Edited by: John J. Foxe, Albert Einstein College of Medicine, USA
Reviewed by: John S. Butler, Albert Einstein College of Medicine, USA; Ming-Zher Poh, Cardiio, Inc., USA
*Correspondence: Christian Duval, Département de Kinanthropologie, Pavillon des Sciences Biologiques, 141 Avenue du Président-Kennedy, Room SB4290, Montréal, QC, Canada H2X 1Y4. e-mail:
This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in other forums, provided the original authors and source are credited and subject to any copyright notices concerning any third-party graphics etc.
Tremor is the most prevalent movement disorder, and can manifest itself in a myriad of pathologies. One of those disorders is Parkinson’s disease (PD) (Calabrese,
The treatment and monitoring of tremor still represents a significant challenge for clinicians as tremor is highly variable in its characteristics within a day and over several days. Currently, tremor is assessed during clinic visits, mostly with the use of clinical rating scales (see Goetz et al.,
In a recent editorial in Nature methods, it was proposed that smart phones could be used to gather data in the laboratory (No authors listed,
To test the real potential of smart phones to assess, characterize, and monitor abnormal tremors, we developed a protocol in two parts. We first determined whether the device could act as a standalone platform for detection and analysis of several tremor characteristics, and whether values obtained with the smart phone were similar to those obtained with a commonly used laboratory measurement tool; i.e., an accelerometer. Secondly, we sought to compare values obtained by the smart phone against those of a clinical evaluation.
In the first part of the study, the objective was to compare a smart phone [Blackberry® Storm™ 9530 (Research In Motion, Ltd., Waterloo, ON, Canada)] to a tremor assessment method normally used in laboratories. In order to do so, tremor was assessed simultaneously with the smart phone and a laboratory accelerometer. One of the authors (Jean-François Daneault) simulated 192 trials of tremor of different amplitudes and frequencies in different conditions (i.e., rest tremor, postural tremor, kinetic tremor, and intention tremor). Analysis of data was then carried out in two steps. In
In
Both time- and frequency-domain properties of tremor, such as tremor amplitude, tremor regularity, power distribution (percentage of power within the 3–7 Hz frequency band), median power frequency, peak power frequency, power dispersion (frequency band containing 68% of total power centered at the median power frequency), power dispersion centered at peak power frequency, and harmonic index were examined. All these measures are known to help categorize abnormal tremors and provide detailed tremor characteristics (Beuter and Edwards,
Bias | SD | CCC | ||||
---|---|---|---|---|---|---|
RMS | Rest | 1.00 | 0.00 | 0.000 | 0.000 | 1.00 |
Post | 1.00 | 0.00 | 0.000 | 0.000 | 1.00 | |
Kin | 1.00 | 0.00 | 0.000 | 0.000 | 1.00 | |
Intention | 1.00 | 0.00 | 0.000 | 0.000 | 1.00 | |
Reg. | Rest | 1.00 | 0.00 | 0.000 | 0.001 | 1.00 |
Post | 1.00 | 0.00 | 0.000 | 0.001 | 1.00 | |
Kin | 1.00 | 0.00 | 0.000 | 0.001 | 1.00 | |
Intention | 1.00 | 0.00 | 0.000 | 0.001 | 1.00 | |
Pow.Dist | Rest | 0.98 | 0.00 | 0.723 | 4.540 | 0.98 |
