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 236

Table 8.11 Stage groupings
and PCA acceptance limits

Group NGI system Acceptance range (%LC)

1 IP and PS 59.2-95.5

2 S1 to S3 7.7-12.4

3 S4 to S6 10.7-20.5

4 S7 and MOC 0.3-2.3

Table 8.12 Summary of classification of data points in test set by EDA metrics and stage grouping

Agreement
with PCA

Type I
error rate Method Pass rate Type II error rate

PCA 934 (54%) - - -

LPM / SPM ratio 1,258 (72%) 809 (46.6%) 125 (7.2%) 449 (28.8%)

LPM / SPM ratio and ISM 872 (50%) 783 (45.1%) 151 (8.7%) 89 (5.1%)

Grouping 1 869 (50%) 693 (39.9%) 241 (13.9%) 175 (10.1%)

Grouping 2 1,068 (61%) 777 (44.7%) 157 (9.0%) 291 (16.7%)

Grouping 3 1,366 (79%) 923 (53.1%) 11 (0.6%) 443 (25.5%)

Grouping 4 1,665 (96%) 929 (53.5%) 5 (0.3%) 736 (42.4%)

All stage groups 635 (37%) 595 (34.2%) 339 (19.5%) 40 (2.3%)

“bad” batch but are accepted by the analysis. Again, a type II error occurred when
an APSD profile was deemed dissimilar from the PCA analysis but was deemed
similar to the 252 clinically relevant profiles by the other methods.
A graphical representation of these errors in terms of the PCA scores plot is
given in Figs. 8.69 and 8.70 for EDA and stage grouping, respectively. Batches or
APSD profiles colored purple that lie inside the Hotelling T 2 ellipse were deemed
similar to the training data set through the PCA analysis, but were deemed dissimi-
lar by EDA or grouped-stage analysis.
Batches or profiles indicated by light blue coloring were deemed outside the
training data set population but were deemed similar to the training data set by the
EDA or grouped-stage analysis.

8.5.3.5 Outcomes

The number of type II errors was slightly lower for stage groupings (2.3%) compared
to EDA metric-based criterion: LPM / SPM + ISM (5.1%), by using PCA as a refer-
ence. There was, however, a significant difference between EDA and stage grouping
with regard to type 1 errors, where EDA gave considerably fewer type 1 errors than
the grouped-stage approach. This analysis demonstrated the potential value of PCA
as a tool in the development of APSD control strategies to allow changes to be