHO algorithm that integrates the levy mechanism and two core operators abstracted from the salp swarm algorithm and grey wolf optimizer to enhance and restore the search capabilities of the HHO. The presented SGLHHO approach is used to tackle the key parameter pair optimization of the SVM, and it is also utilized to grab the optimal feature subset. Regarding the optimal feature subset and the pair parameter simultaneously, SGLHHO-SVM can autonomously predict the DEHP poisoning. The developed SGLHHO was conducted on 23 benchmark problems and compared with other state-of-the-art and competitive methods. The results demonstrate that the designed SGLHHO performs superior to other competitors on most benchmark problems. Furthermore, the proposed SGLHHO-SVM is also compared with other machine learning algorithms on a real-life DEHP sampled data. Statistical results verify the proposal can show better predictive property and higher stability on all for metrics. Therefore, the SGLHHO-SVM may be served as a potential computer-aided tool for the prediction of DEHP poisoning.

ambient DEHP levels were approximately 0.27 mg per day in the USA. He et al. [2] reported that the concentrations and composition profiles of PAEs in biota, air, and water in an agricultural area of western China. It was found that DEHP concentrations are approximately 2.05-52.5 ng/g in fish, 1.00-9.09 ng/g in pig, 1.83-30.5 ng/g in chicken, 0-6.78 ng/g in cattle and 12.6-75.9 ng/g in vegetables. Moreover, in river water, indoor air, and outdoor air, DEHP is 73.3-261 ng/L, 1550-2300 pg/m3, and 713-2600 pg/m3, respectively. DEHP is easily transferred from plastic products to the environment, with annual usage in the world of more than 10 million tons, and it poses a threat to human health [3]. Exposure to DEHP has been found to be associated with liver, ovaries, mammary gland, testicular, and renal dysfunctions [4], [5]. Zhao et al. [6], [7] reported that DEHP exposure resulted in hepatotoxicity and renal injury in quail (Coturnix japonica) by reducing the expression of the heat shock response and antioxidant defense response. Therefore, the evaluation of DEHP exposure in organisms is still a critical issue.
Although some toxicological studies on DEHP have been done, there is little information on the detection and identification of DEHP exposure in rodents. The use of blood and biochemical measurements coupled with machine learning algorithms has emerged as an ideal approach for the identification of toxicity assessments [8], [9]. Williams et al. [10] developed a machine learning model with a Bayesian algorithm to predict drug-induced liver injury(DILI) and the accuracy up to 86%. He et al. [11] built a combined classifier to screen the hepatotoxic ingredients in Traditional Chinese Medicines by the NaiveBayes algorithm from eight machine learning algorithms (NaiveBayes, AdaboostM1, IBK, KStar, RandomForest, LibSVM, J48, and Bagging). They showed the accuracy was 72.8% and 78.3%, respectively [12]. Hammann et al. [13] present four machine learning models (knearest neighbor, artificial neural networks, decision tree induction, support vector machines) for the prediction of clinically relevant DILI from structure, and the accuracy up to 89%.
In this study, an augmented machine learning method is designed for the prediction of DEHP poisoning for the first time. In the developed methodology (SGLHHO-SVM), the levy mechanism and two core operators abstracted from the salp swarm algorithm and grey wolf optimizer strategy integrated harris hawks optimizer (SGLHHO) was designed to devise a reasonable support vector machine (SVM) for prediction of DEHP poisoning. First of all, the proposal SGL-HHO algorithm was compared with several state-of-the-art competitors on 23 benchmark functions. The trial results verified the ascendant property of the SGLHHO approach. Secondly, the proposed SGLHHO-SVM was also conducted against some other metaheuristic algorithms (MEAs) and other methods for the prediction of DEHP poisoning. The motivation of this study is that design an augmented machine learning method for the prediction of DEHP poisoning, and integrating an enhanced Harris hawks optimization with a support vector machine as developed methodology. In this framework, levy mechanism and two core operators abstracted from the salp swarm algorithm and grey wolf optimizer are all integrated into original harris hawks optimization and then used to tackle the key parameter pair optimization of the SVM and grab the optimal feature subset simultaneously for autonomously predicting the DEHP poisoning.
The experimental results exhibited the apparent advantage of the proposed SGLHHO-SVM. The main contribution of this study is as below: • An improved HHO, SGLHHO, is designed by integrating levy mechanism and two core operators abstracted from the salp swarm algorithm and grey wolf optimizer.
• The proposed SGLHHO has performed significant property on 23 function optimization tasks.
• The proposed SGLHHO successfully solved the parameter tuning for SVM for the first time.
• A potential methodology (SGLHHO -SVM) is treated as an efficient auxiliary tool for DEHP poisoning diagnosis in rats using only blood samples The paper is organized as bellow. A brief synopsis of the SVM, Levy mechanism, and two core operators is given in Section 2. Section 3 gives the proposed SGLHHO algorithm. The proposed SGLHHO-SVM model is shown in section 4. Section 5 describes the data collection and experimental setup. The simulation results of SGLHHO on benchmark functions and SGLHHO-SVM on the DEHP dataset are exhibited in Section 6. Section 7 gives a discussion on the results. The conclusion and future direction are exhibited in Section 8.

