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 240

The results of the Tougas OCC approach were summarized qualitatively; that is,
the false accept and false reject regions of the OC curves generated for all products
and for each metric were visually compared. As shown in Tables 8.7 and 8.8 , in the
majority of the cases, the EDA metric demonstrated better performance in making
the correct decision relative to the grouped-stage metrics. The results of the
Christopher-Dey performance evaluation approach was summarized quantitatively
in Table 8.9 by providing the estimated error rates (type I = false reject and type
II = false accept). For both the Tougas and Christopher-Dey approaches, the shape
of the OCC for each product based on LPM / SPM ratio more closely resembled the
shape of the ideal “top-hat”-shaped OCC than those based on grouped stages. The
conclusion from both Tougas and Christopher-Dey approaches was therefore that
the ability of the LPM / SPM ratio metrics to correctly discriminate between accept-
able and unacceptable APSDs was better than the combined-grouped-stage
metrics.
The PCA multivariate approach is also analogous to a decision equivalent proce-
dure. In this instance, the objective was to decide whether or not a set of APSD
profiles were similar or dissimilar (typical versus atypical); whereas, in the Tougas
and Christopher-Dey OCC-based approaches, the focus was on distinguishing
acceptable versus unacceptable through the use of pseudo-product specifications.
As there were two distinct PCA data sets, the original data set was used as the train-
ing or reference data set and the other as the performance evaluation set, and there-
fore, it was not necessary to simulate APSD data. The results of the PCA performance
evaluation method indicate that the false declaration of similarity (type II error) is
slightly lower for the grouped-stage metrics (overall 2.3%) than the EDA metrics
(5.1%). And the false declaration of dissimilarity (type I error) is significantly
higher (19.5%) for the grouped-stage metrics than the EDA metrics (8.7%).

8.7 Conclusions

Because the LPM/SPM ratio is more predictive of particle size changes in the CI
measured APSD than API mass data derived from grouped-stages, the relative pro-
portion of incorrect rejections (type I errors) to incorrect acceptances (type II errors)
is much lower for the EDA approach than for the grouped-stages approach. As a
consequence, when the LPM/SPM ratio acceptance limits are adjusted to control
incorrect acceptance to the same rate for both approaches, the number of incorrect
rejections are much lower for the EDA approach. Figure 8.71 therefore provides the
answer to the question posed in Fig. 8.2 that EDA could be more discriminating at
making correct decisions about changes in OIP APSD that influence specific parti-
cle size ranges, than using the grouped-stage approach currently recommended by
the FDA in OIP quality assessments.