A new diagnostic vestibular evoked response
© Dastgheib et al.; licensee BioMed Central. 2015
Received: 24 June 2014
Accepted: 20 January 2015
Published: 11 April 2015
To describe the development of a new clinically applicable method for assessing vestibular function in humans with particular application in Meniere’s disease.
Sophisticated signal-processing techniques were applied to data from human subject undergoing tilts stimulating the otolith organs and semicircular canals. The most sensitive representatives of vestibular function were extracted as “features”.
After careful consideration of expected response features, Electrovestibulography, a modified electrocochleography, recordings were performed on fourteen Meniere’s patients and sixteen healthy controls undergoing controlled tilts. The data were subjected to multiple signal processing techniques to determine which “features” were most predictive of vestibular responses.
Linear discriminant analysis and fractal dimension may allow data from a single tilt to be used to adequately characterize the vestibular system.
Objective, physiologic assessment of vestibular function may become realistic with application of modern signal processing techniques.
Vestibular disorders are among the most common reasons that patients seek the advice of a physician, yet the diagnosis of dizziness largely relies on the patient history. The patient history is subjective and its reproducibility has not been validated. Significant physiologic disruptions of neurological function should cause repeatable, measureable changes in neural activity. We believe that sophisticated and objective measurement of these changes should be diagnostic and should reveal underlying pathologic mechanisms. This paper outlines the application of advanced statistical signal processing techniques from the fields of engineering and statistics to understand normal and pathologic vestibular function using Meniere’s disease as a prototype.
Evoked potentials have been successfully applied to diagnose auditory disorders but may be difficult for vestibular diagnosis. Auditory evoked potentials typically involve temporal averaging of several hundred auditory stimuli which may be problematic in vestibular stimuli. On the other hand, when observing the averaging process of auditory evoked potentials in real time, the first response or two are often adequate to see the general nature of the response. It would seem then that the first response or two should contain diagnostic information if it could be extracted. With this observation in mind is seems plausible that sophisticated signal processing techniques might be able tease out enough information from a few tilts to permit recognition of repeatable patterns of waveforms that could be diagnostically useful.
Usually several features as biomarkers are extracted from the output of the NEER algorithm on the EVestG signals. Most diagnostic tests measure the signals’ most important parameters to classify a system as normal or abnormal. The “feature” extraction technique utilizes many quantitative criteria from the signal to categorize the response. Extracted criteria may be statistical parameters, calculations of some characteristic of the waveform or derivations from multiple other sources. The technique of “feature extraction” is similar to that used in cochlear implants. Herein, we apply it to vestibular function. A major difficutly in measuring biological electrical potentials is the signal-to-noise ratio. We are trying to detect a small signal in the midst of tremendous electrical noise from nerve, muscle and other cells. In this paper we discuss the clinical utility of NEER algorithm and EVestG extracted signals. First we briefly describe of the key concepts. Further details can be found in the Additional file 1.
In signal processing, features are quantities that are associated with a signal or a process. Features may be statistical measures such as the mean, standard deviation, skewness, kurtosis, etc. of a statistical process, or they may be other quantitative measures representing fractal nature, power distribution, etc. of a signal or process. In addition to statistical features, this report includes features representing fractal dimension (FD) as assessed by the Higuchi fractal dimension (HFD), and entropy-based dimensions such as the Information dimension (DI)  and the Correlation dimension (DC). These features were extracted from the FP and timing intervals of the EVestG recordings from patients undergoing EVestG testing.
Fractal dimension calculation (FD) can be interpreted as the “degree of meandering” (roughness, brokenness, irregularity or singularity) of an object. Another interpretation of a fractal dimension is that it is the critical exponent in a power-law relation [10,11]. Fractal dimension mathematically refers to a non-integer or fractional dimension of a self-similar (or a self-affine) object . The self-similarity (or self-affinity) of the object is confirmed if a portion of the object is exactly (or statistically) a scaled down version of itself.
FD analysis is widely used as an analytical tool in a variety of research areas particularly biological signal processing . It measures the irregularity, the complexity and the self-similarity of a signal. The more complex the signal, the higher the FD value will be. Two effective methods for FD calculation are the Higuchi fractal dimension and Entropy based fractal dimension (see Additional file 1 for details).
Higuchi Fractal Dimension (HFD) is well suited for studying signal fluctuation in one dimension . HFD, proposed in 1988, is an efficient algorithm for measuring the FD of discrete time series . HFD has been established as a method to characterize the morphological complexity of biological signals .
