3.1. Quantitative Measurement of Elicited OKN
In order to verify the accuracy of our method, we invited physicians to assist in marking the correct point when FP occurred and created a graph for comparison (
Figure 6). For verification strategy, we randomly selected the OKN test data from each of the six subjects twice to allow the professional doctors to manually mark the location of the FP occurrence. Next, each intermediate process of our method was used to detection the FP location. For comparison, the position we circled contains the position marked by the doctor (red dot), which is considered to be the correct selection of the FP position, as shown in
Figure 6b.
Table 4 lists the average results of each step for six subjects and also lists
p-values using paired t-test between each intermediate process of our method and ground truth. More specifically, the paired t-test was used to estimate whether the differences between the performance of each intermediate process of our method and ground truth were significant or not. The results showed that 97% of the FP occurrence points of OKN could be found through our method (
Figure 7). The reason why they could not be completely identified was that some special signals were difficult to filter, which is explained in detail in Section C. For normal OKN signals, the location of occurrence could be completely obtained.
3.2. Comparisons with the State-of-the-Art Approaches
The traditional method adopted the Fourier transform combined with band-pass and high-pass filtering and employed the Welch method [
18] to analyze the spectrum (2). It was found that the OKN frequency was about 2 Hz to 4 Hz (
Figure 8). This frequency, compared with the GP3 callback, was consistent with the saccadic state of the subjects; that is, under our experimental environment, the subjects’ occurrence of OKN varied from two to four times per second
The Fourier transform was then used to find the frequency (3) at which FP occurred. High-pass and band-pass FIR filtering were adopted for peak detection. After smoothing, an adjustable threshold was set for the screen, and the filtered peak was found. The adjacent maximum/minimum values were found according to the signal position (
Figure 9) [
19]
This method ensured a certain degree of accuracy, but the parameters of the threshold needed to be adjusted every time. If the parameters were wrong, it would be possible to omit the FP position. Only after repeated tests by professionals could the setting become more accurate (
Figure 10). Moreover, it could be easily affected by dullness and excessive eye movement, which make it difficult to set the threshold.
In the bandpass processing, although subjects were asked to focus on the fixed range of the screen, their eyes were difficult to control when they were tired, which caused eye drift, dullness, and other behaviors. Therefore, even if OKN occurred regularly, the eyes could not be fixed in the same range, and the subjects could easily be distracted (
Figure 11). This phenomenon made it difficult to find the maximum amplitude of the signal either through high-pass or band-pass filtering. In addition, the response of FP in OKN was not consistent, and the amplitude of eye displacement was unstable. All these conditions made it difficult to find the highest (low) point of the fixed frequency in either high-pass or band-pass processing.
After the Fourier transform, we analyzed the high-pass and band-pass signals, and then set the threshold to compare with the signals marked by the physician after artificial observation. It was found that although the threshold was set very close to the maximum average height, there would still be omissions when the signal was elongated, and it varied greatly. As the eye moved away from the original position, it became more difficult to set the threshold and it was not necessarily accurate, which also led to the decrease of accuracy (
Figure 12).
Under the complete dynamic signal, it could be found that an unstable signal would inevitably affect the FP judgment, and the screening based on the slope and amplitude of vibration had a higher identification rate than the band-pass (or high-pass) filtering after the Fourier transform. The reason was that the slope of FP would not affect the threshold judgment with the involuntary eye shift of the subject. Our method could eliminate the need for complex calculations and make the judging simpler and more accurate.
In order to evaluate the performance of the proposed method, the three traditional methods [
6,
8] were implemented for comparisons.
Table 5 shows the comparison and analysis between the proposed method and the three traditional methods [
6,
8], and
p-values using paired t-test between the three traditional methods and the proposed method. More specifically, the paired t-test was used to estimate whether the differences between the performance of three traditional methods and that of the proposed method were significant or not. At first, we thought that we should only use Peak finding [
20] to find the instantaneous high point of FP. However, this method is prone to noise interference and can cause serious misjudgment, resulting in low accuracy. The application of the Fourier transform and filtering [
19] depends on the judgment of values, which requires experienced researchers to select and set. Among them, high-pass filtering is the easiest to select, so the FP position can be correctly determined.
Regardless of the normal OKN identification rate, the uncontrolled noise produced by the subjects affected the determination of the correct FP during the prolonged test time. The OKN with noise includes dullness, fixation position deviation, and other phenomena, which further deepen the difficulty of filter setting and greatly decrease the accuracy. Our method simply filtered by the slope and moving distance, which solved many judgment problems, and screened the FP position from an intuitive angle. It reduced the influence of fixation deviation, leaving only the stagnant fixation problem (the original problem is explained in IV). In addition to the normal FP judgment of OKN and FFT combined with a high-pass filter that can be used well, there were good results on the correct FP judgment of OKN signals with noise.
3.3. Research Contributions and Limitations
Clinical detection of OKN often requires costly instruments, which poses considerable difficulties in clinical applications. Our main contributions were as follows: (1) low-cost instruments could be used to detect OKN, and (2) simple conditional screening and iteration could be used to obtain the accurate FP position, which was conducive to the calculation and analysis of the subject’s eye disease status.
Our method could be used to completely identify the FP position under stable conditions without large eyeball stagnation. However, the reaction of the eyeball cannot be controlled, and abnormal nystagmus and dullness are randomly generated and not included in FP/SP for reference (
Figure 12). Our method has not yet been able to filter these unstable signals. If these phenomena occur more frequently, the accuracy of FP signal identification will decrease (
Figure 13).