Customized Prediction of Respiratory Motion - Prediction and Classification of Respiratory Motion - page 95

Table 5.2 The characteristics of the breathing datasets
Total patients Average records Minimum records Maximum records

130 66 min 25 min 2.2 h

5.4 Experimental Results

5.4.1 Breathing Motion Data

For the prediction of respiratory motion, we used patient breathing datasets
recorded at the Georgetown University CyberKnife treatment facility. Each
breathing recording has three marker breathing datasets, with a 26 Hz sampling
frequency, where each maker has three-coordinates. That means potential inputs
are as follows: (x 1 , y 1 , z 1 ), (x 2 , y 2 , z 2 ), and (x 3 , y 3 , z 3 ). The output is the position of
breathing motion corresponding to 3-coordinates. The total 130 patients breathing
recordings are randomly selected so that breathing datasets can be mixed up with
highly unstable and irregular breathing motions.
Table 5.2 shows the characteristics of the breathing datasets. The breathing
recording times average 66 min in duration, where the minimum and the maxi-
mum recording times are 25 min and 2.2 h, respectively. Each patient's recording
was used to train and predict respiratory motion. We used 5 min sampling data for
the feature extraction.

5.4.2 Feature Selection Metrics

We can derive 120 (= 10 C 3 ) feature combination vectors, i.e., choose three out of the
10 features defined in Table 5.2 , so that we can span three axis vectors corresponding
to the features chosen (shown in the next section). As shown in the following figure,
using results of the minimum value of H(I), we can select the combination number
(105, 106, 107, 108, 109, 110, 117, 118, 119, 120) corresponding to the estimated
feature combination vectors (Î), i.e., the feature combinations with Breath frequency
(BRF), Principal component coefficient (PCA), Maximum likelihood estimates
(MLE), Multiple linear regression coefficient (MLR), and Standard deviation (STD).
This result also confirms that the three chosen axes can provide the distinct
discriminate feature distribution.
Now, we would like to choose the class number (c) with the minimum value of the
objective function (J(c)). The figure shows the clustering of the estimated feature
combination vector (Î) with respect to the class number (c = 2,…, 7). We calculate
the objective function value (J(c)) with a different class number. The class number
(c = 5) is chosen to minimize the criterion J with the corresponding class.
With increasing the cluster number (c), the estimated class number (ˆ)is
selected to get the minimum of the objective function value (J(c)). We can notice