Entropy-Based Fractal Dimensions - Entropy can be defined as the amount of information needed to specify the state of a system . Entropy is known as the measure of disorder in physical systems, or an amount of information that may be gained by observations of disordered systems . Entropy-based fractal dimensions can deal with fractals objects which have non-uniform distributions, while the morphological fractal dimensions such as HFD deal with the shape of a projection of the fractal only. This is understandable because the morphological dimensions are purely metric and not probabilistic concepts. The information dimension (DI) and correlation dimension (DC) are special cases related to generalized entropy concept as introduced by Alfred Renyi in 1955 . Both dimensions are improvements of the geometric definition of a fractal object (See Additional file 1).
The DI reveals the expected spread in the non-uniform probability distribution of the fractal objects, but not its correlation. The DC was introduced to address this problem. Both DI and DC represent a weighted average measure of the actual distribution of self-information over the fractal object (See Additional file 1).
Linear Discriminant Analysis (LDA) is a mathematical technique that utilizes features to classify objects or signals into one or more classifications. Each object/signal has certain features that may be relevant in classifying that object/signal; some of these features can be more important predictors than others. In this study, we are trying to classify patients as with either Meniere’s disease or no Meniere’s disease.
Minimal-redundancy-maximal-relevance (mRMR) feature selection method  is a method of ranking features based on the two criteria of minimum redundancy and maximum relevancy; thus allowing to choose the most relevant and least redundant features as the best set of features for classification.
EVestG research labs have been established for human testing at Alfred Hospital in Melbourne Australia and Riverview Health Center in Winnipeg, Canada. In this study, however, only data recorded at Alfred Hospital in Australia has been used. The EVestG signal acquisition apparatus is illustrated in Figure 1.
EVestG data of 14 Meniere’s patients (54.2 ± 9.7 years, 4 males) and 16 healthy individuals (56.1 ± 5.5 years, 7 males) from the EVestG lab at the Alfred Hospital, Melbourne, Australia, were used as the training data to design the diagnostic algorithm. Ethics approval was granted by the Health Research Ethics Board of Alfred Hospital, and all study subjects signed an informed consent form prior to the experiments.
A complete EVestG recording [4,5] includes passive lateral whole body tilts, up/down movements, rotations from the sitting position and up/down movements and rotations from the supine position. This paper will report data for right and left lateral tilts only. Tilts were symmetric movements moving over 3 seconds from the upright sitting position to a position 45 degrees from the vertical to the right, then upright and then to the left. The EVestG signal was recorded at a sampling rate of 41666 Hz.
Labeling of components of EvestG test
Period of Interest (POI) of the segment (See Figure 2 )
Name of POI
20 s background recording
the final 1.5 s
3 s lateral tilt to the right (about 40 degrees)
first half, 1.5 s, the acceleration phase
second half, 1.5 s the deceleration phase
17 s rest in the tilted position
final 1.5 s just before returning to center
3 s returning back to center
the first 1.5 s
the second 1.5 s
17 s rest at the center position before a new tilt
Transition to Steady State
The NEER  algorithm extracts data from the neural response. Each tilt’s recorded data results in two main signals, an average field potential (FP) and its firing pattern for each time segment (see Table 1) for each ear. In this study, we only used signals of the contralateral and ipsilateral side tilt stimuli that presumably stimulate otolith and semicircular canals.
We investigated the changes in the differences among different time segments for each tilt signal to examine the effects of dynamic changes from resting to acceleration or deceleration phases of the time segment, and also the differences between the two phases (acceleration/deceleration) of chair movement as well as differences between the right and left symmetry (L-R).
The NEER algorithm  removes segments of the original signal that are corrupted by large artifact (due to hydraulic chair, muscle artifact, movements, poor electrode contact, etc.); therefore, not all the segments were precisely of 1.5 s duration. We excluded the segments shorter than 1.36 s. Thus, it was possible that not every subject have all the extracted features.
We calculated the mean, mean of the absolute value, variance (Var), skewness, kurtosis, HFD, entropy-based dimensions such as the Information dimension (DI) and the Correlation dimension (DC), the total energy and the average power of the aforementioned intervals for the range of 100–11000 Hz of the pre- and post-potential regions of every FP signal. Also the depth of the AP point was selected as suggested in .
From the time interval of the FP’s firing pattern signals (Figure 3 lower left), we calculated the mean, standard deviation (Std), skewness, kurtosis, mode, median, the DI, DC, HFD. We also calculated the average number of the time intervals less than 0.5 msec, and the correlation of the probability distribution function (PDF) of the time interval signals (Figure 3 lower right) with the relevant FP signals (Figure 3 top) as another feature.
