Table of contents
Abstract   Introduction   Materials and Methods   Results and Discussion   Conclusion   Acknowledgments   References  

Wilson WI, Peng Y, Augsburger LL. Generalization of a Prototype Intelligent Hybrid System for Hard Gelatin Capsule Formulation Development. AAPS PharmSciTech.  2005; 6(3): Article 56. DOI:  10.1208/pt060356

Generalization of a Prototype Intelligent Hybrid System for Hard Gelatin Capsule Formulation Development
Wendy I. Wilson,1 Yun Peng,2 and Larry L. Augsburger1

1Department of Pharmaceutical Sciences, University of Maryland - Baltimore, Baltimore, MD 21201
2Department of Computer Science and Electrical Engineering, University of Maryland Baltimore County, Maltiomre, MD 21250

Correspondence to:
Wendy I. Wilson
Tel: (410) 939-7273
Fax: (410) 939-7301
Email: wendyiwilson@yahoo.com

Submitted: February 4, 2005;  Accepted: July 1, 2005;  Published: October 22, 2005


The aim of this project was to expand a previously developed prototype expert network for use in the analysis of multiple biopharmaceutics classification system (BCS) class II drugs. The model drugs used were carbamazepine, chlorpropamide, diazepam, ibuprofen, ketoprofen, naproxen, and piroxicam. Recommended formulations were manufactured and tested for dissolution performance. A comprehensive training data set for the model drugs was developed and used to retrain the artificial neural network. The training and the system were validated based on the comparison of predicted and observed performance of the recommended formulations. The initial test of the system resulted in high error values, indicating poor prediction capabilities for drugs other than piroxicam. A new data set, containing 182 batches, was used for retraining. Ten percent of the test batches were used for cross-validation, resulting in models with R2 ≥ 70%. The comparison of observed performance to the predicted performance found that the system predicted successfully. The hybrid network was generally able to predict the amount of drug dissolved within 5% for the model drugs. Through validation, the system was proven to be capable of designing formulations that met specific drug performance criteria. By including parameters to address wettability and the intrinsic dissolution characteristics of the drugs, the hybrid system was shown to be suitable for analysis of multiple BCS class II drugs.

Keywords: in silico modeling, capsule formulation, artificial neural networks, expert systems, low solubility drugs


Artificial Intelligence

The use of artificial intelligence, such as artificial neural networks (ANNs) and expert systems, provides an opportunity to systematically approach formulation in an efficient manor. ANNs are computer-based programs that attempt to simulate some features of the biological brain such as learning, generalizing, or abstracting from experience.1 ANNs are parallel information processing systems that can develop adaptive responses to environmental information.2 ANN models, such as back propagation learning networks, may be viewed simply as multiple nonlinear regression models. The experimental data and information generated may be transformed relatively easily into knowledge that can be used in the construction of domain specific rules by the formulator.

There are several advantages to using ANNs. Unlike statistics-based analysis, these programs do not require experimentally designed data. Incomplete or historical data can be used successfully to train ANNs. These programs can also model nonlinear and discontinuous functions. In spite of these advantages, there are disadvantages to using this type of modeling. The model generated is very specific and is dependent on experimental conditions. The ability of the program to successfully model the relationships hidden in the data is dependent on the quality of the data used to train the system. Overtraining of the ANN is also possible, resulting in “memorized” patterns instead of derived relationships.

An expert system is a computer program that emulates expert thought to solve significant problems in a particular domain of expertise. These intelligent computer programs use knowledge and inference procedures to solve problems. A unique characteristic of these problems is that they are often difficult enough to require significant expertise for their solutions. The knowledge necessary to perform at such an expert level as well as the inference procedures used are often thought of as a model of the expertise of the best practitioners of that given field.3

The use of expert systems (ES) presents several advantages. Well-functioning ES can facilitate an increased distribution of expertise in a given field and present a new communication channel for this knowledge. Moreover, the system provides protection of this knowledge by establishing a coherent and durable existence that can easily be accessed, modified, and updated. This knowledge base can serve as a valuable training aid for novices, allowing experts involved in training to focus on other issues. ES also provide for more consistent approaches to resolution of problems.3 On the other hand, these systems suffer from such limitations as lacking creativity in problem resolution. ES can only deal with issues that have been anticipated and included in the knowledge base.