A. SUPPORT VECTOR MACHINE (SVM)
SVM is the most widely accepted machine learning based on risk minimization and VC dimensional theories. An excellent tradeoff must be achieved between minimization of the error of the training set and maximizing the margin to obtain the best generalization ability and avoid the phenomenon of overfitting. Owing to its satisfactory training speed and high classification accuracy, especially these small sample data sets, SVM has been applied in many cases [14]- [16].
The SVM can separate the pending samples with highest classification accuracy by hyperplane when a suitable w and b can be obtained. Regarding of kernel techniques, SVM can also solve nonlinear classification. The non-linear function can be modeled as bellow: where K (x, x i ) is the kernel function and K (x, is Gaussian kernel. For more details, one can refer to [17]. VOLUME 8, 2020

B. LEVY MECHANISM AND TWO CORE OPERATORS
The levy mechanism [18] can be defined as follows : where z is the step size and β is the levy index, A/B ∼ N (0, σ 2 ) means the samples originate from a Gaussian distribution, mean and variance are 0 and δ 2 , and (·) indicates the Gamma function. To further reduce the possibility of stagnation, the levy mechanism was adopted to update the positions of the agents, while the β is set to 3/2 in this study.
In addition, another two operators abstracted from salp swarm algorithm [19](SSA) and grey wolf optimizer [20] to further enhance and restore the search capability of the HHO. The core character of SSA is the mathematical model as follow: where x 1 j is the position of the leader salp in the jth dimension, F j means the position of food source in the jth dimension, ub j and lb j are the upper bound and lower bound of jth dimension, c 1 is is defines as equation 6, where l is the current iteration, L is the maximum number of iterations and c 2 , c 3 are random number uniformly generated in [0, 1].
The position of the followers can be updated by equation 7, where i ≥ 2, x i j means the position of ith follower salp in jth dimension, t is the time, v 0 is the initial speed, a = v final v 0 and v = x−x 0 t . For more details, one can refer to [19].
Another core operator is abstracted from the grey wolf optimizer, which is composed of social hierarchy, encircling prey, hunting, attacking prey, and search for prey parts. The core idea is the application of hierarchy. The main mathematical model is as follows : where the D means the distance between grey wolves and the prey, A and C are coefficient vectors by regarding of A = 2 a · r 1 − a and C = 2 · r 2 respectively, X α , X β and X δ are the positions of three closest wolves to current prey, the t indicates the current iteration.For more details, one can refer to [20].
In the first step, the initialization of the population x i (i = 1, 2, 3, · · · , N ) with N agents are randomly generated within the established range [lb, ub], then the fitness of each agent is calculated; finally, the agent who has minimum value is treated as the optimal solution (prey).
During the iteration, the candidate position of each agent in the population will be calculated according to its inherent mechanism of HHO, including exploration phase-based control parameter |E|and q, and exploitation phase-based soft besiege and hard besiege, then if corresponding candidate position is better than its current one, it is updated by candidate one; otherwise, no change will be made.
Secondly, the core characteristic of SSA is performed by the population; The specific process can be shown : For the current population x i (i = 1, 2, 3, · · · , N ) in the proposed methodology, each x i will be calculated F j +c 1 ((ub j −lb j )c 2 + lb j ) or F j − c 1 ((ub j − lb j )c 2 + lb j ) decided by the value of random c 3 , and then update x i regarding of the elite strategy. In addition, c 1 = 2e −(4l/L) 2 can further increase the depth of particle search.
Thirdly, the core idea of hierarchy in the original GWO is also used to evolve the population, and the detailed can be seen in Figure 1. Through this processing mechanism, each agent in the population can learn from three leading wolves to promote the mechanism of the proposed algorithm by this X (t + 1) = X 1 + X 2 + X 3 3 equation. During this process, these three leading wolves are dynamically changing according to their respective fitness values.
Finally, the levy mechanism is adopted to adjust each agent in the population. When the position of ith agent X (t + 1) is updated, and the new candidate solution can be formulated as X l (t + 1) = X (t + 1) + levy(β), at the same the greedy strategy is also used to determine whether the current particle needs to be updated. This structure can effectively reduce the probability of the algorithm falling into local optimum. The detailed algorithm SGLHHO can be seen algorithm 1, and please refer to the original manuscript [53] for detailed formula explanation in this algorithm.

