The prediction of the optimum compositions of a parenteral nanoemulsion system loaded with a low water solubility drug for the treatment of schizophrenia by artificial neural networks

Abstract

Aripiprazole was encapsulated in a palm kernel oil esters nanoemulsion for the purpose of brain delivery via intravenous administration. High shear and high pressure homogenizers were applied for formulating the low water solubility drug in the nanoemulsion system, which was stabilized using different emulsifiers; lecithin, Tween 80 and glycerol. The artificial neural networks (ANNs) modeling of the nanoemulsion formulation was carried out to achieve the minimum particle size. The effects of the amount of palm kernel oil ester (PKOE) (3–6%, w/w), lecithin (2–3%, w/w), Tween 80 (0.5–1%, w/w), glycerol (1.5–3%, w/w), and water (87–93%, w/w) on the particle size were considered as inputs to the network. The particle size of the samples with various compositions was measured as an output. To obtain the optimum topologies, the ANNs were trained using Incremental Back Propagation (IBP), Genetic Algorithm (GA), Batch Back Propagation (BBP), Quick Propagation (QP), and Levenberg–Marquardt (LM) algorithms for the testing data set. The topologies were determined by the indicator of the minimized root mean squared error (RMSE) for each algorithm. According to the results, QP-5-4-1, GA-5-12-1, IBP-5-11-1, BBP-5-10-1, and LM-5-9-1 were selected as the optimized topologies. It was found that the optimal algorithm and topology were the quick propagation and the configuration with 5 input, 4 hidden and 1 output nodes. Conclusively, ANN models were developed for the prediction of the particle size of nanoemulsions loaded with aripiprazole and its stable nanoemulsion system, which could be used effectively for intravenous administration.

---

Introduction

Schizophrenia is a serious, chronic and debilitating mental illness for which most patients require long term treatment with antipsychotic treatment.¹ This mental illness is characterized by positive symptoms (e.g., hallucinations, delusions and deranged thoughts), negative symptoms (e.g., loss of motivation, restricted emotional experience, poverty of speech) and cognitive impairment. The first hypothesis formulated that could explain the pathophysiology of this mental illness, is namely the “dopamine theory”, which attributed a prominent role to dopamine system dysregulation. In recent years, a medication (aripiprazole) has received a growing attention towards the treatment of schizophrenia.^2–5

Schizophrenia disease is mostly related to the Central Nervous System (CNS) in the body. All organisms with a well-developed CNS have a blood–brain barrier (BBB).⁶ The BBB acts as a barrier to prevent any macromolecules from entering the brain.⁷

Aripiprazole is regarded as a third generation antipsychotic drug (APD) with excellent therapeutic efficacy in controlling the symptoms of schizophrenia, a low incidence of extra-pyramidal side effects (EPS) and weight gain side effects.^8–11

Aripiprazole possesses poorly-water soluble characteristics in the form of its salt. To ensure that it can reach the target cell effectively, a higher dosage of aripiprazole is needed. However, in some clinical cases, a higher dosage of aripiprazole can cause some side effects.¹² A nanoemulsion-based aripiprazole carrier could improve the solubility of the drug in the dispersed phase, drug penetration through the blood–brain barrier (BBB) and target cells due to its extremely small size. Thus, a smaller dosage of aripiprazole is preferred to reduce its side effects. For enhancing the poorly-water soluble molecule's dissolution rate, the surface area is purposely increased by decreasing the particle size of the drug molecule. The performance can extremely affected by increasing the surface area.¹³ In this context, the development of new drug nanodelivery systems to increase drug bioavailability and reduce adverse effects has been claimed as a good option.