Artificial intelligence is not a newcomer in the arena of pharmaceutical sciences. ANNs have been used to predict dissolution profiles and for formulation optimization.4-10 Their use has also been employed in predicting model granulation and tablet characteristics based on material and process variables11 as well as in estimating the aqueous solubility of structurally related drugs.12 ANNs have been used to assess in vitro-in vivo correlations13 and parameters such as crushing strength and disintegration time have been optimized.14 In preformulation, ANNs have been used to characterize the physiochemical properties of amorphous polymers.15

ES have been used to recognize complex relationships between formulation variables and in vitro drug release.16 They have also been used in the area of solid dosage development, especially in the areas of tableting and film coating.3 ES have also been applied to troubleshooting pharmaceutical processing equipment, such as rotary tablet presses.17 Prototype ES have been developed for the use in formulary decision making18,19 as well as selecting the most appropriate pharmaceutical powder mixer.20

Capsugel’s expert system (CES) for formulation support is a centralized system incorporating worldwide industrial experience to support formulation of powders in hard gelatin capsules.21 The most serious limitation is that the CES provides only a suggested formulation. The system provides no guidance or assurance that the predicted formulation will meet any particular dissolution, content uniformity, and/or weight variation requirement within user specified limits. It was hypothesized by Guo et al that the development of a hybrid system linking the current expert system to an ANN would effectively address this limitation and provide a facilitated way of generating new rules based on “learning.” The development of such a hybrid system that integrated an ANN with an ES could take advantage of the strengths of both the ANN and the ES while avoiding the weaknesses of either. By combining both of these systems, the knowledge of an ES can be used to design a formulation that could subsequently be optimized by the ANN. The concept of an expert network (EN) has been proven viable on a small scale by Guo et al using piroxicam as a model drug.19

Biopharmaceutics Classification System

The work of Amidon et al resulted in a scientific method to identify drugs based on their solubility and permeability.22,23 The biopharmaceutics classification system (BCS), introduced in 1995, consists of 4 drug categories: class I, class II, class III, and class IV. Class II drugs demonstrate high permeability but are poorly soluble. These compounds have the potential for enormous therapeutic success; however, absorption and, hence, the effectiveness may be limited by the rate of dissolution of the drug. Owing to solubility issues, the dissolution behavior of class II compounds is one of the most critical variables for this category of substances. In contrast to class I and class III, multipoint dissolution specifications are recommended for class II drugs. In addition, a complete characterization of the entire dissolution profile may be necessary to ensure quality control.24

The widely understood and studied drugs carbamazepine, chlorpropamide, diazepam, ibuprofen, ketoprofen, naproxen, and piroxicam were used as model drugs in this project. These drugs were selected because of their low solubilities and high permeabilities. These drugs were classified as BCS class II drugs based on their dissolution rate-limited absorption behavior. Owing to their low solubility, it was important to characterize the properties of the drugs that were associated with solubility and wettability. The intention of this project was to provide a more systematic approach to capsule formulation of BCS class II compounds by expanding the prototype EN for use in the analysis of multiple BCS class II drugs.

Materials and Methods


The following drugs were used as received from the suppliers: carbamazepine United States Pharmacopeia (USP) (lot RF1355, Spectrum Chemicals, Gardena, CA); chlorpropamide USP (lot 7812B, ICN Biomedicals, Aurora, OH); diazepam BP (lot RD0991, Spectrum Chemicals); ibuprofen USP (lot SE0196, Spectrum Chemicals); ketoprofen USP (lot RC0730, Spectrum Chemicals); naproxen (lot E2JA099, courtesy of Syntex Pharmaceuticals, Clarecastle, Ireland); and piroxicam (lot C21P903, courtesy of Pfizer Inc, Groton, CT).

The following materials were used as received from the suppliers: anhydrous lactose (lot 60679, Quest International, Hottman Estates, IL); citric acid (lot RT0330, Spectrum Chemicals), croscarmellose sodium NF (Ac-Di-Sol, lot T226N, courtesy of FMC Biopolymer, Newark, DE); cyclohexane (lot 000696, Fisher Scientific, Fair Lawn, NJ); fumed silica dioxide (Cab-O-Sil M-5P, lot 1J079, Cabot Corp, Billerica, MA); hydrochloric acid (lot 993660, Fisher Scientific; lot 43169, EMD, Gibbstown, NJ); hard gelatin capsules (lot 590311, courtesy of Capsugel, Greenwood, SC); microcrystalline cellulose (Emcocel 90M, lot E9B0B17, courtesy of PenWest Pharmaceutical, Patterson, NY); potassium citrate (lot QU0707, Spectrum Chemicals); potassium phosphate, monobasic (lot RP0462, Spectrum Chemicals; lot L49147, JT Baker Chemicals, Phillipsburg, NJ); and potassium phosphate, dibasic (lot PX0482, lot SM1234, lot SS0455, Spectrum Chemicals; lot 7080KAHA, Mallinckrodt, St Louis, MO); sodium lauryl sulfate (Stepanol, lot 1-354 87, Stepan Co, Northfield, IL); and sodium stearyl fumarate (Pruv, lot 305-01X, Mendell, Patterson, NJ).