IV. PROPOSED SGLHHO-SVM MODEL
To predict DEHP poisoning, an improved SVM model based on SGLHHO (SGLHHO-SVM) was proposed in this study. The framework of SGLHHO-SVM is exhibited in Figure 2. In addition, the inner 5-fold and outer 10-fold cross-validation (CV) scheme are used simultaneously, which means the K 1 = 5 and K 2 = 10 in this framework. The core part of this model is SVM. In the input space, the radial basis function (RBF) kernel is used to map the aggregate data into a kind of hidden layer space. The process involves two key parameters (penalty coefficient C and kernel width γ ), and n features subset, penalty parameter C determines the tradeoff

Algorithm 1 SGLHHO
Input : The population size N and maximum number of iteration T Output: The best position of the hawks X best and its fitness value fitness best 1 Initialize the random population X i (i = 1, 2, . . . , N ) ; 2 begin 3 Calculate the fitness value of hawks; 4 Set the X rabbit as the location of rabbit (best position); 5 foreach (X i ) do 6 Update the initial energy E 0 = 2rand() − 1 and jump strength J ; if q 0.5 then 10 Update the position using Update the position using if r > 0.5, and |E| 0.5 then 15 Update the position using if r > 0.5 and |E| < 0.5 then 17 Update the position using if r < 0.5, and |E| 0.5 then 19 Soft besiege with progressive rapid dives; 20 if r < 0.5, and |E| < 0.5 then 21 Hard besiege with progressive rapid dives; 22 The population performs the core characteristic of SSA and the hierarchy in original GWO; 23 The levy mechanism is adopted to adjust the population; 24 return X best , fitness best 25 End-Loop; between fitting error minimization and model complexity and the kernel bandwidth γ defines the non-linear mapping from the input space to some high-dimensional feature space. The SGLHHO algorithm evolves these two parameters and the optimal feature subset. In addition, the continuous space was converted through the sigmoid function into the binary space, and if the value is less than 0.5, the feature will be assumed to be chosen; otherwise, the features will be discarded. Finally, SVM makes an accurate diagnose of DEHP poisoning based on the obtained optimal C, γ , and feature subset.

V. EXPERIMENTAL DESIGNS A. DATA COLLECTION
In this study, a total of 133 adult male SD rats weighing between 220 ± 20 g were from the Laboratory Animal Centre (production license: SCXK(Zhe)2015-0001), Wenzhou Medical University, Wenzhou, China. Rats were kept in the animal house facility of the institute in standard polypropylene cages at 23 ± 3 • C temperature with 40-70% relative humidity, 12 h light-dark cycles (light: 07:00-19:00 h, dark: 19:00-07:00 h) with free access to ad libitum food and water. This work had received approval for research ethics from the Administration Committee of Experimental Animals, Laboratory Animal Center, Wenzhou Medical University ethically, and a certificate of approval is available upon request with the inspection form number of wydw 2019-0448. All procedures were conducted in accordance with the guidelines of the Institutional Animal Care. Di-2-Ethylhexyl phthalate (DEHP, CAS no. 117-81-7, purity ≥ 98%) was obtained from Aladdin Industrial Corporation (Shanghai, China). All other solvents and reagents were obtained from commercial sources at the highest grade of purity.
In addition, histopathological examination was also conducted. Liver sections were collected from the experimental rats and immediately fixed in 4% formalin solution for 24 h, dehydrated in 70% ethanol. Liver tissues embedded in a paraffin block, and cut with a microtome to yield sections of 4 µm thickness for histopathological analysis. Sections were stained with hematoxylin-eosin (H&E) and blindly examined for the extent of liver damage under bright field microscope (Nikon Eclipse E200). Table 1 lists the detail of these features.