An emulsion consists of a mixture of two immiscible liquids, which are dispersed throughout each other. Water and oils is considered as its basic components, which requires a surfactant to decrease the interfacial tension and maintain the stability of the emulsion. An emulsion can be prepared in the form of oil-in-water or water-in-oil.¹⁴ Palm kernel oil esters (PKOEs) consist of a higher amount of short chain esters when compared to other oils. In the pharmaceutical industry, palm kernel oil esters are used in formulations because they can be a good carrier for active components. Due to its lower slip melting point and higher saponification value, palm kernel oil esters have unique properties for nanoemulsions and can be used in many applications such as cosmetics¹⁵ and pharmaceuticals.¹⁶

A nanoemulsion is thermodynamically stable transparent (or translucent) system of oil, water, and surfactants, having a droplet size usually in the range of 20–200 nm.^17,18 It has been proven that a very small particle size of emulsion can provide effective encapsulation for delivery in the human body.¹⁹ Nanoemulsions present good drug delivery and parenteral delivery due to their small nanometer size particles, biocompatibility, relative stability, ability to solubilize high quantities of hydrophobic compounds, ability to reduce the toxicity of drugs, and ability to protect drugs from hydrolysis and enzymatic degradation under physiological conditions.²⁰

Nanoemulsions are classified as non-equilibrium systems. To form the nanoemulsion, energy input from a mechanical device or from the chemical potential of the components is needed.²¹ There are two emulsification methods that have been used to make nanoemulsions. The first emulsification method is called the dispersion or high energy emulsification method, which uses mechanical energy (such as high-pressure homogenization).²² This method is widely used in industry to produce small and uniform droplet size of emulsions. The second emulsification method is the low energy emulsification technique. This method consists of the stepwise addition of one component to a mixture of the other components at a constant temperature.²³

In recent years, artificial neural networks (ANNs) were introduced in pharmaceutical applications as an effective tool to solve complex multivariate non-linear relationships.^24,25 ANNs are artificial neurons used to simulate the method in which biological neurons process information.²⁶ ANNs consist of an interconnected group of neurons in the modeling process, which are used to give predictions on the behavior of a given system, designing a new process and analyzing existing processes. Moreover, this process can be carried out in a short period computing with high potential for adaptive performance of adequate quality.^27–30 In pharmaceutical research, ANNs have been successfully used in analysis and modeling applications such as forming the controlled release drug delivery systems, enhancing the understanding of the formation nanoemulsions and the evaluation of the stability of nanoemulsions.³¹

In this study, the optimization of the composition of nanoemulsions containing aripiprazole with respect to the amount of oil, lecithin, Tween 80, glycerol and water was carried out. The response, which is the particle size, was studied to find the best model in the ANNs.

Methodology

Materials

Palm kernel oil esters (PKOEs) were synthesized in our laboratory via the enzymatic transesterification of palm kernel oil and oleyl alcohol.¹⁵ Pure soy bean lecithin (Lipoid S75) was purchased from Lipoid GmbH, Ludwigshafen, Germany. Glycerol was purchased from JT Baker, USA. Polysorbate 80 (Tween 80) was obtained from Fluka, Sigma-Aldrich Chemie GmbH, Germany. Aripiprazole was purchased from Laboratory & Scientific Enterprise, Malaysia. Water was deionized using a Milli-Q filtration system, Milipore, USA.

Determination of the solubility of aripiprazole in oil

The solubility of aripiprazole in the PKOEs was determined. Different amounts of drug were added into the oil containing lecithin (3%). The solutions were kept under moderate magnetic stirring for 24 h to reach equilibrium. The samples were then centrifuged at 4500 rpm for 15 min. The best amount of drug was observed to be 0.1% in the composition formulation.

Preparation of the emulsion formulation using low shear rate emulsification

Emulsions were prepared via low shear rate stirring emulsification using an overhead stirrer (IKA® RW20 Digital, Nara, Japan) at 300–305 rpm. Aripiprazole (0.1%) was dissolved in the oil phase, which was comprised PKOEs (3.0–6.0%) containing lecithin (2.0–3.0%) as the surfactant. Tween 80 (0.5–1.0%) was then added into the oil phase as a co-surfactant after the aripiprazole was completely dissolved. The oil phase was added dropwise into an aqueous phase consisting of glycerol and was continuously stirred to form a coarse emulsion. The mixing of the emulsions was carried out with a shear rate of 300–305 rpm for 3 h.