Aqueous Solubility

Aqueous solubility was determined using the shaker-flask method. Two grams of neat drug was added to 50 mL of reagent grade water and was shaken for 24 hours at 25°C ± 1°C. Filtered samples were analyzed spectrophotometrically at the wavelength of maximum absorption for each drug. Each sample was analyzed in triplicate.

Contact Angle

Augustine Scientific (Newbury, OH) determined contact angles using the sessile-drop method. One gram of neat drug was compressed at 500 psig in a lab-scale Carver press. Ten drops of pure distilled water of volume 1 μL were placed on each compact surface and analyzed using a Kruss Drop Shape Analysis System (model DSA10, Kruss GmbH, Hamburg, Germany). The contact angles reported are the mean of 10 determinations.

Specific Surface Area

Single-point BET determinations of specific surface area were conducted by the Materials Analysis Laboratory at Micromeritics Inc (Norcross, GA) using nitrogen. Most actives were degassed at 100°C. Ibuprofen (IBU) was degassed at 60°C and ketoprofen (KET) was degassed at 80°C owing to their melting points being below 100°C.

Intrinsic Dissolution Rate

A quantity of 250 (± 1) mg of drug was compressed at an average compression force of 1000 lbs for 3 minutes to make nondisintegrating compacts using intrinsic dissolution rate (IDR) dies and a Carver Press (model 4687, Sterling Inc, Menomonee Falls, WI). The surface area of the compacts was 0.950 cm2. Compacts were tested in 900 mL of media maintained at a temperature of 37°C ± 1°C in a VanKel VK7000 (VanKel Industries, Edison, NJ) dissolution system fitted for IDR die attachments and were rotated at 100 rpm. Samples were analyzed with recirculation every 2 minutes over a time period of 1 hour and analyzed using an in-line Shimadzu spectrophotometer (model UV160U, Shimadzu, Kyoto, Japan) at the maximum absorbance wavelength for each active tested. The flow rate used was one mL/min. Based on the dissolution profiles obtained, the intrinsic dissolution rate was calculated using the following equation9:
G = d w d t 1 S (1)
where G is intrinsic dissolution rate (mg/min/cm2); dw is the change in drug dissolved (mg); dt is the change in time (minutes); and S is the surface area of the compact (cm2). The cumulative amount dissolved was plotted vs time for each vessel. The linear region of this plot (R2 ≥ 0.95) was determined using linear regression. The slope of the linear region was taken as dw/dt.


Hard gelatin capsules were manufactured using a capsule filling simulator. Fifteen-gram batches were blended for 10 minutes without lubricant. Lubricant was added and the batches were mixed for an additional 3 minutes. A size 1 tamping piston was used to compress 200 mg of formulated batch at 100 to 120 N using a laboratory scale Carver Press (model 4687, Sterling Inc). Compression force was monitored using a load cell (model 13, Sensotec, Columbus, OH), strain gauge conditioner (model 2160, Measurements Group Inc, Raleigh, NC), and digital oscilloscope (model 310, Nicolet Instrument Corp, Madison, WI). Twenty-five capsule plugs were formed and pushed into the empty capsule bodies (ConiSnap Gelatin Capsules, lot 590311, Capsugel) and closed by hand.


Dissolution testing was performed using a VanKel VK 7000 dissolution apparatus with a Shimadzu UV spectrophotometer (model UV-160). UV cells with a path length of 1 cm were used (model 175, Hellma, Plainview, NY). Automated sampling with recirculation was performed every 5 minutes for 45 minutes at a flow rate of 1 mL/min. A quantity of 900 mL of appropriate medium was maintained at 37°C ± 1°C for each capsule tested. The weakly acidic drugs (naproxen [NAP], KET, chlorpropamide [CHL]) were tested in 0.1 M pH 6.8 potassium phosphate (K PO4) buffer using USP apparatus II (paddles) with a rotation speed of 50 rpm. Capsules were deterred from floating using capsule sinkers (model 0500-0473, Epoxy Capsule Weights, Distek, North Brunswick, NJ). The weakly basic drugs (carbamazepine [CAR], diazepam [DIA]) were tested in 0.1 N HCl using USP apparatus I (baskets) with a rotation speed of 100 rpm. The percentage dissolved values at 10, 30, and 45 minutes reported are the average of 6 determinations.