B. EXPERIMENTAL SETUP
Several other algorithms were involved as competitors, including EHHO [58], OBLHHO [69], HHO, ALO, MFO, WOA, GWO on a common benchmark function. The parameters of these algorithms are set up regarding the original papers, and the details can be seen in table 2. In the next experimental, the liver hypertrophy and unpaired Student t-test or one-way analysis of variance (ANOVA) (Graph-Pad Prism Software, San Diego, CA, USA, version 5.01) were both further investigated thoroughly. Then, the proposed algorithm SGLHHO was used to optimize the best parameter combination and feature subset for SVM, the resultant SGLHHO-SVM used for the prediction of DEHP poisoning on the collection data. SGLHHO-SVM was also compared with several popular machine learning methods, including original SVM, GWO-SVM, WOA-SVM, MFO-SVM, KELM, and KNN. The two key parameters including the C and γ SVM were set as [−2 ∧5 , 2 ∧5 ] and [−2 ∧5 , 2 ∧5 ] respectively. The data were scaled into the range [-1, 1] before the construction of the classifiers, to avoid the instability of the experiment caused by large data.
It should be noted that the simulation experiments were carried out in MATLAB environment ran on a Windows Server2016 R2 operating system with Intel (R) Xeon (R) CPU E5-2660 v3 (2.60 GHz) and 16GB of RAM. To make unbiased and objective results, the classification performance was assessed using the 10-fold cross-validation (CV) analysis. Furthermore, to assess the performance of the SGLHHO-SVM, four commonly used evaluation criteria, including specificity, sensitivity, classification accuracy (ACC), and Matthews Correlation Coefficients (MCC), were employed.  It can be defined as follows: where TP and FN are the numbers of true positive cases and false negatives respectively. True negatives mean cases that are correctly categorized in the negative class.

A. BENCHMARK FUNCTION VALIDATION
In this part, 23 benchmark functions were used to verify the performance of the SGLHHO, which is composed of 7 unimodal cases, 6 multimodal cases, and ten fixed dimensional multimodal cases. These benchmarks have been adopted in many works [70]- [72]. In addition, 30 independent tests were conducted to exclude the influence of random factors.
The presented SGLHHO in this study is compared with several algorithms, including EHHO [58], OBLHHO [69], HHO, ALO, MFO, WOA, GWO. All these methods were conducted on these 23 benchmarks by 30 independent executions. The detailed experimental results are exhibited in table 3, where the mean indicates the average of the results, and STD is the standard deviation. It can be seen that the average results of the presented SGLHHO are the lowest among these benchmarks, and it shows the apparent superiority. Furthermore, in terms of Wilcoxon's test, the SGLHHO is much   better than other peers in 13 of 23 functions in a significant manner. Table 4 exhibiters the sorting results of the Friedman test for SGLHHO versus all other competitors. According to the average ranking value of the involved algorithms, it can be seen that the SGLHHO shows the second-best performance on these benchmark functions, HHO, GWO, EHHO, WOA, OBLHHO, MFO, and ALO show the worse search performance. A reason can account for that the core characteristics of the involved SCA, GWO, and levy in this study can make original HHO obtain a better between exploration and exploitation.
To show the specific property of the SGLHHO, figure 3 records the convergence curves of these algorithms on some benchmark functions. It can be observed that the SGLHHO algorithm perform fast conference ability, and is superior to all other competitors in these functions. It should be noted that the SGLHHO can perform fast convergence search, especially F3, F7, which means it can get the theoretical optimal value quickly, but other competitors need more exploration and need to be accelerated. Furthermore, the same convergence trend can also be observed in F11, F15, F20, F21, F22, and F23. In sum, a conclusion can be made that these VOLUME 8, 2020  involved mechanisms can significantly enhance the property of the original HHO.