Preparation of the nanoemulsion using high shear and high pressure homogenization

The emulsion prepared using low shear rate emulsification was homogenized with a high shear homogenizer (Kinematic Switzerland) at high speed (3500 rpm) for 15 minutes. The samples were further homogenized using a high-pressure homogenizer 1000 psi for 14 cycles. The final products were placed into sample bottles.

Stability study

The freshly prepared samples were placed in a container and then stored in the refrigerator at ±5 °C for nine months. To test the stability of the samples, the samples were centrifuged at 4500 rpm for 15 min. The samples were then observed to observe if there any precipitation had occurred.

Particle size measurement

The particle size distribution was measured by a diffusion method using a dynamic light scattering (DLS) particle analyzer (Zetasizer Nano ZS, Malvern Instruments, Malvern, UK). The size distribution using the diffusion of scattered laser light by the particles was measured. The measurement was conducted using the Photon Correlation Spectroscopy (PCS) principle.^32,33

Transmission electron microscopy (TEM)

The particle size was also measured using transmission electron microscopy (Hitachi H7100, Japan). A formvar coated copper grid was placed on top of a drop of the sample and left at room temperature (25 ± 0.5 °C) for 5 minutes. The filled copper grid was stained for 2 minutes using 2% phosphotungstic acid and air dried prior to TEM analysis.

Experimental design

The modeling and optimization of the nanoemulsions containing aripiprazole were carried out using NeuralPower software version 2.5.^28,29 In a mixture design wherein the composition was the factor of interest, the levels cannot be chosen arbitrarily. All fractions of the components must sum to unity.³⁴ As Table 1 shows, a total of 27 experimental points have been randomly divided into two data sets for training (20 experiments) and testing (4 experiments), and a validation set (3 experiments). The software facilitated the option of randomization. The training and testing data sets were used to compare and ensure the robustness of the network parameters, respectively. Moreover, the testing set was utilized to avoid over fitting by controlling errors.³⁵

Table 1 The experimental design that consists of the training and testing data sets

Run no.	PKOEs (%)	Lecithin (%)	Tween 80 (%)	Glycerol (%)	Water (%)	Particle size (nm)
						Actual	Predicted
Training set
1	6.00	3.00	0.50	1.50	89.00	124.97	127.67
2	6.00	2.00	1.00	3.00	88.00	87.62	87.62
3	4.50	2.00	0.75	2.25	90.50	86.55	86.56
4	5.25	2.50	0.63	1.88	89.75	103.43	103.44
5	6.00	2.00	1.00	1.50	89.50	86.18	88.34
6	3.00	2.50	0.50	1.50	92.50	81.35	81.35
7	3.00	2.00	0.50	3.00	91.50	84.43	84.43
8	6.00	3.00	0.50	1.50	89.00	130.37	127.67
9	6.00	2.00	1.00	1.50	89.50	90.48	88.34
10	3.00	2.00	1.00	2.25	91.75	66.20	64.22
11	6.00	3.00	1.00	2.25	87.75	89.59	89.59
12	3.00	3.00	0.50	2.25	91.25	81.39	81.39
13	3.00	2.00	0.75	1.50	92.75	71.56	71.56
14	6.00	2.00	0.50	2.25	89.25	124.57	124.57
15	3.00	3.00	1.00	1.50	91.50	65.25	65.25
16	5.25	2.75	0.88	2.63	88.50	89.07	89.07
17	4.50	2.50	0.75	3.00	89.25	87.12	87.12
18	4.50	3.00	1.00	1.50	90.00	80.26	80.26
19	3.00	3.00	1.00	3.00	90.00	65.55	65.55
20	4.50	2.00	0.50	1.50	91.50	106.13	106.14
Testing set
1	5.50	2.50	0.50	1.50	89.25	120.40	121.78
2	4.00	2.50	1.00	1.50	90.00	72.66	73.74
3	4.50	2.00	0.75	2.50	91.00	92.53	87.02
4	4.25	2.75	0.60	2.10	90.25	92.64	93.78