Results and Discussion

Model Drugs

Table 1 includes the results of the determination of aqueous solubility, contact angle, specific surface area, and aqueous intrinsic dissolution rate. The contact angles for these drugs are all in the range of 90° to 100° with the exception of KET having a value of 68°. Drugs that are conducive to wetting generally have low contact angles. These values, along with the low values for aqueous IDR and aqueous solubility supported the selection of these drugs as models for the low solubility BCS class II drugs. The marginal values for the specific surface areas of these drugs, with the exception of KET, indicated that wetting of these drugs was problematic due to the limited surface area available for contact with the solvent.

Table 1. Physiochemical Properties of the Model BCS II Drugs*

Contact Angle (o)† Specific Surface Area (m2/g) Aqueous Solubility
Intrinsic Dissolution Rate (mg/min/cm2)†

CAR 93.3° (0.2) 0.5924 0.039 (0.001) 0.054 (0.004)
CHL 106.7° (0.3) 0.7107 0.049 (0.001) 0.038 (0.008)
DIA 101.5° (0.3) 0.1397 0.015 (0) 0.002 (0.001)‡
IBU 98.8° (0.3) 0.1892 0.035 (0) 0.015 (0.002)§
KET 67.5° (0.2) 2.3045 0.036 (0.001) 0.035 (0.003)
NAP 105.7° (0.4) 0.3486 0.006 (0) 0.008 (0.001)‡
PIR 90.4° (0.3) 0.7264 0.010 (0.001) 0.003 (0.0003)‡

*CAR indicates carbamazepine; CHL, chlorpropamide; DIA, diazepam; IBU, ibuprofen; KET, ketoprofen; NAP, naproxen; and PIR, piroxicam.
†Reagent grade water
‡Run time = 6 hours
§Run time = 3 hours

Expert Network

System design and architecture

A rule-based ES was developed and integrated with an ANN. Several assumptions were made in the development of the model ES for simplification purposes: (1) only directly fillable formulations will be considered (ie, granulation is outside of the scope of this program); (2) all excipients are compatible with the active ingredients; (3) a simplified blend uniformity model can be applied; and (4) diluents can be simplified to microcrystalline cellulose / lactose (MCC/LAC) blends (low dose, more LAC; high dose, more MCC). The ES was used as the decision module and the ANN served as the prediction module. These components were connected using 2 information exchange paths to form a loop. These systems, along with a control module (CM), formed the 3 major components of the hybrid EN. The flowchart detailing the functions and interrelationships of the 3 major components of the hybrid EN are detailed in Figure 1.

Figure 1 Overview of the Prototype Expert Network

Based on information provided by the user in the input package, the ES recommended a capsule-based formulation for the drug of interest. The CM transmitted the recommended formulation to the ANN, where the ANN predicted the dissolution performance of the recommended formulation. The ANN then returned the predicted dissolution performance to the CM. The user was then allowed to compare the predicted results with the target dissolution properties for the formulation. If the dissolution performance was not acceptable, the CM provided guidance to improve the dissolution and sent the new information to the ES for reformulation. The CM guided the optimization process until a satisfactory formulation was achieved or the optimization cycle was terminated by the user.


The ANN used in the prediction module was a back propagation learning system that computed output based on the forward pattern established by the training. The training of the back propagation network involved 3 stages: the feed forward of the input training pattern, the calculation of the output and back propagation of the associated error, and the adjustment of the weights associated with the variables. After training, the ANN computed the outputs using the feed forward method. By increasing or decreasing the weight associated with a given variable, the effect of that variable on the model developed was altered to reduce the error between the calculated values and the actual data. Variables found to be insignificant in the model were weighted less. Variables that contributed significantly to the model were weighted more.

The optimization of several ANN training parameters was the key to the success of the program. These parameters included the number of hidden layers, number of hidden nodes, type of training function, training time, training rate, and training slope. Different combinations of these parameters were evaluated to determine the optimum values for training to provide minimal system error for the predictions. For this study, the maximum system error allowed was 0.00002. The optimized ANN parameters used are listed in Table 2. A sigmoid function with a learning rate of 0.02 and maximum iterations of 30 000 was used for training. Sufficient training time along with a complimentary training rate ensured that the program would develop models for the data that it was presented. Seven input nodes were used to model the 7 input variables (% LAC, % disintegrant, % lubricant, % wetting agent, specific surface area (SSA), contact angle, and IDR). Three output nodes represented the 3 output variables, Q10, Q30, and Q45. Twelve hidden nodes were used. The number of hidden layers and nodes per layer were dependent on factors such as the number of input and response variables as well as the number of samples and the required prediction accuracy.