B. PREDICTION RESULTS OF DEHP POISONING
To investigate whether DEHP has toxic effects on rat liver, male adult rats were orally administrated 3000 mg/kg DEHP or corn oil as the control for 15 consecutive days. As shown in Figure 4, relative liver weights, the indicator of liver hypertrophy, were significantly increased in the DEHP exposure group compared with the control group. The results revealed that there was no difference in relative kidneys weights between DEHP exposure group and the normal control group. Figure 5 shows the transformation in collagen and tissue structure. There were pathological changes, including inflammation in the liver of the group that received DEHP. Hepatic tissues of DEHP treated rats revealed hydropic degeneration of hepatocytes, cell swelling. However, no significant changes were observed in the liver sections of the vehicle control rats. Histological considerations presented that the control group had a normal structural design of liver tissue as well as a normal monolayer of hepatocytes radially organized.
The blood parameters of the control group and the DEHP group were shown in Table 5. We observed that the levels of ALT and AST were significantly elevated in the DEHP exposure group when compared with the control group (P < 0.05). These results suggest that DEHP exposure induced rat liver injury.
In the next part, the effectiveness of the SGLHHO-SVM method for the prediction of DEHP poisoning was evaluated deeply. Table 6 shows the detailed results. It can be observed from this table that the classification accuracy fulfilled by the SGLHHO-SVM is 96.7%, Matthew's correlation coefficient is 92.88%, sensitivity is 10%, specificity is 97.5%, and their variance is 0.0532, 0.1146, 0 and 0.0791 respectively. Besides, in this experiment, we can observe that the proposed SGLHHO-SVM can automatically obtain the optimal parameters of the SVM model, which is mainly due to the enhanced SGLHHO that can find the optimal parameters effectively.
To further verify the property of this model, the proposed methodology is compared with the other competitors SVM, HHO-SVM, GWO-SVM, WOA-SVM, MFO-SVM, KELM, and KNN. The comparison in terms of the four metrics and standard deviation is shown in figure 6. The results show that the SGLHHO-SVM algorithm shows superior to other competitors in four evaluation metrics, including ACC, MCC, sensitivity, and specificity, and its corresponding standard deviation is also the smallest among all models. This indicates that SGLHHO-SVM performs better performance and stability compared with the original model. In terms of the ACC evaluation metric, the SGLHHO-SVM model is the best. The results of the original SVM and original KELM perform the same property approximately. It is noted that the SVM-based original HHO has the second performance and just below the SGLHHO-SVM. The calculation results of the GWO-SVM and MFO-SVM are very close, and the KNN shows the worst; Regard to the MCC metric, SGLHHO-SVM model, still achieve the best results followed by the WOA-SVM, MFO-SVM, GWO-SVM, original SVM, KELM, and KNN. KNN still has the worst outcomes than other competitors and also the most substantial standard deviation among all these involved approaches. In terms of sensitivity   The improved SGLHHO-SVM model is the best, and the variance is the smallest one. Regarding specificity meristic, the proposed SGLHHO-SVM still performs the best property, not only the mean value but the standard deviation. Besides, in the experiment, we can observe that SGLHHO-SVM can automatically obtain the best property among all these competitive models, which is mainly due to the improved SGLHHO that can find the optimal parameters for SVM automatically. A preliminary conclusion can be drawn that the proposed SGLHHO-SVM shows not only highest diagnostic accuracy but also best stability compared to other algorithms.
In addition, it should be noted that the SGLHHO proposed in this paper is not only used to optimize the two critical parameters of SVM but also used for the optimal feature selection. In the process of feature selection, the 10-fold CV method is also adopted and 10 repetitions of ten-fold CV was used to construct the final averaging results. The selected number of each feature and statistical values via each run of ten-fold CV on the DEHP poisoning database is exhibited in table 1. It can be observed from this table that SGLHHO-SVM shows obvious superiority over other competitors. The features X 3 , X 8 ,X 9 , X 11 , X 14 , X 18 , X 19 , X 20 , and X 21 were selected with values 0.9430, 0.8765, 0.8956, 0.8578, 0.9129, 0.8321, 0.7902, 0.9654, and 0.8931 respectively, while the number of other features were selected with less relatively. However, the other competitors were unsatisfactory on these features. Therefore, a conclusion may be made that these features which appear frequently may perform a superior ability to discriminate DEHP poisoning from healthy individuals than other features with low frequency. Accordingly, the features of ALP, ALT, AST, UA, HGB, MCHC, PLT, LYM% and NEU% should be given more consideration in practical medical cases due to the underlying information contained in these frequency features.