Post | 0.99 | 0.00 | −0.743 | 2.557 | 0.99 | |
Kin | 0.99 | 0.00 | −0.246 | 3.813 | 0.99 | |
Intention | 0.94 | 0.00 | 1.466 | 7.877 | 0.94 | |
MPF | Rest | 0.99 | 0.00 | 0.016 | 0.194 | 0.99 |
Post | 0.99 | 0.00 | −0.031 | 0.138 | 0.99 | |
Kin | 0.98 | 0.00 | −0.077 | 0.242 | 0.98 | |
Intention | 0.98 | 0.00 | −0.054 | 0.213 | 0.98 | |
HI | Rest | 0.96 | 0.00 | 0.026 | 0.027 | 0.96 |
Post | 0.96 | 0.00 | 0.029 | 0.033 | 0.96 | |
Kin | 0.97 | 0.00 | 0.023 | 0.019 | 0.97 | |
Intention | 0.95 | 0.00 | 0.023 | 0.030 | 0.95 | |
Peak | Rest | 0.73 | 0.00 | −0.115 | 1.157 | 0.73 |
Post | 0.86 | 0.00 | −0.096 | 0.933 | 0.86 | |
Kin | 0.95 | 0.00 | −0.020 | 0.540 | 0.95 | |
Intention | 0.94 | 0.00 | 0.093 | 0.478 | 0.94 | |
Disp. | Rest | 0.99 | 0.00 | 0.039 | 0.348 | 0.99 |
Post | 1.00 | 0.00 | 0.031 | 0.216 | 1.00 | |
Kin | 0.98 | 0.00 | 0.189 | 0.510 | 0.98 | |
Intention | 0.99 | 0.00 | 0.059 | 0.304 | 0.99 | |
Disp.Peak | Rest | 0.98 | 0.00 | −0.216 | 0.767 | 0.98 |
Post | 0.98 | 0.00 | −0.242 | 0.692 | 0.98 | |
Kin | 0.97 | 0.00 | −0.070 | 0.695 | 0.97 | |
Intention | 0.99 | 0.00 | −0.125 | 0.596 | 0.99 |
Our results show correlation coefficients above 0.80 for time-domain tremor properties; namely tremor amplitude and tremor displacement regularity (see results in Table
Without threshold |
With threshold |
|||||||
---|---|---|---|---|---|---|---|---|
Bias | SD | CCC | ||||||
RMS | Rest | 0.99 | 0.00 | 0.99 | 0.00 | 0.024 | 0.017 | 0.99 |
Post | 0.98 | 0.00 | 0.98 | 0.00 | 0.022 | 0.044 | 0.98 | |
Kin | 0.99 | 0.00 | 0.99 | 0.00 | 0.071 | 0.032 | 0.99 | |
Intention | 0.99 | 0.00 | 0.99 | 0.00 | 0.066 | 0.061 | 0.99 | |
Reg. | Rest | 0.92 | 0.00 | 0.95 | 0.00 | −0.009 | 0.053 | 0.95 |
Post | 0.81 | 0.00 | 0.90 | 0.00 | 0.033 | 0.060 | 0.90 | |
Kin | 0.88 | 0.00 | 0.88 | 0.00 | −0.004 | 0.046 | 0.88 | |
Intention | 0.91 | 0.00 | 0.98 | 0.00 | 0.018 | 0.039 | 0.98 | |
Pow.Dist | Rest | 0.96 | 0.00 | 0.97 | 0.00 | −9.266 | 8.215 | 0.97 |
Post | 0.97 | 0.00 | 0.98 | 0.00 | −9.044 | 9.121 | 0.98 | |
Kin | 0.97 | 0.00 | 0.97 | 0.00 | −7.137 | 7.208 | 0.97 | |
Intention | 0.96 | 0.00 | 0.92 | 0.00 | −11.259 | 6.476 | 0.92 | |
MPF | Rest | 0.82 | 0.00 | 0.99 | 0.00 | −0.103 | 0.210 | 0.99 |
Post | 0.84 | 0.00 | 0.95 | 0.00 | −0.244 | 0.452 | 0.95 | |
Kin | 0.59 | 0.00 | 0.59 | 0.00 | −0.828 | 1.268 | 0.59 | |
Intention | 0.92 | 0.00 | 1.00 | 0.00 | −0.160 | 0.069 | 1.00 | |
HI | Rest | 0.92 | 0.00 | 0.92 | 0.00 | 0.003 | 0.010 | 0.92 |
Post | 0.77 | 0.00 | 0.90 | 0.00 | 0.001 | 0.012 | 0.90 | |
Kin | 0.81 | 0.00 | 0.81 | 0.00 | 0.023 | 0.034 | 0.81 | |
Intention | 0.89 | 0.00 | 0.89 | 0.00 | 0.004 | 0.007 | 0.89 | |
Disp. | Rest | 0.75 | 0.00 | 0.55 | 0.00 | |||
Post | 0.85 | 0.00 | 0.81 | 0.00 | ||||
Kin | 0.81 | 0.00 | 0.81 | 0.00 | ||||
Intention | 0.96 | 0.00 | 0.89 | 0.00 | ||||
Disp.Peak | Rest | 0.73 | 0.00 | 0.52 | 0.00 | |||
Post | 0.81 | 0.00 | 0.83 | 0.00 | ||||
Kin | 0.79 | 0.00 | 0.79 | 0.00 | ||||
Intention | 0.92 | 0.00 | 0.87 | 0.00 |
To verify this, we removed all trials having tremor amplitude below 1 mm from the correlation analysis (see Table
As such, the smart phone application could be considered as an adequate measurement tool to provide power distribution, median power frequency, and harmonic index of tremors having amplitude above 1 mm while not performing a kinetic task.
