Performance Characterization of EDA and Its Potential to Improve Decision Making in Product Batch Release - Good Cascade Impactor Practices, AIM and EDA for Orally Inhaled Products - page 183

was to determine the ability of EDA to discriminate or detect differences in APSD
on the basis of MMAD as the measure of central tendency of each APSD. MMAD
by itself does not describe the APSD. However, once such a distribution is given (as
in examples used in this study), the MMAD value is an important characteristic as
well as a key indicator of any change to that APSD.
The advantage arising from using full-resolution CI results is that both the EDA
metrics and stage groupings can be compared to MMAD values determined on the exact
same CI run. This situation eliminates the need to obtain replicate measurements, and
the measurement variability can be directly estimated by assuming all the measurement
error resides with the EDA or stage grouping metrics when applied to the estimation of
MMAD . This latter assumption is reasonable considering the expected precision of
determining MMAD from the cumulative APSD obtained through the CI-based mea-
surement. Further, since one goal was to compare the performance of EDA to stage
groupings, any observed difference in the relative performance of the metrics should be
unaffected by any slight noise contribution from the MMAD determination.
In practice, values of MMAD were determined for each individual CI run according to
the method outlined by Christopher et al. [ 8 ]. A logistic model was fit to the cumulative CI
stage data and the resulting model was used to determine the MMAD of that particular
measurement. Next, the LPM / SPM ratio and stage groupings were also computed for each
corresponding individual CI measurement. The LPM / SPM ratio was obtained at a previ-
ously determined optimum boundary between LPM and SPM (see Chap. 7 ) .
The particular stage groupings were preselected consistent with FDA recom-
mendations outlined in Table 8.1 (Sect. 8.1 ). Regression analysis was then per-
formed on the metrics versus MMAD , on a product-by-product basis. The
relationship between LPM / SPM and MMAD was already known to be nonlinear,
and therefore, simple power functions were observed to provide an adequate fit to
the data. An a priori model for the relationship between stage groupings and
MMAD was unknown. Linear models appeared to provide an adequate fit and
were used to fit stage grouping results to MMAD . In all cases, 95% prediction
bands were computed for each regression. An example of the results is depicted
in Figs. 8.9 , 8.10 , and 8.11 for a CFC suspension MDI (IPAC-RS blinded database
code: w9kw01 ).
Table 8.4 is a summary of the full results of the regression analyses for all eight
products that were evaluated. The expressions for the power and linear models were
of the form

metric = b MMAD b
0

[ ] (8.1)

1

and

metric= bbMMAD
0 1 [

] (8.2)

respectively.
An examination of Table 8.4 yields some general observations. In all cases, the model
fit for the ratio metric was superior to any grouped stage (group 1-group 4) for all prod-
ucts, as evidenced by the relative magnitudes of the coefficients of determination ( R 2 ).