VII. DISCUSSIONS
DEHP is a material widely used as a plasticizer, which can leach from plastics and be available for biological exposure, and thus influence human health. Here, for the first time, we put our focus on the developed methodology (SGLHHO-SVM) to assess DEHP safety. The feature selection results show that ALP, ALT, AST, UA, HGB, MCHC, PLT, LYM%, and NEU% are very important for the prediction of DEHP poisoning in rats.
The liver is the principal detoxifying organ and maintains hepatic metabolic homeostasis; therefore, toxic responses occur more in the liver than in other organs. Liver injury is characterized by elevated levels of serum biomarkers such as ALP, ALT, and AST may suggest hepatotoxicity. In the present study, significantly (P-value < 0.05) higher activities of ALP, ALT, and AST recorded in DEHP exposure group as compared to the control group is suggestive of DEHP-mediated hepatic damage, which was also corroborated by the significantly (P-value < 0.01) higher relative liver weight in DEHP exposure group than in the control group. Increased levels of ALP, ALT, and AST indicate tissue damage by toxicants such as DEHP. Zargar et al. [73] reported that lead exposure caused an increase in ALP, ALT and AST activity, and liver weight in lead acetate-intoxicated Wistar rats. Among all commonly measured enzymes in the liver function tests, ALP is used as a marker enzyme for liver function and integrity. The increased levels of serum ALP activity may be attributed to disturbances in hepatic secretory activity or the transport of metabolites or may be due to cellular leakage. He et al. [74] found evidence of an inverse association between vitamin D deficiency and higher ALP levels in individuals with viral hepatitis.
As an organic metabolite with dual effects of antioxidation and pro-oxidation, UA is an effective antioxidant in the extracellular environment and protects the erythrocyte membrane by scavenging oxygen radicals. Elevated levels of Urea, UA, and Crea have been reported of hyperuricemia induced endothelial dysfunction in rats [75]. In the present study, decreased (P-value < 0.01) Crea levels in the plasma suggested that DEHP had impaired renal function; however, an interesting phenomenon was reported, there was no difference in UA between DEHP exposure group and the normal control group.
Compared with the normal control group, lower values of hematological parameters such as HGB, HCT, and RBC were observed in the DEHP exposure group (P-value < 0.05). The results of the red blood cell parameters (HGB, HCT, and RBC) were significantly reduced in the DEHP exposure group may reveal disturbances of peripheral phagocytosis and red blood cell apoptosis [76], [77]. There was no statistical difference (P > 0.05) in MCHC, PLT, LYM%, and NEU% between the DEHP exposure group and the normal control group.
However, even the proposed SGLHHO-SVM has performed competitive, potential, and progressive results in comparison with other algorithms in this study, an reality that could be recognized is that some rooms to be further improved still existed. Firstly, the experimental samples is a little less, which needs further collection. It is expected to cooperate with more specialized hospitals in clinical research, so as to increase the number of experimental samples. Secondly, the design of the new algorithm may face more complex computing resource consumption due to the meta heuristic algorithm is a stochastic algorithm intrinsically. Parallel programming for the designed SGLHHO-SVM may be a very worthwhile work. Furthermore, the proposal should also be attempted to apply more medical or other poisoning cases although it performs well on Di-2-ethylhexyl phthalate toxicity in this study.

VIII. CONCLUSION AND FUTURE WORK
In this study, an efficient auxiliary tool SGLHHO-SVM for DEHP poisoning diagnosis in rats using only blood samples is constructed. The main novelty of the proposed methodology is that the developed SGLHHO, which integrates the levy mechanism and two core operators abstracted from the salp swarm algorithm and grey wolf optimizer to enhance and restore the search capability of the original HHO. The effectiveness of SGLHHO was strictly verified on 23 benchmark functions compared with several other competitors. The experimental results have illustrated that the proposed SGLHHO can perform a much better property than other peers on these function optimization. Moreover, the proposed SGLHHO was also employed for searching for the optimal parameters and the feature subset for SVM; the resultant SGLHHO-SVM has been used successfully for predicting DEHP poisoning. A rigorous comparison was also executed between the SGLHHO-SVM and other competitive algorithms. Furthermore, the results have also verified that the SGLHHO-SVM has performed higher prediction accuracy with the more stable property.
For future work, several aspects still need to be investigated further. It should be noted that more influencing factors and coefficients are introduced, and so the parallel computing can also be introduced to reduce the computational burden in the specific application process. In addition, more data samples can be collected to construct a more effective and reliable framework. Moreover, the SGLHHO-SVM can be used to predict more diseases to expand the application scope of this algorithm, such as clustering and CT image segmentation.