![](bg2.png)
classifier assigns an object x to the class k with the
largest a posteriori probability p(ω
k
|x). In One-Class
classification only p(x|ω
k
), the probability density of
the target class, ω
k
, is known. Estimating the
probability density from the training data and given
a threshold, the classifier accepts or rejects the test
samples. The 1-NN classifier assigns an object x to
its nearest class, with closeness measured by the
Euclidean distance between the vectors of inputs.
This paper is composed of 4 sections, besides the
current one. The next section presents the data
acquisition system from which ECG records were
obtained. Section 3 describes the authentication
system, detailing the implementation of the
classifiers. An overview and discussion of results is
provided in Section 4. Section 5 finalises the paper,
drawing the main conclusions.
2 DATA ACQUISITION AND
PROCESSING
The ECG data analysed in this work was acquired
within the scope of the HiMotion Project (HiMotion,
2008). The HiMotion Project consisted on the
design, implementation and administration of a set
of computer based experiments with cognitive tests
related to memory, concentration, association,
intelligence and insight (discovery). The underlining
idea is that these activities produce noticeable
changes in the physiological characteristics of
subjects, which, on one hand, are task dependent,
and therefore global task-related dynamics/features
can be recognized, and, on the other hand, individual
behavioural traits may be present in the acquired
data, and thus contribute for human authentication.
A set of physiologic signals was continuously
acquired during the realization of the tests:
electrodermal activity, blood volume pressure,
electroencephalography and ECG. A population of
24 male and female volunteers, with a mean age of
23.4±2.5 years, was asked to complete the series of
tests in individual sessions, designed to take, in
average, 30 minutes.
ECG measurements were taken using a surface
mount triode placed on the V2 pre-cordial
derivation. Each heartbeat waveform was
sequentially segmented from the full recording, and
then all individual waveforms were aligned by their
R peaks in segments of equal temporal length. The
mean wave for groups of 10 heartbeat waveforms
(without overlapping), was computed to minimize
the effect of outliers. A labelled database composed
by 137 samples was compiled, in which each pattern
corresponds to a mean wave. For each mean
waveform (Figure 1), the latency and amplitude for
each of the P QRS T peaks were extracted, along
with a sub sampling of the waveform itself,
providing a feature representation space of 53
features. In this work, only the latencies and
amplitudes of P, Q, S and T complexes were used,
resulting in 8 features, Table 1.
Table 1: Description of features.
Feature Description
1 Latency of P complex
2 Latency of Q complex
3 Latency of S complex
4 Latency of T complex
5 Amplitude of P complex
6 Amplitude of Q complex
7 Amplitude of S complex
8 Amplitude of T complex
Concluding, the available ECG data comprises
24 classes (each corresponding to each one of the
subjects under test) and 8 features.
3 AUTHENTICATION SYSTEM
The purpose of ECG-based authentication systems is
to attest that the user of a system is who he claims to
be, through the monitoring of its ECG records. In
this work, three classifiers were implemented using
Matlab (Matlab, 2007): MAP classifier, One-Class
classifier and 1-NN classifier.
The MAP classifier algorithm was constructed as
follows. Two mutually exclusive sub-sets from the
137 sample set are created, with 1 pattern for test
and the remaining 136 for training (“leave-one-out”
method). Then, density of the training data, p(x|ω
k
),
is estimated according to a maximum likelihood
technique later explained. p(ω
k
|x) is subsequently
computed for each test sample according to (1) and
the classifier decides on accepting test samples if (2)
is verified. This process is repeated for all the 137
samples. It is important to state that a Naive Bayes
model is considered for used features, thus assuming
statistical independence between them, (3). Also,
classes are assumed to be equiprobable, (4).
24
1
(| )( ) (| )( )
(|)
()
(| )( )
kk kk
k
ii
i
px p px p
px
px
px p
ωωω
ω
ω
=
==
∑
(1)
(|)
k
px
λ
>
(2)
BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing
164