To address the objective of this part of the study, we opted to create and validate a clinical rating scale for tremor amplitude (see Table
The purpose of this experiment was to characterize the properties of the clinical scale used to assess tremor amplitude of the hand. The scale incorporated a six level ordinal scale; each associated with a predefined tremor amplitude (0 = no visible tremor; 1 = up to 1.5 cm; 2 = 1.5–3 cm; 3 = 3–4.5 cm; 4 = 4.5–6 cm; and 5 = 6 cm and above) | |||
Group 1 | Group 2 | ||
Rater 1 | 0,967 (0.000) | 0,963 (0.000) | |
Rater 2 | 0,954 (0.000) | 0,968 (0.000) | |
Rater 3 | 0,907 (0.000) | 0,969 (0.000) | |
Rater 4 | 0,951 (0.000) | 0,966 (0.000) | |
Rater 5 | 0,957 (0.000) | 0,965 (0.000) | |
Validity of the clinical scale was assessed against a laboratory accelerometer. As such, tremor was assessed simultaneously with both instruments by a group of five raters. This was repeated with a second group of five raters. Pearson correlations were performed between the clinical score and the result obtained from the accelerometer ( |
|||
Sensitivity of the clinical scale was assessed again against a laboratory accelerometer. Trials were grouped by clinical score and the results obtained from the accelerometer for contiguous clinical scores were compared using repeated |
|||
Group 1 | Group 2 | ||
Intraclass correlation | 0,943 | 0,963 | |
95% Confidence interval | 0,912–0,965 | 0,945–0,976 | |
Inter-rater reliability was assessed using an intraclass correlation from the data of the two groups of five raters with the corresponding 95% confidence interval | |||
The results presented above demonstrate that the clinical scale used to evaluate hand tremor in the current study is valid, provides good sensitivity and has very good inter-rater reliability. As such, the use of this scale to provide a clinical score of hand tremor is appropriate as per the study parameters |
Results presented in Table
Rest | RMS | 0.762 | <0.000 |
Postural | RMS | 0.851 | <0.000 |
Intention | RMS | 0.880 | <0.000 |
Kinetic | RMS | 0.086 | 0.557 |
Pow.Dist | 0.700 | <0.000 |
Mobile technology is currently used to transmit or store data obtained from laboratory instruments (Barroso Junior et al.,
The current study provides a comparison between the use of a smart phone application and laboratory tools for the characterization of tremor. The results demonstrate that the smart phone application can provide similar results to laboratory tools for measurements of time-domain and spectral characteristics albeit with some limitations. While the smart phone always provides valid tremor amplitude values, evaluation of spectral characteristics and regularity of tremor requires an amplitude threshold above 1 mm. This smart phone is therefore not suitable for characterizing the spectral characteristics of normal physiological tremor, but can be used to monitor pathological tremor characteristics. This is expected since accelerometers used in laboratory settings are highly sensitive to low amplitude oscillation but can be deleteriously affected by shocks, which would be impractical in smart phones. Furthermore, our results demonstrate that spectral characteristics of tremor should currently not be assessed during movement. The exact mechanism behind the poorer results obtained during movement is not known, but we hypothesize that it could stem from crosstalk between the three axes of the smart phone accelerometer, which would be absent from the one-axis laboratory accelerometer.
There are numerous ongoing studies examining pathological tremor worldwide. The use of a portable smart phone based measurement tool could therefore provide much needed information on the characteristics of tremor beyond what can currently only be gathered in laboratory setting. Amongst many applications for tremor evaluation, smart phones could provide valid measures of prevalence of ET in different populations, help determine the effect of different medications on tremor frequency and amplitude in PD over the patient’s day, and evaluate the prevalence of tremor as a side-effect of certain medications.
