Other Applications: Sequential Dependence Modelling and Data Mining - On Statistical Pattern Recognition in Independent Component Analysis Mixture Modelling - page 157

Table 7.2 Main data corresponding to the sleep staging experiment

Patient Testing sample

size (number
of epochs)

Training sample
size (number
of epochs)

Estimated
stage transition
probabilities

Percentage
error
ICAMM
(%)

Percentage
error
SICAMM
(%)

1 445 (3.7 h)

424 ''stage 1''
21 ''stage 2''

445 (3.7 h)
402 ''stage 1''
43 ''stage 2''

P11 = 0.97
P22 = 0.69

20 9

2 454 (3.8 h)

403 ''stage 1''
51 ''stage 2''

454 (3.8 h)
361 ''stage 1''
93 ''stage 2''

P11 = 0.96
P22 = 0.80

36 20

stage is made every 30 s. The selected four features extracted from the PSG
signals were: the amplitude, the dominant rhythm, and the theta-slow-wave index
(TSI), which were estimated from the C3-A2 EEG channel, and the alpha-
slow-wave index (ASI), which was estimated from the O2-A1 EEG channel. The
dominant rhythm was estimated as the pole frequency of the second-order auto-
regressive (AR) model; the ASI was the ratio of power in the alpha band
(8.0-11 Hz) to the combined power in the delta (0.5-3.5 Hz) and theta
(3.5-8.0 Hz) bands; and the TSI was the ratio of power in the theta band to the
combined power in the delta and alpha bands. These features are commonly used
in computerized PSG analysis [ 10 , 11 ].
Two patients with apnea were considered for the experiment. The main data are
included in Table 7.2 . The parameters A k ; b k ; p ½ s k k ¼ 1...K were estimated
from the training record in a supervised form using the JADE algorithm [ 12 ]. The
labelling of stages ''1'' and ''2'' that is required for the supervised training was
made taking into account the manual score done by the expert. The probabilities of
transition between stages were also estimated from the training record. Note that
the probabilities of permanence in the same class are clearly above 0.5, so the use
of SICAMM seems to be justified in this application. The reference hypnograms
were also obtained by an expert using conventional non-automatic procedures.
Obviously, there is a clear improvement of SICAMM with respect to ICAMM
when we compare the percentage of error in the automatic computation of the
hypnograms with respect to the reference hypnogram.
To have a better understanding of the results, in Figs. 7.3 (patient 1) and 7.4
(patient 2), we show the hypnogram estimated by the expert, together with the
hypnograms computed by SICAMM and ICAMM. Essentially, SICAMM reduces
the number of false detections of arousals so that a ''cleaner'' hypnogram is obtained.
In patient 1, we verified that 9 of the misclassified stages obtained with SICAMM
actually corresponded to ''stage 2'' and 31 corresponded to ''stage 1''. With
ICAMM, 9 of the misclassified stages obtained actually corresponded to ''stage 2''
and 81 corresponded to ''stage 1''. Similarly, in patient 2, we verified that 31 of the
misclassified stages obtained with SICAMM actually corresponded to ''stage 2'' and
61 corresponded to ''stage 1''. With ICAMM, 31 of the misclassified stages obtained
actually corresponded to ''stage 2'' and 165 corresponded to ''stage 1''.