Neural computing in pharmaceutical products and process development - Computer-Aided Applications in Pharmaceutical Technology - page 136

reliability. The model tree approach is simple yet informative and is
expected to gain much interest in the fi eld of formulation and
process development.

Example 4 (Petrovi ´ et al., 2011)

Fluidized-bed granulation process was analyzed using the decision
tree approach. Infl uence of various input parameters (product, inlet
and outlet air temperature, consumption of liquid-binder, granulation
liquid-binder spray rate, spray pressure, drying time) on granulation
output properties (granule fl ow rate, granule size determined using
light scattering method and sieve analysis, granules Hausner ratio,
porosity, and residual moisture) was assessed. In order to perform
classifi cation analysis on the fl uidized-bed granulation study, input
and output parameters were discretized by size in bins containing
either four or six elements. Decision trees were built using an
algorithm similar to Quinlan's C4.5. Default software parameters
for the training algorithm were used - pruning, subtree raising, and
nonbinary splits were allowed. A single tree was built for each
granule property separately. The decision tree module underwent
cross-validation, whereby the data set was divided into a random
subset of cases and one set was left out whilst decision trees
were built. Developed decision trees were applied on the left out
test set, and the predicted results were recorded. Distribution of
classes between the training and the validation sets was similar.
Apart from internal cross-validation, additional external validation
of the built decision trees was performed, using the test examples.
Once the trees were built, rules were inducted and information
gain was the criterion used for selecting attributes and numerical
splits.
Two decision trees, built for medium and median particle size of
granules determined by using the dynamic light scattering method,
are represented in Figures 5.12 and 5.13. Top-down decision trees
enable prediction of categories, that is, ranges of particle sizes as

￿

￿ ￿