---

The ANN description

Artificial Neural Networks (ANNs) are computer programs, which are used for the simulation of some roles in the human brain using dissimilar learning algorithms. ANNs possess the extraordinary information processing attributes of the human brain, such as non-linearity, high parallelism, robustness, fault and failure tolerance, learning, capability to manage imprecise and distorted information, and the ability to generalize. ANNs, are able to manage multiple independent and dependent variables simultaneously in one model.³⁶ The effectiveness of the ANN model is shown when modeling non-linear relationships between dependent and independent variables using an approach that is similar to a ‘black box’.³⁷ This is because the performance of the ANN model is not needed for the end user to have extended understanding. The ANN model can predict and formulate optimization capabilities and also be updated with new raw data. After the models are trained, the response to new experimental conditions can be predicted using the ANN model.³¹

ANNs that contain input, hidden and output layers provide the mathematical free functionalization of a complicated practical process.³⁸ The layers, which consist of several nodes, are connected by multilayer normal feed-forward or feed-back connection formula.³⁹ The hidden layer could be more than one parallel layer; however, the single hidden layer is universally suggested. The connection is that the nodes of a particular layer are connected to the nodes of the next layer. The nodes are simple artificial neurons, which simulate the behavior of biological neural networks.⁴⁰ The nodes in the input layer are qualified by sending data via special weights to the nodes of the hidden layer and then to the output layer.^39,41 The qualification is carried out by associated weights during the learning process via well-known learning algorithms.

The learning process

In the learning process, the weights are calculated by the weighted summation of the received data from the former layer and transferred to the next layer.⁴² The number of hidden nodes is obtained by a trial and error training calculation, which is examined from one to n nodes. In the process, the output of the hidden nodes in turn, acts as input to final (output) layer's nodes, which undergo a similar or different transformation. The universal learning algorithms are QP, IBP, GA, BBP and LM whereas the multilayer is the nodes' connection type.⁴³ The usual transfer function is the logarithmic sigmoid for both hidden and output layers that is bounded from (0–1).^29,44 The sigmoid bounded area is used to normalize the input and output data that is provided by the software scaling. The scaled data are passed into the first layer, propagated to the hidden layer and finally meet the output layer of the network. Each node in the hidden layer or in the output layer firstly acts as a summing junction, which modifies the inputs from the previous layer using the equation as follows:where y_i is the input of the network to j node in the hidden layer, i is number of nodes, x_i is the output of previous layer, while w_ij is the weight of the connection between the ith node and jth node. The bias associates with node j and is presented by b_j. The main aim of the process is to find the weights for minimizing the error of the RMSE, which is obtained from difference between the network prediction and actual responses.where n is number of the points, y_i is the predicted values and y_di is the actual values. Therefore, the learning process with an algorithm is continued until finding the minimum RMSE, which is called the topology. To avoid random correlation due to the random initialization of the weights, learning of a topology is repeated several times. As a result, the topology with the lowest RMSE is selected to compare with other nodes' topologies.⁴⁵ Therefore, the topologies for n numbers of hidden layers for the considered algorithms are obtained in same way. Finally the topologies of the algorithms were compared to select the provisional model by the maximum R² (eqn (3)), minimum RMSE and average absolute deviation (AAD) (eqn (4)),where n is the number of points, y_i is the predicted value, y_di is the actual value and y_m is the average of the actual values.

Results and discussion

The topologies of the algorithms

The hidden layers are in between the input and output layers providing a link between the input and output layers. The structure of the hidden layer was constructed by examining a series of topologies with varied node number from 1 to 15 for each algorithm, which were the QP, IBP, BBP, GA, and LM algorithms. To determine the minimum value of the RMSE as an error function, the model learning was performed for the testing data set. However, to obtain the best model, it was carried out using 10 repetitions for each node to avoid random correlations due to the random initialization of the weight.

To determine the optimized topology for each algorithm, the training was performed using the same steps for the QP, IBP, BBP, GA and LM algorithms. The minimum value of the RMSE was chosen among the 10 learning repetition data for each node and plotted in graphs of the minimum value of RMSE versus the nodes of the algorithms' hidden layer, which are shown in Fig. 1.

Fig. 1 The selected RMSE vs. node number for the composition of the nanoemulsion containing an aripiprazole network's hidden layer for QP, IBP, BBP, GA and LM.

In a comparison of the best topology, one node of the 15 topologies in each algorithm was selected as the lowest RMSE. The selected topologies were 5-4-1, 5-11-1, 5-12-1, 5-10-1 and 5-9-1 for the QP, IBP, GA, BBP and LM algorithms, respectively. As shown in Fig. 2, the topology of QP-5-4-1 presented the lowest RMSE among the other topologies chosen as the provisional model for the nanoemulsion containing aripiprazole composition.

Fig. 2 The scatter plots of the predicted particle size versus actual particle size for the testing data set that shows the performed R² of the optimized topologies, QP-5-4-1, BBP-5-10-1, IBP-5-11-1, GA-5-12-1 and LM-5-9-1.

The model selection

To train the neural networks in the multivariate optimization of reaction, a gradient descent back-propagation algorithm in the QP version was applied. The experimental data of the mixed experimental design were divided into two sets: 20 of the data experiments were used as the training set, 4 of the data experiments were used as the test set and the remaining 3 data experiments used as the validation set (Table 1).

The ANNs modeled the raw data and the best model created to provide better quality predictions for the particle size of the nanoemulsions. The RMSE, R², and AAD of the model for the training and testing sets for the particle size are shown in Table 2. To select the final model for the particle size of the nanoemulsions, the values of RMSE, R² and AAD were studied for the topologies of QP-5-4-1, BBP-5-10-1, IBP-5-11-1, GA-5-12-1 and LM-5-9-1. The best value of the RMSE was chosen based on 10 repeated runs. As shown in Table 2, QP had the lowest of the RMSE values, which was 2.614 in comparison with the other algorithms. Therefore, the performance of QP with the 5-4-1 topology was more effective than the IBP, BBP, GA and LM algorithms.

Table 2 The performance results for the optimized topologies

Learning algorithms	Architecture	Training data			Testing data
		RMSE	R^2	AAD	RMSE	R^2	AAD
GA	5-12-1	4.639	0.940	4.114	2.822	0.993	2.631
LM	5-9-1	1.228	0.996	0.728	2.801	0.982	2.350
BBP	5-10-1	1.244	0.996	0.744	2.712	0.979	2.182
IBP	5-11-1	1.327	0.996	1.095	2.706	0.984	2.203
QP	5-4-1	1.228	0.996	0.730	2.614	0.972	2.402

---

As the scatter plots depicted, in the comparison with the other topologies, the QP-5-4-1 topology presented an R² value of 0.972, which was the best performance with a minimum RMSE. Then, the performed results of the topologies were used to calculate the RMSE, R² and AAD. To calculate the R² value, the prediction of the topologies and actual values of the particle size were plotted for the testing and training data sets, as shown in Fig. 2 and 3, respectively.

Fig. 3 The scatter plots of the predicted particle size versus actual particle size for the training data set that shows the performed R² of the optimized topologies, QP-5-4-1, BBP-5-10-1, IBP-5-11-1, GA-5-12-1 and LM-5-9-1.

The QP-5-4-1 network

Fig. 4 shows a schematic of a multilayer perceptron feed-forward network of ANNs based on QP consisting of 5 inputs, one hidden layer with 4 nodes and one output. It illustrates the structure of the QP-5-4-1 topology as a final model for nanoemulsions containing aripiprazole. The input layer with 5 nodes (PKOEs, lecithin, Tween 80, glycerol and water) is the distributor for the hidden layer with 4 nodes, which were determined by the learning process. The input data of the hidden nodes were calculated by weighted summation (eqn (5)). Then, the output data of the hidden layer were transferred to the output layer (particle size) using a log-sigmoid function (eqn (6)).where S is the summation, b is a bias, I_i is the ith input to the hidden neuron and W_i is the weight associated with I_i. The bias shifts the space of the non-linear properties.where f(x) is the hidden output neuron. As a result, the QP-5-4-1 topology was the best result and used to determine the optimum and importance values of the input variables of the composition of nanoemulsions containing aripiprazole to obtain the smallest particle size.