In addition to providing confirmation of the usefulness of smart phones for tremor research, the current study provides for the first time evidence of clinical utility for these devices in tremor disorders. We found a strong relationship between the amplitude of tremor measured by the smart phone application and the amplitude of tremor measured by a clinical rating scale. Furthermore, mean tremor amplitude recorded with the smart phone application for each level of clinical scale rating was significantly different, which indicates that clinicians could obtain a valid profile of tremor severity over time. This demonstrates that the smart phone can provide a reliable measure of tremor of different amplitudes. This is important as it is well known that while not all PD patients exhibit tremor, those that have tremor can exhibit highly variable tremor amplitude and that this amplitude can fluctuate over short periods of time (Duval et al.,
The current study examined the feasibility of evaluating tremor using a Blackberry® smart phone. While the data presented in this study clearly demonstrates that several tremor characteristics can be assessed using this device, and that the results provided are clinically meaningful, the algorithms have solely been tested on one Blackberry® model. Indeed, other Blackberry® devices are equipped with the hardware necessary for the proposed applications. We have not yet verified whether other models of Blackberry® smart phones or phones from other manufacturers may indeed be able to characterize the spectral properties of tremor having amplitude below 1 mm. While the feasibility of such applications was briefly demonstrated on the iPhone® (see above), to our knowledge there are currently no such studies on other widely used platforms such as Android™, Java ME, and Windows Phone®. The implementation of the method demonstrated in the current study on all major platforms is essential in order to maximize the impact of such a tool. We have no reason to believe that other platforms could not be used for tremor assessment, except if their hardware or software configuration does not allow access to the accelerometer data. It is quite possible that accelerometer chips within smart phones may soon be as precise as laboratory tools and because of their relative affordability, smart phones may eventually replace some laboratory assessment tools.
The current study provides the framework for the use of smart phones for the evaluation of motor behavior. While we presented data on the assessment of tremor, with the proper algorithms, a myriad of other movements, voluntary or involuntary, in health and disease, can be quantified using these devices. The only barrier to the use of smart phones in research settings is creativity and the ability to implement the proper software within the devices to achieve the set goals. The implementation of smart phones in medical settings clearly requires more research since it will directly impact the lives of patients. One foreseeable avenue from the data presented here is the development of applications tailored to the long-term monitoring of patients with tremor. These applications could incorporate testing schedules, data transmission to the physician for remote monitoring, and many other useful specifications. The goal of these applications would be to provide meaningful information to the clinician in order to improve personalized care and reduce burden on the health-care system. Such devices could also greatly improve the efficiency of clinical trials in which tremor monitoring is needed.
The current study demonstrates for the first time that, with the proper analytical algorithms, smart phones can be used not only for data storage during experiments, but also as a data gathering and analysis tools. Furthermore, we demonstrate that smart phones can provide meaningful and powerful data for clinical evaluations. Smart phones may therefore revolutionize scientific research and greatly improve patient care.
The analysis of this part of the study is subdivided into two section;
Jean-François Daneault performed the tests. He is free of any neurological disorders that could have influenced movements or understanding of the tasks. Jean-François Daneault is right-handed according to the Edinburg Handedness Inventory. This protocol was approved by the ethics board of the Université du Québec à Montréal.
Tremor was assessed using a Blackberry® Storm™ 9530 (Research In Motion, Ltd., Waterloo, ON, Canada). A choice was made to code the required algorithms for the Blackberry® platform as several were readily available in the laboratory for testing and it was the programming language in which we were the most proficient. Furthermore, the security provided by this platform for storing and transmitting sensitive information that could be medically relevant is currently much greater than on other platforms. Finally, tremor was also quantified using a one-axis accelerometer (TLB333B42, PCB Piezotronics, NY, USA) which was affixed to the back of the smart phone.