Table 2ANN Architecture Parameters

Input layer 7
Output layer 3
Hidden layer 12
Activation function Sigmoid
Slope 0.1
Learning rate 0.02
Error limit 0.00001
Maximum number of iterations 30 000

Data Analysis

Initial test of hybrid system

The capacity of the initial hybrid system to accurately predict the dissolution performance of model BCS class drugs other than PIR was tested before any modifications to the system were made. Input data for CAR, CHL, DIA, KET, and NAP were processed by the ES, and recommended formulations for each drug were manufactured. Dissolution testing was performed on these formulations to determine the experimental Q10, Q30, and Q45 values. These values were then compared with the predicted values generated by the hybrid system. The results of these tests are listed in Table 3.

Table 3Initial Test of Expert Network*


Diluent F-InSol F-InSol F-InSol F-InSol F-InSol
% Diluent 72 67 82 82 69
% Glidant 1 1 1 1 1
% Disintegrant 0.5 0.5 0.5 0.5 0.5
% Lubricant 8 8 8 4 8
% Wetting agent 0.1 0.1 0.1 0.1 0.1
Q10 20.3 67.3 86.8 94.8 85.1
Q10 Predicted 52.3 52.2 53.1 48.4 52.7
Q10 Error -32 15.1 33.7 46.4 32.4
Q30 40.3 98.9 98.3 97.5 98.2
Q30 Predicted 68.6 68.8 67.8 70.8 68.1
Q30 Error -32 15.1 33.7 46.4 32.4
Q45 50.6 98.8 97.8 98 99.2
Q45 Predicted 73.4 73.8 71.9 78.8 72.6
Q45 Error -32 15.1 33.7 46.4 32.4

*CAR indicates carbamazepine; CHL, chlorpropamide; DIA, diazepam; IBU, ibuprofen; KET, ketoprofen; NAP, naproxen; and F-InSol, blend of 75% anhydrous lactose and 25% microcrystalline cellulose.

The system recommended a filler system of 75% LAC and 25% MCC (F-InSol). The recommended formulations for each model drug are listed in Table 3. The differences in the values for predicted vs actual are included in the table as the error values. As evidenced by the high error values, the initial system was not very successful in predicting the dissolution performance of drugs other than PIR. Based on these results, retraining of the ANN was conducted using a new training data set.

ANN training

In order to form the causal associations between the formulation parameters and dissolution performance, the ANN was trained using experimental data. To ensure sufficient prediction power, it was extremely important to include sufficient experimental data from well-designed experiments. For this research, a Box-Behnken experimental design (Table 4) was used to develop a data set for the variables drug, excipient levels, and dissolution performance in the most efficient manner. The quality of the training data and the number of batches used dramatically affected the prediction power of the ANN. The variables and levels used in the training set data are listed in Table 5. In total, 182 batches were included in the training set data.

Table 4. Box-Behnken Experimental Design*


1 55 12 0.85 1 15 10 12 0.85 0.55
2 55 8 0.85 0.55 16 10 8 0.2 0.55
3 55 4 1.5 0.55 17 10 4 0.85 0.55
4 55 4 0.2 0.55 18 100 8 0.85 0.1
5 10 8 1.5 0.55 19 55 12 1.5 0.55
6 100 4 0.85 0.55 20 55 4 0.85 1
7 100 8 0.2 0.55 21 55 8 0.2 1
8 55 8 0.85 0.55 22 100 8 1.5 0.55
9 55 12 0.85 0.1 23 55 8 0.2 0.1
10 10 8 0.85 0.1 24 55 8 1.5 0.1
11 55 8 0.85 0.55 25 100 8 0.85 1
12 55 4 0.85 0.1 26 55 12 0.2 0.55
13 100 12 0.85 0.55 27 100 8 0.85 1
14 55 8 1.5 1

*%LAC indicates percentage of lactose in the MCC/LAC blend; %ADS indicates the percentage of disintegrant (Ac-Di-Sol); %SSF indicates the percentage of lubricant (sodium stearyl fumarate); and %SLS indicates the percentage of wetting agent (sodium lauryl sulfate).

Table 5. ANN Training Data Set Variables

182 Experimental Batches

3 Responses
7 Independent Variables
% Lactose in MCC/LAC blend (10%, 55%, 100%)
% Disintegrant (4%, 8%, 12%)
% Lubricant (0.2%, 0.85%, 1.5%)
% Wetting agent (0.1%, 0.55%, 1%)
Specific surface area (m2/g)
Contact angle (o)
Intrinsic dissolution rate (mg/min/cm2)

ANN validation

Validation of the ES was conducted to assess predictability and functionality. Ten percent of all available batches were randomly selected to serve as batches for validation and were not included in the training. The program was trained under optimal conditions and then used to predict the dissolution performance of the validation batches. The predicted dissolution was then compared with the experimentally determined dissolution data for these batches. Model statistics were determined during the training and are listed in Table 6. The target R2 value was ≥70%. The system was extremely accurate in modeling the training set data, resulting in R2 values > 99%. The system was successful in modeling the test data with increasing predictive capabilities as the amount of drug dissolved increased. Based on the training parameters chosen, the system error and R2 value indicate that the model determined by the ANN was very predictive of the dissolution behavior of the model drugs.

