Instrumentation & Measurement Magazine 24-6 - 9

is related to the module of the three acceleration components
and the Gyroscope output recorded at the user's hip. Scores
are generated by computing the maximum value of the
cross-correlation between the unknown pattern and ADLs Signatures,
while a threshold classification paradigm is used. As
it can be observed, although good performance in recognizing
the unknown pattern is clearly demonstrated, some cases of
misclassification or unclassified events may still occur.
Fig. 3. Example of an ROC Curve. Each point has been estimated by fixing a
threshold value.
value, TP, FP, TN, and FN indexes are computed to obtain the
TPR and FPR metrics needed to build the curve. Taking into
account that a well-performing classifier should be characterized
by a high TPR and a very low FPR, the top-left corner of
the graph represents the most convenient condition. When the
ROC curve has been obtained, different methods can be adopted
for the identification of the optimal threshold, such as
the value minimizing the distance to the top-left corner of the
ROC curve.
It must be observed that the result of a threshold-based
classification strategy can lead to multiple classifications (an
unknown pattern could
be recognized as belonging
to different classes) and
unclassified events (an unknown
pattern could not
be classified as belonging
to any of the considered
classes). In such cases, the
use of a post-processing
strategy could help. As an
example, in the case of fall
detectors, the analysis of
the static value of acceleration
components (recorded
after a potential fall) can
help to take out misclassifications
from the set of
multiple classifications
[16].
As an example, Fig. 4
shows results of an eventdriven
approach for ADLs
classification (the class of
events is the same abovementioned)
[17]. The main
adopted rough information
September 2021
Reliability of Classifier
One of the most mandatory requirements for the development
of an effective classifier is the definition of a reliability index,
which should be computed each time the classification paradigm
identifies one or more classes to which an unknown
pattern may belong. It is important to remember that in this
case, a priori knowledge about the right class of belonging of
the unknown pattern is not known.
A suitable way to define such Reliability Index, in case of
threshold classifier, could be to weight the relative distance
between score(s) values and pre-defined threshold(s) [18].
This index can be very effective, not only to assess the reliability
of the classification performed by the paradigm, but also
to help discriminate among multi-classifications in case this
should happen. Sometimes, although results of the classification
process may appear straightforward in defining for each
unknown instance's relevant class of belonging, the Reliability
Index may confirm or not the situation.
As an example, Fig. 5a shows the output of a classifier
adopted to identify stable/unstable postural behaviours
[18], based on a threshold classifier and scores generated by
Fig. 4. Results of event classification by the data fusion algorithm. FF: Forward Fall; LF: Lateral Fall; BF: Backward Fall;
SUR,L, SDR,L: Step Up/Down by the Right/Left leg; LD: Lying Down; SI: Sitting [17] (©2016 IEEE).
IEEE Instrumentation & Measurement Magazine
9
Instrumentation & Measurement Magazine 24-6

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-6