Tremor was recorded simultaneously with both methods. The accelerometer was fixed to the back of the smart phone using two-sided tape. Smart phone data acquisition rate was set at 60 Hz, and the accelerometer acquisition rate was set at 2048 Hz. For the purpose of the current experiment tremor was always recorded in the same axis (i.e., front-to-back axis of the smart phone). Four tasks were performed by the participant while he was seated: (A) Resting tremor. This task consisted of having the participant sitting with his arm hanging by his side as tremor was being recorded. (B) Postural tremor. This task consisted of having the participant keep his arm and hand outstretched in front of him and parallel to the ground. (C) Intention tremor. This task consisted of having the participant keep his arms and hands in front of him while trying to bring the tips of his fingers as close as possible to each other. (D) Kinetic tremor. This task consisted of starting in the same position as in B and then, bringing the phone to one’s ear and back at a relatively slow velocity. During each task, the participant held the smart phone in his hand. We opted for this approach instead of taping or strapping the smart phone to the participant’s hand because we wanted a more ecological design where they would hold the phone as they would in real life. The participant performed 48 trials of every task while simulating tremors of different amplitude and frequency. No specific instruction on tremor amplitude and frequency was given for each trial; only that after completion, there should be a high variability in tremor amplitude and frequency between trials. Tremor recordings were coded to last 10 s, while only the last 8.5 s of recording were taken into account for analysis in order to minimize the impact of any movement that could have occurred as a result of placing the hand in the required position. A 1-s vibration (∼140 Hz) indicated to the participant the end of each trial. Ten seconds were allotted between trials to minimize fatigue.
The analysis algorithms usually employed in our laboratory, using the S-Plus software (Mathsoft, Seattle, WA, USA), were coded in Java™ in order to run on the Blackberry® operating system. Each trial yielded two files of data on the smart phone. One containing the raw time series of a given trial and another containing the results obtained by the algorithms implemented in the smart phone. The time series was analyzed with both the smart phone algorithms and the algorithms usually employed within the laboratory to obtain tremor characteristics, such as: tremor amplitude, tremor regularity, power distribution, median power frequency, peak power frequency, power dispersion, power dispersion centered at peak power frequency, and harmonic index. Each characteristic and their computation are described below:
First, to remove the influence of gravity, the time series was demeaned. Then, a root mean square was applied to the signal.
First, the original time series was normalized, i.e., the mean of the time series was removed from each point and then, each point was divided by the SD of the time series. After, the signal was divided into epochs of 1 s and the amplitude (root mean square) was calculated for each epoch. Finally, the total SD from the amplitude of all epochs was computed. This yielded a measure of signal amplitude stability over time. A more regular signal was associated with a lower value.
Spectral characteristics of tremor were then evaluated. To do so, a fast-Fourier transform (FFT) was performed on the time series. Codes were written in such a way that the power spectral density function yielded only the power lying between 1 and 20 Hz as these are the prominent frequencies of tremor.
Represents the sum of the power within a specific frequency band located between 3 and 7 Hz; divided by the total power. This frequency band harbors the majority of power in most pathological tremors (McAuley et al.,
Represents the frequency where 50% of the power lies below it and the remaining 50% lies above it.
Represents the frequency where the maximum power was observed.
Represents the width of a frequency band containing 68% of total power; centered at the median power frequency.
Represents the width of a frequency band containing 68% of total power; centered at the peak power frequency.
Represents a ratio considering a rectangle bounded on the sides by the frequency band of interest (0–20 Hz), and vertically from 0 to the height of the highest peak. The harmonic index is the proportion of the area of this rectangle lying above the power spectrum itself.
While the coding of the algorithms mentioned above provide the analytical basis of the application, several lines of codes were also required to implement the graphical user interface, to assign a specific id to each recording, to record files in appropriate folders on the phone, and many more functional issues that are required when using a smart phone application. These will not be discussed in detail here as they can be modified to suit the needs of each study and do not directly impact the data being analyzed.
In order to identify whether the algorithms written for the smart phone provide the same results as those used in the laboratory, a Pearson’s correlation was performed on the data. The results provided by the smart phone were compared to results obtained by analyzing the raw time series of the smart phone offline using custom-designed algorithms within the S-Plus software (Mathsoft, Seattle, WA, USA). These analyses have been used on several occasions by our group to characterize tremor in healthy and pathological populations (see Duval et al.,
Next, Bland–Altman analyses were performed between data from the smart phone and the accelerometer for each variable (bias and SD of the difference between both methods of analysis).