Table 6. ANN Training Model Statistics


Source of Variation Sum of Squares Degrees of Freedom Mean Squares Computed f ratio
Model 100938.846 109 926.044461 51.365618
Error 973.538386 54 18.028489
Total 100857.607 163
Train set R2 99.03474
Test set R2 69.053528
Source of Variation Sum of Squares Degrees of Freedom Mean Squares Computed f ratio
Model 66676.8402 109 611.714131 72.094966
Error 458.181267 54 8.484838
Total 66788.0279 163
Train set R2 99.313977
Test set R2 70.236563
Source of Variation Sum of Squares Degrees of Freedom Mean Squares Computed f ratio
Model 53297.3053 109 488.966103 139.130902
Error 189.779332 54 3.514432
Total 53231.717 163
Train set R2 99.643484
Test set R2 88.278314

The results for the validation batches are shown in Table 7. The predicted percentage drug dissolved was compared with the actual dissolution values and the error calculated. Considering the normal variability of real dissolution data, percentage error values ≤ ±10% for the comparison data were acceptable. Based on the data included in Table 6, it was determined that the ANN had a reasonable capability of modeling the relationship among the formulation parameters and dissolution performance.

Table 7. ANN Validation Results

% Lac % Disint % Lub %
Wetting Agent
SSA (m2/g) Contact Angle (o) IDR (mg/min/cm2) Q10 Q10 Pred Q10 Err Q30 Q30 Pred Q30 Err Q45 Q45 Pred Q45 Err Pattern Error

B13 10 12 0.85 0.55 0.592 93.3 0.057 15.9 12.3 3.6 38.8 26.6 12.2 47.8 47.1 0.7 80.5
B22 55 8 1.5 0.1 0.592 93.3 0.057 21.0 9.8 11.2 39.3 51.9 −12.6 46.1 54.4 −8.3 177.1
B44 55 8 1.5 0.1 0.711 106.7 1.154 97.2 91.2 6.0 98.2 98.3 −0.1 98.4 99.4 −1.1 18.5
B48 55 8 0.85 0.55 0.140 101.5 1.265 75.9 73.5 2.4 94.1 95.1 −1.0 98.0 97.8 0.2 3.4
B52 100 8 0.2 0.55 0.140 101.5 1.265 83.5 45.4 38.1 97.8 98.2 −0.5 96.7 97.7 −1.1 728.3
B57 100 12 0.85 0.55 0.140 101.5 1.265 83.7 80.6 3.0 99.0 98.0 1.0 97.9 96.9 1.0 5.6
B65 100 8 1.5 0.55 0.140 101.5 1.265 80.1 81.1 −1.0 96.7 98.2 −1.5 96.1 92.9 3.2 6.7
B78 10 8 0.85 0.1 2.304 67.5 1.511 96.8 101.8 −5.0 98.2 100.0 −1.8 98.4 92.2 6.2 33.2
B90 55 8 1.5 0.1 2.304 67.5 1.511 92.3 86.3 5.9 97.6 91.0 6.6 98.4 98.7 −0.3 39.3
B104 100 12 0.85 0.55 0.348 105.7 0.434 86.2 90.5 −4.3 98.0 97.5 0.5 99.5 100.0 −0.5 9.6
B128 100 5 1.5 1 0.161 90.4 0.033 76.2 83.9 −7.7 86.5 91.3 −4.8 88.9 88.5 0.5 41.7
B144 50 5 1 0.5 0.246 90.4 0.033 67.1 66.7 0.4 80.7 82.2 −1.5 85.2 86.9 −1.7 2.7
B156 0 4 0.9 0.6 0.277 90.4 0.033 63.7 66.7 −3.1 92.0 78.9 13.1 97.7 86.6 11.1 151.8
B165 100 8 0.9 0.6 0.277 90.4 0.033 81.7 90.9 −9.2 98.8 91.9 6.8 100.0 96.7 3.3 71.1
B167 0 4 0.6 1 0.277 90.4 0.033 64.1 61.0 3.1 84.2 81.3 2.9 89.9 78.8 11.1 70.6
B174 0 6 0.8 0.3 0.277 90.4 0.033 69.9 45.7 24.2 92.1 79.5 12.7 97.0 88.4 8.6 410.4
B176 0 6 0.2 1.1 0.277 90.4 0.033 70.9 72.2 −1.3 89.1 69.6 19.6 94.6 87.5 7.1 217.5
B179 0 9 0.2 0.7 0.277 90.4 0.033 75.7 85.4 −9.8 93.1 68.0 25.2 97.9 91.1 6.8 387.8

*Lac indicates the percentage of lactose in the MCC/LAC blend; [% Disint] indicates the percentage of disintegrant (Ac-Di-Sol); [% Lub] indicates the percentage of lubricant (sodium stearyl fumate); SSA - specific surface area; IDR - intrinsic dissolution rate; Pred - Predicted; and Err - Error.