Fig. 4 The network architecture (5-4-1) of the multilayer normal feed-forward connection type for the Quick Propagation algorithm, which consists of 5, 4 and 1 nodes in the input, hidden and output layer, respectively.

Model verification

Table 3 shows that three random formulations were prepared to validate the final model. Verification of the model was carried out to examine the adequacy of the predicted particle size. The results indicated the good agreement between actual and predicted values.

Table 3 The validation set for the nanoemulsion containing aripiprazole

Run no.	PKOEs (%)	Lecithin (%)	Tween 80 (%)	Glycerol (%)	Water (%)	Particle size (nm)
						Actual	Predicted
1	3.75	2.00	0.75	2.50	91.00	76.73	78.76
2	4.25	2.25	0.50	2.25	90.75	96.18	98.49
3	4.50	2.50	0.60	2.10	90.30	98.04	96.13

---

Optimization of the compositions

Table 4 suggests the optimum formulation of nanoemulsions containing aripiprazole using the final ANN model. The smallest particle size (62.23 nm) was obtained using a nanoemulsion composition of 3% PKOEs, 2% lecithin, 1% Tween 80, 2.25% glycerol, and 91.75% water.

Table 4 The optimum compositions derived from the ANN based on QP model for the particle size of the nanoemulsion

Methods	Optimal composition					Particle size (nm)
	PKOEs (%)	Lecithin (%)	Tween 80 (%)	Glycerol (%)	Water (%)	Actual	Predicted	RSE%
ANN (QP)	3.00	2.00	1.00	2.25	91.75	62.23	64.38	3.34

---

Importance of the effective variables

Fig. 5 shows that the importance percentage of the input variables on the particle size of the nanoemulsions. The Tween 80 content at 29.45% is the most important factor controlling the particle size, followed by water at 19.34%, glycerol at 18.57%, lecithin at 17.39%, and PKOE at 15.25%. The presence of Tween 80 appeared to be most influential on the particle size. On the other hand, the effects of other variables such as water, glycerol, lecithin, and palm kernel oil esters were very strong on the particle size. As a result, none of the variables is neglectable in this study.

Fig. 5 The relative importance of the particle size of the nanoemulsion containing aripiprazole input variables of Tween 80, water, glycerol, lecithin and PKOEs.

Morphological studies

Because the range of size normally expands over the capacity of any advanced instrument, the measurement of particle size must be taken at least with two corresponding techniques. Dynamic light scattering and transmission electron microscopy are the most recommended tools to measure the particle size, which is lower than 1 μm. Transmission electron microscopy was implemented to confirm the DLS data and to gain more information on the shape and size of the nanoemulsion oil globules. Fig. 6 shows the particle size of the nanoemulsion containing aripiprazole as determined by TEM. It was found that the average of particle size is smaller than 100 nm (between 65 nm and 80 nm) with almost spherical shape oil globules. It is observed that the particle shape is in accordance with the results measured using dynamic light scattering (DLS).

Fig. 6 TEM photomicrographs of the freshly prepared aripiprazole loaded nanoemulsion. The scale bar represents 200 nm.