Overall, we calculated over 40 features to consider. The features are grouped in three categories based on which signal they were calculated: 1) the features from from the field potential signals, 2) the features from one of the firing pattern representations, and 3) the features from the correlation calculation between the pdf of the time interval signals and FPs. The names of the features are summarized for the sake of space. For example, the names of “Pre Kurtosis”, “Pre mean abs”, or “Pre Energy” show that the features are found by calculation of the kurtosis, mean of absolute value, or total energy (from the entire frequency range) of the pre potential interval of the mentioned original segment.
A t-test was used to assess the statistical significance of differences for CT tilt and IT tilt between Meniere’s patients and controls. Then, we ran the mRMR algorithm on these 39 statistically significant features, and selected 5 top features (from every tilt) as the best features for classification.
Classification (Average Voting Classifier)
Each selected feature was used in a single feature classifier using linear discriminant classification algorithm (LDA) . We used leave-one-out routine  for training classifiers. Then, we considered each feature as a symptom, and used a heuristic method for a final classification, called Average Voting Classifier, in which every feature has “a vote” for the test subject as either Meniere’s (vote = 1) or non-Meniere’s (vote = 0) based on the LDA classifier, and the final classification is based on the average vote of all the selected best features. In this way, the final vote represents the probability that the test subject is a Meniere’s or non-Meniere’s patient. If that probability is greater than 0.5, the subject is classified as a Meniere’s patient; otherwise as non-Meniere’s.
Five best features for CT (Feature 1–5), and IT (Feature 6–10) tilts
OnBB - R
RTC OnAA - RTC OnBB - L-R
ID of Time Interval Signal
OnBB – L + R
OnBB - L-R
RTC OnBB - R + L
RTC BGi - L
RTC OnAA - R
BGi - OnBB - R
CD of Time Interval Signal
BGi - OnBB - R
RTC OnBB - L
The features that extracted in this study represented some consistent variations. Fractal dimension calculation (DI, and CI) over both firing patterns and FP signals showed higher values for control subjects compared to those of the Meniere’s subjects. This may imply the higher complexity in control subjects’ data compared to Meniere’s patient, which is congruent with the observed pattern of fractal dimension features calculated from other biological signals .
Moreover, the correlation between the FP and the pdf of FP’s firing signal demonstrated positive values in Meniere’s patients, while negative ones were observed in controls. Also, the AP was lower (wider FPs) in Meniere’s patients compared to Control ones. This may talk about possible slower conductivity of the stimulus in the vestibular organ of Meniere’s patients.
Tilts cause angular (rotational) accelerations and changes in the direction of gravity. For these reasons EVestG testing involves a combination of head movement responses that may or may not be consistent among patients with Meniere’s disease. Our long-term goal is to develop an objective test that is diagnostic and specific for Meniere’s disease. Our current tests, electronystagmography, rotary chair, and others, are supportive or helpful, but have no features that are unique to Meniere’s disease. We observed a distinct difference in the pre- and post-potential parts within the period time of the interest of the average FP curve of the Meniere’s and control subjects. Our finding that the fractal dimensions showed more complexity in Meniere’s patients than controls is consistent with the general physiologic literature regarding chaos theory in other organ systems – that abnormal systems lose their variability [20,21].
The results of this study show a new potential of EVestG signals toward generating an adequate set of bio-features as a diagnostic and monitoring aid for dizziness related diseases, especially Meniere’s disease. We suspect, but cannot prove at this point, that our data identify features unique to Meniere’s disease as opposed to some general findings of reduced vestibular function. If EVestG turns out to be a general method of quantifying vestibular function, it should be clinically useful. This is an ongoing study, and we hope to confirm these results with other populations. The results may lead to a simple, objective and non-invasive clinical assessment of Meniere’s disease. We acknowledge that this small dataset is not adequate to recommend clinical use without further development. The method must be tested in larger populations in future studies, which is currently under investigation at the EVestG lab at Winnipeg, Canada.
Modern signal processing techniques such as EVestG may identify neural firing patterns that are diagnostic in patients with vestibular disorders but much more work needs to be done.
This research has been supported by Neural Diagnostics Canada Ltd. and Mathematics of Information Technology and Complex Systems (MITACS).