The other facet of validation was assessing the ability of the EN to recommend a formulation based on sound design criteria and predict the dissolution performance accurately. The system should be able to perform these functions for drugs not included in the training as well. To determine the competence of the system in terms of these criteria, recommended formulations were manufactured and tested for the model drugs included in the training set. This information was used to investigate the accuracy of the system predictions. Input data for IBU, a drug not included in the training, was processed by the EN, and a recommended formulation was also manufactured and tested to determine if the system was capable of adequately formulating and predicting performance of drugs not included in the training.

The results of the external validation are shown in Table 8. The moderate success of the system can be attributed to the increased error in prediction at Q10 for CHL, IBU, and PIR. The result of 13.2% for IBU may not be considered extremely significant since the acceptance criteria is 10%. CHL and PIR showed significant deviations from the 10% acceptance criteria. The system was able to predict the behavior at Q30 and Q45 with much greater success (%error ≤ 10%) for all drugs with the exception of PIR. One factor that may have contributed to this lack of predictability for PIR was the fact that the experimental values for the input data for PIR were close to the values used for CAR. The system may have treated both drugs as one and made its predictions based on the patterns learned for CAR instead of PIR. The addition of another variable in the training set to further distinguish between drugs would overcome this issue. The results were somewhat promising in that the system was able to predict the performance of IBU with some success. It was also encouraging that the prototype EN was expanded to include several other BCS class II drugs along with the initial drug, PIR.

Table 8. Expert Network Validation Results*

Q10 Q10%error Q30
Q30 Q30%error Q45
Q45 Q45%error

CAR 21.2 20.3 -4.2 40.8 40.3 -1.2 48.3 50.6 4.8
CHL 94.8 67.3 -29.0 99 98.9 -0.1 99.5 98.8 -0.7
DIA 80.3 86.8 8.1 95.8 98.3 2.6 97.3 97.8 0.5
IBU 77.2 87.4 13.2 93.3 97.5 4.5 96.3 97.8 1.6
KET 91 94.8 4.2 97.2 97.5 0.3 97.9 98 0.1
NAP 87.3 85.1 -2.5 97.6 98.2 0.6 99 99.2 0.2
PIR 20.1 92.1 358.2 40.8 96.8 137.3 47.5 97.4 105.1

**CAR indicates carbamazepine; CHL, chlorpropamide; DIA, diazepam; IBU, ibuprofen; KET, ketoprofen; NAP, naproxen; and PIR, piroxicam.


From this research, the EN was expanded to include several other BCS class II drugs. Through validation, the EN was proven to be capable of recommending formulations for the model drugs that met specific drug performance criteria. By including parameters to address wettability and the intrinsic dissolution characteristics of the drugs, the EN was expanded to include multiple BCS class II drugs.


The authors would like to thank Pfizer, Inc (Groton, CT) and Syntex Pharmaceuticals (Clarecastle, Ireland) for their donations of PIR and NAP, respectively, and FMC Biopolymer (Newark, DE) and Penwest Pharmaceuticals (Patterson, NY) for their donations of croscarmellose sodium and microcrystalline cellulose, respectively. Funding for this project was provided by the Graduate School of University of Maryland-Baltimore, Capsugel (Greenwood, SC), and the National Organization for the Professional Development of Black Chemists and Chemical Engineers in conjunction with DuPont, Inc (Wilmington, DE).


1. Bourquin J, Schmidli H, van Hoogevest P, Leuenberger H.  Application of artificial neural networks (ANN) in the development of solid dosage forms. Pharm Dev Tech. 1997;2:111-121.

2. Achanta AS, Kowalski JG, Rhodes CT.  Artificial neural networks: implications for pharmaceutical sciences. Drug Dev Ind Pharm. 1995;21:119-155.

3. Rowe RC.  Expert systems in solid dosage development. Pharm Ind. 1993;55:1040-1045.

4. Ebube NK, McCall T, Chen Y, Meyer MC.  Relating formulation variables to in vitro dissolution using artificial neural network. Pharm Dev Tech. 1997;2:225-232.

