Discrimination of physiological tremor from pathological tremor using accelerometer and surface EMG signals

1. Introduction

Tremor is defined as a rapid involuntary back-and-forth or up-and-down movement of a body part and is a readily apparent and easily quantified motor phenomenon found both in normal individuals and as a pathological symptom [1]. Physiological tremor (PH) is the most well-known form of tremor; it has a mechanical component that consists of the damped oscillation, which is driven by several factors: resonance at the Eigen-frequency, cardio ballistic effects, and the discontinuities of innervation [2]. The resonance frequency can be identified by spectral analysis of the accelerometer signal from the affected extremity or the EMG from the antigravity muscle [2].

The two most frequent and common pathological tremor forms encountered in clinical neurology are essential tremor (ET) and Parkinson’s disease (PD)-related tremor [3, 4]. It has been shown that up to 20% of those two tremors are misdiagnosed during the early years of the disease [3].

The key features that help clinicians to separate the different tremor entities are the tremor frequency, the conditions under which the tremor is activated, the medical and family history (e.g., hereditary tremor, factors influencing the tremor) and the neurological examination (e.g., akinesia, rigidity, peripheral neuropathies) [2]. Those criteria are mostly good enough to separate the conditions. However, especially in the beginning of the disease, significant errors may occur.

There is an overlap in the clinical presentation between ET and PD that can make the differentiation difficult on purely clinical grounds [5, 6, 7, 8]. In such overlap cases, SPECT (Single-Photon-Emission Computer Tomography) technology (FP-CIT) is used which comes with radiation. Note that careful pre-administration of iodine can limit the radiation. Therefore, classical spectral analysis of tremor time-series recorded by accelerometer and surface EMG has been used to replace SPECT [9, 10, 11].

During visual assessments, there are some groups in an essential tremor that, at one extreme, are misdiagnosed as physiological tremor and, at the other extreme, are misdiagnosed as Parkinsonian tremor. Thus, there is a need for neurophysiological assessment based on spectral analysis of accelerometer and sEMG signals [12].

The spectral analysis based on fast Fourier transform (FFT) is most often used. However, FFT has two limitations: first, it assumes that tremor is a linear and periodic signal, which is not the case; second, it cannot discriminate between different tremors, when their frequency spectrum overlap, such as essential tremor and physiological tremor [13].

To increase the efficiency of discrimination, methods beyond the standard spectral analysis of the recorded tremor time-series have been applied to separate different pathological tremors [3, 14, 15, 16, 17, 18, 19].

In this study we have used a wavelet-based soft-decision algorithm to estimate the power spectral density of both accelerometer and surface EMG signals to discriminate between PH, ET and PD [19]. It is concluded that the power spectra of short (30 s) EMG and accelerometer time-series recorded in a laboratory setting carry important information in the different frequency bands and this information could be sufficient for the discrimination between different tremors during diagnosis.

2. Methods and patients

3. Results

An efficient discrimination technique between physiological tremors and pathological tremors (essential and Parkinsonian tremors) based on the computation of the approximate power entropy of the soft-decision wavelet-based frequency bands is implemented in this work. The technique uses as a classification factor the sum of power entropies of bands 6 and 11 out of 256 bands (obtained by 8-decomposition stages) estimating the power-spectral density of the accelerometer and two EMG signals. An accuracy of discrimination between PH and ET on test data is found to be 89.6%. The same accuracy value of discrimination of 89.6% is obtained between PH and PD. An excellent accuracy of discrimination between PH and (ET or PD) of 93.8% is obtained on test data and using a voting between the three signals (accelerometer, EMG1, and EMG2).

It was found that the two frequency bands 6 and 11 have to be used in the discrimination, covering the ranges between 7.8125 and 9.375 Hz, and between 15.625 and 17.1875 Hz, respectively. These bands are very close to the frequency regions in which the first and second harmonic peaks are located. In [26, 27], it has also been observed that the harmonic peaks in PD tremor may be also related to a second tremor generator at twice the tremor frequency. Such observations have not been encountered to the same extent in ET patients [28, 29], and this could thus be an explanation for the differences found in these frequency bands in our present results. It is known that the harmonic peaks are typically broader than those at the basic tremor frequency and may well cover also the frequency bands found in the present study.

Regarding the tremor frequency, tremors can be classified into three main types: low-frequency tremor (less than four cycles per second or Hertz), middle frequency-tremor (between 4 and 7 Hz), and high-frequency tremor (more than 7 Hz) [30]. Physiological tremor usually presents high frequency (8–12 Hz). Essential tremor reaches frequencies between 6–12 Hz, while tremor from Parkinson’s disease usually has frequencies between 4 and 6 Hz [30]. So the frequency bands that we investigated may cover important information about both PH and pathological tremors (PD or ET). According to previous literature [31, 32] the frequencies overlap between ET and PH [13]. For all three tremors analyzed here, there are hints that they are governed by a central oscillator which is shown earlier with electro-physiological studies in PD and ET [33] and in PH [34]. Most of the previous literature in PH had recorded tremor using an accelerometer signal only [35] and restricted the analyses to spectral power of the accelerometer signals [36, 37, 38, 39, 40, 41, 42, 43, 44]. All of them found spectral power in high frequency components. The accelerometer signal, which is a second derivative of displacement, emphasizes high-frequency components [44]. To get comprehensive analyses of the tremor in this study, we looked at both the accelerometer and EMG signals from flexor and extensor in all the patients. Recently also shown in a retrospective cohort of 210 tremor patients [45], the sensitivity increases to 84% when three features are combined, namely, the effect of loading, tremor frequency, and frequency variability, pointing to a voting system of at least 2 positive tests out of 3 for clinical diagnostic purposes.

In summary, a weighted-voting process between the results of the three signals according to their efficiency in discrimination between PH and ET & PD of the trial set can be also used to discriminate between the tremors in the test stage. Such a process may assume different effects of the three signals on the final decision and may have better results than the normal voting approach applied in this paper, especially if it is used with a larger data size. Combination of time-domain parameters and frequency-domain parameters could be a good solution for better discrimination of a larger data size, too. For this purpose, a neural network may be a good solution to be used for trial and feature matching steps.