- Ferraro JA. Electrocochleography: a review of recording approaches, clinical applications, and new findings in adults and children. J Am Acad Audiol. 2010;21:145–52.View ArticlePubMedGoogle Scholar
- Ferraro J, Best LG, Arenberg IK. The use of electrocochleography in the diagnosis, assessment, and monitoring of endolymphatic hydrops. Otolaryngol Clin North Am. 1983;16:69–82.PubMedGoogle Scholar
- Kim HH, Kumar A, Battista RA, Wiet RJ. Electrocochleography in patients with Meniere’s disease. Am J Otolaryngol. 2005;26(2):128–31.View ArticlePubMedGoogle Scholar
- Lithgow B. “A neural response system”, Australia Patent WO/2008/144840. Government of Australia Patent Office. Patent (WO 2008/144840, priority date June 2007)Google Scholar
- Lithgow B. A methodology for detecting field potentials from the external ear canal: NEER and EVestG. Ann Biomed Eng. 2012;40:1835–50.View ArticlePubMed CentralPubMedGoogle Scholar
- Lithgow BJ, Garrett A, Heibert D. “EVestG: a measure for Meniere’s Disease”. Conf Proc IEEE Eng Med Biol Soc. 2008;2008:4162–5.PubMedGoogle Scholar
- Dastgheib ZA, Lithgow B, Moussavi Z. Diagnosis of Parkinson’s disease using electrovestibulography. Med Biol Eng Comput. 2012;50:483–91.View ArticlePubMedGoogle Scholar
- Garrett A, Heibert D, Lithgow B. “Electrovestibulography: the “DC” potential used to separate Meniere’s disease and Benign Paroxysmal Positional Vertigo,”. Conf Proc IEEE Eng Med Biol Soc. 2007;2007:2381–4.PubMedGoogle Scholar
- Haghgooie S, Lithgow BJ, Winograd-Gurvich C, Kulkarni J. “EVestG: a diagnostic measure for schizophrenia,”. Conf Proc IEEE Eng Med Biol Soc. 2008;2008:4142–5.PubMedGoogle Scholar
- Kinsner W. A Unified Approach to Fractal Dimensions. 2009.View ArticleGoogle Scholar
- Phothisonothai M, Nakagawa M. EEG signal classification method based on fractal features and neural network. Conf Proc IEEE Eng Med Biol Soc. 2008;2008:3880–3.PubMedGoogle Scholar
- Sullivan R, Holden T, Tremberger Jr E, Cheung E, Branch C, Burrero J, et al. “Fractal dimension of breast cancer cell migration in a wound healing assay”. Int J Biol Life Sci. 2011;7(Issue 3):170.Google Scholar
- Higuchi T. Approach to an irregular time series on the basis of the fractal theory. Physica D: Nonlinear Phenomena. 1988;31:277–83.View ArticleGoogle Scholar
- Gómez C, Mediavilla A, Hornero R, Abásolo D, Fernández A. Use of the Higuchi’s fractal dimension for the analysis of MEG recordings from Alzheimer’s disease patients. Med Eng Phys. 2009;31:306–13.View ArticlePubMedGoogle Scholar
- Kimcheng K, Sourina O, Kulish V, Nguyen M. An algorithm for fractal dimension calculation based on Renyi entropy for short time signal analysis. In: Information, Communications and Signal Processing, 2009. ICICS 2009. 7th International Conference on. 2009. p. 1–5.Google Scholar
- Rényi A. On a new axiomatic theory of probability. Acta Math Acad Sci Hung. 1955;6:285–335.View ArticleGoogle Scholar
- Peng H, Long F, Ding C. Feature selection based on mutual information: criteria of max-dependency, max-relevance, and min-redundancy. IEEE Trans Pattern Anal Mach Intell. 2005;27:1226–38.View ArticlePubMedGoogle Scholar
- Duda RO, Hart PE, Stork DG. Pattern classification. New York: Wiley; 2001.Google Scholar
- Ehtiati T, Kinsner W, Moussavi ZK. Multifractal characterization of the electromyogram signals in presence of fatigue. In: University and Industry - Partners in Success, Conference Proceedings Vols 1–2. 1998. p. 866–9.Google Scholar
- Halley JM, Kunin VE. Extinction risk and the 1/f family of noise models. Theor Popul Biol. 1999;56:215–30.View ArticlePubMedGoogle Scholar
- Kinsner W. A unified approach to fractal and multifractal dimensions. Winnipeg, Manitoba, Canada DEL94-4: University of Manitoba; 1994.Google Scholar
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.