Finally, concordance correlation coefficients (CCC) were computed between data from both methods. These last two analyses were performed to identify whether there was a good agreement (reproducibility) between both tremor analysis methods.
Note that for this part of the study, the data from the accelerometer was compared to the results calculated by the smart phone; not from post-processing of the phone time series. Time series from the accelerometer were recorded on a computer for post-processing.
First, in order to identify whether the data recorded by the phone was congruent with measures obtained with the accelerometer, the time series obtained with the accelerometer were analyzed using the S-Plus software (Mathsoft, Seattle, WA, USA). Data from the accelerometer were down-sampled to 60 Hz using a moving average. The same analysis performed in
In order to identify whether the results stemming from the algorithms written for the smart phone are significantly correlated with results stemming from the analysis of the time series of a laboratory accelerometer, a Pearson’s correlation was performed. First, a Pearson’s correlation was performed on all trials pooled together. Then, a Pearson’s correlation was performed on trials within each condition.
Then, to assess whether a threshold of tremor amplitude would improve the correlation between instruments, the same correlations were performed only on trials having tremor amplitude above 1 mm.
Next, Bland–Altman analyses were performed between data from the smart phone and the accelerometer for each variable showing high correlation between methods (i.e., RMS, regularity, MPF, distribution, and harmonic index).
Finally, CCC were computed between data from both methods for the same variables showing high correlations. These last two analyses were performed to identify whether there was a good agreement (reproducibility) between both tremor quantification methods.
The objective of this part of the study was to evaluate whether there was a relationship between the results from the smart phone application and a clinical scale to evaluate tremor amplitude.
Sixteen patients were recruited for this experimentation. Twelve patients were diagnosed with idiopathic PD. Three patients were diagnosed with ET and one patient was diagnosed with multiple sclerosis. When patients exhibited clinically visible tremor within one limb, tremor recordings were performed on that limb otherwise, tremor was recorded on their dominant side. The experimental protocol was approved by the institutional ethics board of the Montreal Neurological Hospital and Institute. Fourteen patients were right-handed and two were left-handed.
Tremor was concurrently evaluated using two methods: the smart phone and a clinical scale performed by Jean-François Daneault. The clinical scale was custom-designed to evaluate tremor amplitude of the upper limb. The rater was asked to assign a value ranging from 0 to 5 to the participant’s tremor. Each value was representative of a specific tremor amplitude [0 = no visible tremor; 1 = up to 1.5 cm (roughly the equivalent of one finger width based on a previous study; Peters et al.,
Patients were asked to perform the four tasks that were evaluated in
Tremor amplitude from the smart phone was analyzed using the same method as described in
In order to identify whether the amplitude results stemming from the smart phone are significantly correlated with results stemming from a clinical rating scale, a Pearson’s correlation was performed within each condition. Then, we evaluated whether the smart phone was able to detect clinical differences. For this, the amplitude of each trial assigned a given clinical score were pooled. Then, the mean tremor amplitude was computed for each clinical rating (0–5). The mean value for each contiguous score was compared using a
Jean-François Daneault, Benoit Carignan, Carl Éric Codère, and Christian Duval have partial ownership of Medapplets, Ltd.; a company created for the development of medical applications for mobile devices. This company has not commercialized any application for mobile devices as of the date of manuscript submission. No commercial application has emerged from the data presented in this manuscript as of the date of submission.
The authors of the present study wish to thank the participants who volunteered their time for this study. This research was funded by Natural Science and Engineering Research Council of Canada through a Doctoral scholarship (Carignan) and operating grant (Duval). This research was also made possible by a Fonds de la Recherche du Québec-Santé Doctoral scholarship (Daneault). Dr. Duval is supported by a Fonds de la Recherche du Québec-Santé salary grant.
Jean-François Daneault, Benoit Carignan, and Christian Duval designed and performed experiments, analyzed data, and wrote the manuscript; Carl Éric Codère provided technical guidance, created the smart phone application, adapted the code to the device, and helped with manuscript editing; Abbas F. Sadikot recruited patients, advised on clinical issues, and contributed to manuscript editing.
1Smart Phone Sales Statistics. According to figures for 2010, smart phones accounted for 297 million (19%) of the 1.6 billion mobile phones sold that year (