Pattern Classification Techniques for Lung Cancer Diagnosis by an Electronic Nose - Computational Intelligence in Healthcare: Advanced Methodologies - page 389

- Fractional . This approach combines the differential and relative manipu-
lation, substracting the baseline x s (0) from the sensor response x s ( t )and
dividing the modified sensors response by the baseline x s (0):
y s ( t )= x s ( t )

x s (0)
x s (0)

−

(3)

It is important to highlight that fractional manipulation provides not only
dimensionless, but also normalized measurements.
The choice of baseline manipulation is strongly related to the specific sensor
technology and the particular application; for MOS sensors chemoresistors it
has been shown [7] that the fractional change in conductance provides the best
pattern recognition performance.

Feature extraction. The second stage in preprocessing consists in identifying
a number of descriptors able to condense all the response information in few
features. This new data will represent the olfactory blueprint of the analyzed
substance. This is a very crucial step, since all further pattern analysis will be
performed on these data: the lack of information must therefore be avoided.
In most electronic nose works the feature extraction operation is performed by
keeping one single parameter (usually the steady-state, i.e., the final, maximum
or minimum value reached during the response to the stimulus). However this
approach is very reductive since the useful information could be elsewhere: tran-
sient analysis could indeed provide significant information and thus improve the
final performance. In this perspective there are several approaches to detect
features in the transient and steady-state response [5]:
- Sub-sampling methods : sampling the sensor response and/or its derivative at
different times during the acquisition of the analyzed gas
- Parameter extraction methods : extracting a number of features able to cap-
ture the information contained in the response curve (e.g., integral, maxi-
mum or minimum in certain time intervals, rise time, slope etc.)
- System identification methods : fitting a theoretical model (e.g. exponential or
autoregressive) to the experimental transient and use the model parameters
as features

As demonstrated in [5,4], the consideration of the transient, besides the steady-
state, in the feature extraction operation, can lead to many advantages, such as
an improvement of classification performance (if the discriminative information
is in the transient), as well as a potentially better parameters repeatability w.r.t.
static descriptors, as exhibited in some cases [6]. Moreover if the discriminative
information is in the transient there will be no need to reach the steady-state of
the response, allowing a reduction of measurement time and, consequently, an
increase of sensor life time.
In the lung cancer diagnosis application, we extracted the features using all
the three approaches and taking into consideration both the steady-state and
the transient information. The ten descriptors that we extracted from the sensor
responses were based on: