3.2.1 Materials
The experimental data were acquired from the sleep center of Tzu-Chi Hospital in Taichung, Taiwan. There are 2 records from 1 male and 1 female. The ground truth ta-ble is shown in Tata-ble 3.5.
3.2.2 Experimental results
In the classification of phases subtypes, Figure 3.12 shows the distribution of A-phase subtypes of RemLogic system scoring result. The following Table 3.6 presents the accuracy of the proposed subtypes classified method which uses RemLogic system scoring and RPSGT manual scoring as reference. The classified accuracy of subtypes is 65.5% to 77.9% in subtypes A1, 16.7% to 41.5% in subtypes A2, and 54.3% to 75% in subtypes A3.
Figure 3.12: The row dimension means times of all night and the column dimension means fractal dimension value. The blue line is fractal dimension by Higuchi’s method and the subtypes of A1, A2, and A3 are presented by red, green, and black point.
The variable length template is a flexible method to adaptive range of time. In Fig-ure 4.1, we plan to investigate the influence of A-phase detection with the variable length template or not. The experimental result without using template method is more similar with ground truth than use that. And in the Figure 4.4 and 4.5, the accuracy of A-phase detection and CAP sequence detection without template is higher than using template. But the results demonstrate that may not adjustment A-phase length if we do not use template.
The A-phase length range is 2 second 6 second too short to correspond the previous results with 2 second 60 second.
Figure 4.2 and Figure 4.3 shows the distribution of average descriptor of A-phase and non A-phase. They are the proposed method without using template and the proposed method. The row shows the time dimension and the column shows descriptor mean value in each phase. Compare with the proposed method without using template and the proposed method, the distribution of A-phase and non A-phase of the proposed method converges to a constant respectively.
4.1.2 Threshold adjustment
In the previous methods, the exist threshold (equal to 1) and the length threshold (equal to 0) are used to distinguish A-phase from background signal. But it is different for each subject’s physiological state. In our experiment, we set exist threshold to 0.8 and length threshold to 0.3 to compare with previous studies set. This result of accuracy of A-phase detection and CAP detection is shown in Figure 4.4 and Figure 4.5.
Figure 4.1: The five rows are sleep stage, EEG signal, CAP detection result without tem-plate, RPSGT result, and RemLogic system result. The third row shows that the duration of A-phase is reduced to 2 second - 6 second.
4.1.3 Forward function
The forward function is a strategy to overcome that the template method only extends the length to the right. This function is turned on to extend the template to the left for 2 s.
We also compare the accuracy for CAP detection in Figure 4.4 and Figure 4.5. The orig-inal method detects CAP without using the forward function. Compare with the accuracy with forward function, turning off this function is better. So this function may bring more unnecessary influence than benefit.
4.1.4 Band selection
In the previous studies, the delta band was the most frequency band used in CAP detec-tion. Some studies dropped sigma band for simplicity, due to the secondary or negligible role of this rhythm for CAP detection in normal subjects. In our experiment, we test dif-ferent band set in exist threshold and length threshold for CAP detection. Figure 4.4 and Figure 4.5 show the result of the accuracy for CAP detection.
Figure 4.2: The distribution of average descriptor of CAP phases on Subject1. The figure on the top is the proposed method and the figure at the bottom is the proposed method without using template
Figure 4.3: The distribution of average descriptor of CAP phases on Subject2. The figure on the top is the proposed method and the figure at the bottom is the proposed method without using template.
Figure 4.4: Show the Subject1 result which use the proposed result as test and the registered PSG technologists manual scoring as reference. The top figure is A-phase detection result and the bottom figure is CAP sequence detection result. These eight lines from top to down are the proposed method without template, exist threshold set to 1 and length threshold set to 0, the proposed method with forward function, the length band without using delta band, the exist band using delta and theta band, the exist band using delta, higher delta and, lower delta band, the exist band using delta, higher delta, lower delta, and theta band, and the proposed method.
Figure 4.5: Show the Subject2 result which use the proposed result as test and the registered PSG technologists manual scoring as reference. The top figure is A-phase detection result and the bottom figure is CAP sequence detection result. These eight lines from top to down are the proposed method without template, exist threshold set to 1 and length threshold set to 0, the proposed method with forward function, the length band without using delta band, the exist band using delta and theta band, the exist band using delta, higher delta, and lower delta band, the exist band using delta, higher delta, lower delta, and theta band, and the proposed method.
RemLogic A 653 921
4.1.5 Vertex sharp wave
We had developed a vertex sharp wave detection method to improve CAP detection.
But it performed negative influence in the A-phase detection. In Table 4.1 and Table 4.2, the confusion matrix of A-phase detection result shows that the true positive in non A-phase is decreasing.