Conclusions

The compositions of the nanoemulsions containing aripiprazole, such as PKOEs, Tween 80, lecithin, glycerol, and water as effective variables, were modeled at different compositions using ANNs to define the desirable particle size of the nanoemulsions containing aripiprazole. To obtain the qualified network, five different algorithms, including QP, IBP, BBP, GA and LM were learned using training and testing data sets. The results of the learning program were 5 topologies: QP-5-4-1, BBP-5-10-1, IBP-5-11-1, GA-5-12-1 and LM-5-9-1. The performance of the topologies was optimized using RMSE, AAD and R². The topology (QP-5-4-1) with the lowest RMSE was selected as the provisional network of the composition of the nanoemulsion containing aripiprazole. The importance of the variables included the Tween 80 content at 29.45% as the most important factor controlling the particle size, followed by water at 19.34%, glycerol at 18.57%, lecithin at 17.39%, and PKOE at 15.25%, which shows none of the variables was neglectable in this study. As a conclusion, the ANN is an efficient quantitative tool that is able to model the effective input variables used to predict the desirable particle size of a nanoemulsion containing aripiprazole.

Conflict of interest

The authors have declared no conflict of interest.

Acknowledgements

The author would like to express acknowledgment to University Putra Malaysia for granting this project under Research University Grant Scheme (RUGS).