5. Peh KK, Lim CP, Quek SS, Khoh KH.  Use of artificial neural networks to predict drug dissolution profiles and evaluation of network performance using similarity factor. Pharm Res. 2000;17:1384-1388.
PubMed  DOI: 10.1023/A:1007578321803

6. Kandimalla KK, Kanikkannan N, Singh M.  Optimization of a vehicle mixture for the transdermal delivery of melatonin using artificial neural networks and response surface method. J Control Release. 1999;61:71-82.
PubMed  DOI: 10.1016/S0168-3659(99)00107-8

7. Takayama K, Fujikawa M, Nagai T.  Artificial neural networks as a novel method to optimize pharmaceutical formulations. Pharm Res. 1999;16:1-6.

8. Takayama K, Takahara J, Fujikawa M, Ichikawa H, Nagai T.  Formula optimization based on artificial neural networks in transdermal drug delivery. J Control Release. 1999;62:161-170.
PubMed  DOI: 10.1016/S0168-3659(99)00033-4

9. Leane MM, Cumming I, Corrigan OI.  The use of artificial neural networks for the selection of the most appropriate formulation and processing variables in order to predict the in vitro dissolution of sustained release minitablets. Pharm Sci Tech. 2003;4:218-229.

10. Takayama K, Morva A, Fujikawa M, Hattori Y, Obata Y, Nagai T.  Formula optimization of theophylline controlled release tablet based on artificial neural networks. J Control Release. 2000;68:175-186.
PubMed  DOI: 10.1016/S0168-3659(00)00248-0

11. Kesavan JG, Peck GE.  Pharmaceutical granulation and tablet formulation using neural networks. Pharm Dev Tech. 1996;1:391-404.

12. Huuskonen J, Salo M, Taskinen J.  Neural network modeling for estimation of the aqueous solubility of structurally related drugs. J Pharm Sci. 1996;86:450-454.
DOI: 10.1021/js960358m

13. Dowell JA, Hussain A, Devane J, Young D.  Artificial neural networks applied to the in vitro-in vivo correlation of an extended release formulation: initial trials and experience. J Pharm Sci. 1998;88:154-160.
DOI: 10.1021/js970148p

14. Rocksloh K, Rapp FR, Abu Abed S, et al.  Optimization of crushing strength and disintegration time of a high-dose plant extract tablet by neural networks. Drug Dev Ind Pharm. 1999;25:1015-1025.
PubMed  DOI: 10.1081/DDC-100102264

15. Ebube NK, Owusu-Ababio G, Moji Adeyeye C.  Preformulation studies and characterization of the physiochemical properties of amorphous polymers using artificial neural networks. Int J Pharm. 2000;196:27-35.
PubMed  DOI: 10.1016/S0378-5173(99)00405-6

16. Hussain AS, Yu X, Johnson RD.  Application of neural computing in pharmaceutical product development. Pharm Res. 1991;8:1248-1252.
PubMed  DOI: 10.1023/A:1015843527138

17. Murray FJ.  The application of expert systems to pharmaceutical processing equipment. Pharm Technol. 1989;13:100-110.

18. Greer ML.  RXPERT: A prototype expert system for formulary decision making. Ann Pharmacother. 1992;26:244-250.

19. Guo M, Kalra G, Wilson W, Peng Y, Augsburger LL.  A prototype intelligent hybrid system for hard gelatin capsule formulation development. Pharm Technol N Am. 2002;26:44-60.

20. Lai FKY.  A prototype expert system for selecting pharmaceutical powder mixers. Pharm. Technol. 1998;12:22-31.

21. Expert System for Formulation Support. 1996. 

22. Amidon GL, Lennernas H, Shah VP, Crison JR.  A theoretical basis for a biopharmaceutic drug classification: the correlation of in vitro drug product dissolution and in vivo bioavailability. Pharm Res. 1995;12:413-420.
PubMed  DOI: 10.1023/A:1016212804288

23. FDA. In Vivo Bioavailability and Bioequivalence Studies for Immediate Release Solid Oral Dosage Forms Based on a Biopharmaceutical Classification System. Guidance for Industry. Rockville, MD: Food and Drug Administration; 2001. 

24. FDA. Dissolution Testing of Immediate Release Oral Solid Dosage Forms. Guidance for Industry. Rockville, MD: Food and Drug Administration; 1997. 

A publication of the American Association of Pharmaceutical Scientists
2107 Wilson Blvd., Suite 700, Arlington, Virginia, 22201, USA
703-243-2800, Fax: 703-243-9650, aaps@aaps.org
Copyright ©2003. All Rights Reserved. ISSN 1530-9932.
Legal Disclaimer