1. Results and Discussion
A Hybrid Expert System-Neural Network (“Expert Network”) for Capsule Formulation Support
1Gunjan Kalra, 2Mintong Guo, 1Yun Peng, 2Larry L. Augsburger
University of Maryland, Department of Computer Science and Electrical Engineer, Baltimore County; 2 University of Maryland, School of Pharmacy, Baltimore
Conclusion
Training Data Set
Introduction
GUI: interface
C functions
Preliminary results indicate that the PEN is a working system. Good predictive power of the NN module requires sufficient training samples and a cross validation process.
The objective was to construct a prototype intelligent hybrid Prototype Expert Network (PEN) for capsule formulation, which may yield formulations meeting specific running and drug delivery performance design criteria for BCS II drugs. To that end, a rule-based expert system (MES) was developed to specifically address BCS Class II drugs and integrated with a neural network (NN). These two components, which comprise the decision module and the prediction module, respectively, are connected together by two information exchange paths to form a loop. The system is believed to have the power to design a suitable capsule formulation to meet both requirements of quality control and dissolution.
Prolog Engine
BCS II N
CAPEX Y
Further research will be directed toward:
Validation and refinement of PEN
Automation of the parameter adjustment as a process of optimization.
Generalization of PEN to other drugs in BCS Class II.
CU: calculate PS to meet content-uniformity limit
OM: if PS is small, add diluent and use blend style
DF: choose excipients types
Final Formulation: calculate capsule size, % excipients, and final formulation
Acknowledgement
This work is being supported by Capsugel. We also gratefully acknowledge Pfizer Central Research for the gift of piroxicam.
Input Package
Dose/Sol
>250?
Permeability
> ?
Dose
Low < 50mg
CU Module
Eval_HalfDose?
Using the given equation to
calculate required PS
to achieve required tolerance
Use New PS
Ask Mixing Style
from user ?
Interactive Physical Blending
Liquid Addition
CAPEX
Diluent MINsol for OM
User Input: tapped & bulk density of OM
Computer Carr Index
Dose Volume > 1000
User Input: Bulk density of dose
Compute Carr’ Index
Compute Capsule Size
Diluent needed? N Y
High mg
Mod mg
V. high >1000mg
PS < 10 um
Old PS
New PS
Diluent MPS for OM
Choose Disintegrant
Choose Lubricant
Choose Glidant
BCS -IV
BCS -I or III
BCS -II
OM Module
DF Module
Choose Diluent
SSM
Flowability
Adhere to
metal?
Lubricant
Granulate
Fair
Bad
Good
V. Good
Low or Medium
Large or
V. Large
Test for DF (Plug Formation)
Diluent F-PS
Dose Volume
> mL
>1000mL
Particle
Size
>150μm
Ask User
for new
particle size
> μm
>10-50μm
Diluent F-Insol
Diluent M-InSol
Diluent M-PS
Drug
Poorly Soluble ml
Insoluble
>1000ml
4% Sodium starch glycolate
Croscarmellose
8% Sodium starch glycolate
Wettable?
Wetting Agent
Sodium Lauryl Sulfate
Prediction Engine
User:
Acceptable? N Y
Parameter
Adjustment
Final formulation
Predicted dissolution rate for the current formulation
Reformulate
compute
Materials and Method
MES
Prediction
Control
ANN
result
Microcrystalline cellulose (Avicel PH102 (FMC), Emcocel 90M (Penwest)), anhydrous lactose (direct tableting grade, Quest International), piroxicam (donated from Pfizer), magnesium stearate, Explotab (Penwest), Ac-Di-Sol ( FMC) and sodium lauryl sulfate have been used in the study. An instrumented Zanasi LZ-64 was used for the encapsulation process, and the compression force was maintained at 100 ~ 200N to achieve the specific target weight. The plug height was adjusted at 14mm. The dissolution testing was conducted on a Vankel 5000 dissolution station, and followed the USP procedure. The percentage dissolved in 10, 30 and 45 minutes were recorded as the measurements for the dissolution rate. Sixty-three batches have been generated to train and validate the system.
A rule-based expert system was developed in Prolog by followed the decision procedures in the flow chat, and integrated with a neural network (NN). These two components, which comprise the decision module and the prediction module, respectively, are connected together by two information exchange paths to form a loop.  
Results and Discussion
An expert system (MES) in the decision module (based on a decision tree modeled after the Capsugel Expert System1 [CAPEX]) was developed to provide decision rules for formulation recommendation. The NN in the prediction module (using backpropagation learning) was developed to provide predictive capability for the expected outcomes of the recommended formulation. The NN was trained with experiment data to capture the causal associations between the formulation and the outcome. The training was conducted with two experimental datasets using piroxicam as a model drug. The datasets represent two response surface designs for the capsule formulation which were developed to reflect the mapping from such variables as filler type/ratio, lubrication systems, drug particle size/specific surface area, disintegrants and surfactants to dissolution of the model compound. The capsules were filled using dosator-type automatic filling machines.
1 S. Lai, F. Podcek, J.M. Newton, and R. Daumesnil. An expert system to aid the
development of capsule formulations. Pharm. Tech. Eur., 8:60-65 (1996).