Notes and references

1. M. Wilson, J. E. Helzer, H.-G. Hwu, E.-K. Yeh, M. McEvoy and J. E. Tipp, Am. J. Psychiatry, 1991, 148, 1697–1704.
2. H. Li, J. Luo, C. Wang, S. Xie, X. Xu, X. Wang, W. Yu, N. Gu and J. M. Kane, Schizophr. Res. Treat., 2014, 157, 112–119.
3. S. G. Potkin, A. R. Saha, M. J. Kujawa, W. H. Carson, M. Ali, E. Stock, J. Stringfellow, G. Ingenito and S. R. Marder, Arch. Gen. Psychiatry, 2003, 60, 681–690.
4. J. P. McEvoy, D. G. Daniel, W. H. Carson, R. D. McQuade and R. N. Marcus, J. Psychiatr. Res., 2007, 41, 895–905.
5. H.-Y. Chan, W.-W. Lin, S.-K. Lin, T.-J. Hwang, T.-P. T. Su, S.-C. Chiang and H.-G. Hwu, J. Clin. Psychiatry, 2007, 68, 29–36.
6. N. J. Abbott, Cell. Mol. Neurobiol., 2005, 25, 5–23.
7. N. J. Abbott, A. A. Patabendige, D. E. Dolman, S. R. Yusof and D. J. Begley, Neurobiol. Dis., 2010, 37, 13–25.
8. R. B. Mailman and V. Murthy, Curr. Pharm. Des., 2010, 16, 488.
9. M. Wood and C. Reavill, Expert Opin. Invest. Drugs, 2007, 16(6), 771–775.
10. E. Stip and V. Tourjman, Clin. Ther., 2010, 32, S3–S20.
11. J. Bhattacharjee and H. G. El-Sayeh, Aripiprazole versus typical antipsychotic drugs for schizophrenia, The Cochrane Library, 2008.
12. B. Chavez and R. A. Poveda, Ann. Pharmacother., 2006, 40, 2265–2268.
13. E. Merisko-Liversidge and G. G. Liversidge, Adv. Drug Delivery Rev., 2011, 63, 427–440.
14. F. Shakeel and W. Ramadan, Colloids Surf., B, 2010, 75, 356–362.
15. P. Keng, M. Basri, M. Zakaria, M. A. Rahman, A. Ariff, R. A. Rahman and A. Salleh, Ind. Crops Prod., 2009, 29, 37–44.
16. N. Salim, M. Basri, M. B. Rahman, D. K. Abdullah and H. Basri, Int. J. Nanomed., 2012, 7, 4739.
17. H. Nakajima, S. Tomomasa and M. Okabe, In Proceedings of First Emulsion Conference, 1993, Paris, pp. 1–11.
18. H. Nakajima, Surfactant Sci. Ser., 1997, 66, 175–197.
19. C. Solans, P. Izquierdo, J. Nolla, N. Azemar and M. Garcia-Celma, Curr. Opin. Colloid Interface Sci., 2005, 10, 102–110.
20. S. M. Jafari, Y. He and B. Bhandari, Food Res. Int., 2007, 40, 862–873.
21. P. Walstra, Chem. Eng. Sci., 1993, 48, 333–349.
22. S.-Y. Lee, S.-J. Chun, I.-A. Kang and J.-Y. Park, J. Ind. Eng. Chem., 2009, 15, 50–55.
23. N. Anton, J.-P. Benoit and P. Saulnier, J. Controlled Release, 2008, 128, 185–199.
24. M. Rezaee, M. Basri, R. N. Z. R. A. Rahman, A. B. Salleh, N. Chaibakhsh and H. R. F. Masoumi, Ind. Crops Prod., 2014, 52, 506–511.
25. A. Amani, P. York, H. Chrystyn and B. J. Clark, Pharm. Res., 2010, 27, 37–45.
26. E. Akbari, Z. Buntat, A. Enzevaee, S. J. Mirazimiabarghouei, M. Bahadoran, A. Shahidi and A. Nikoukar, RSC Adv., 2014, 4, 36896–36904.
27. H. R. F. Masoumi, A. Kassim, M. Basri, D. K. Abdullah and M. J. Haron, Molecules, 2011, 16, 5538–5549.
28. Y. Abdollahi, N. A. Sairi, M. K. Aroua, H. R. F. Masoumi, H. Jahangirian and Y. Alias, J. Ind. Eng. Chem., 2014, 25, 168–175.
29. Y. Abdollahi, A. Zakaria, M. Abbasiyannejad, H. R. F. Masoumi, M. G. Moghaddam, K. A. Matori, H. Jahangirian and A. Keshavarzi, Chem. Cent. J., 2013, 7, 96.
30. Y. Abdollahi, S. B. M. Said, N. A. Sairi, M. F. B. M. Sabri, A. Zakaria, E. Abouzari-lotf and M. Dorraj, J. Ind. Eng. Chem., 2015, 29, 400–407.
31. Y. Sun, Y. Peng, Y. Chen and A. J. Shukla, Adv. Drug Delivery Rev., 2003, 55, 1201–1215.
32. H. Geers and W. Witt, in Direct calculation of the volume based particle size distribution from PCS or PCCS measurements, Proceedings of the Particulate Systems Analysis, Stratford-upon-Avon, UK, 2008.
33. W. Witt, H. Geers and L. Aberle, Particulate Systems Analysis, Harrogate, UK, 2003.
34. R. D. Snee, J. Qual. Tech., 1971, 3, 159–169.
35. M. G. Moghaddam, F. B. H. Ahmad, M. Basri and M. B. A. Rahman, Electron. J. Biotechnol., 2010, 13, 3–4.
36. M. Dorraj, Y. Abdollahi, S. B. Mohd Said, M. F. Bin Mohd Sabri, N. A. Sairi, W. P. Meng and E. Abouzari-lotf, RSC Adv., 2015, 5, 21384–21395.
37. S. Askari, R. Halladj and M. J. Azarhoosh, RSC Adv., 2015, 5, 52788–52800.
38. M. Ghaedi, A. Ghaedi, E. Negintaji, A. Ansari and F. Mohammadi, J. Ind. Eng. Chem., 2014, 20, 4332–4343.
39. A. Ghaffari, H. Abdollahi, M. Khoshayand, I. S. Bozchalooi, A. Dadgar and M. Rafiee-Tehrani, Int. J. Pharm., 2006, 327, 126–138.
40. P. Shabanzadeh, R. Yusof and K. Shameli, J. Ind. Eng. Chem., 2015, 24, 42–50.
41. D. G. Garson, AI Expert, 1991, 6, 47–51.
42. M. Khare and S. S. Nagendra, Artificial neural networks in vehicular pollution modelling, Springer, 2006.
43. D. Salari, N. Daneshvar, F. Aghazadeh and A. Khataee, J. Hazard. Mater., 2005, 125, 205–210.
44. P. Kaur, V. K. Sangal and J. P. Kushwaha, RSC Adv., 2015, 5, 34663–34671.
45. M. Ghaedi, S. Hajati, M. Zare and S. Y. Shajaripour Jaberi, RSC Adv., 2015, 5, 38939–38947.

---