This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR mHealth and uHealth, is properly cited. The complete bibliographic information, a link to the original publication on http://mhealth.jmir.org/, as well as this copyright and license information must be included.
Hand tremor typically has a negative impact on a person’s ability to complete many common daily activities. Previous research has investigated how to quantify hand tremor with smartphones and wearable sensors, mainly under controlled data collection conditions. Solutions for daily real-life settings remain largely underexplored.
Our objective was to monitor and assess hand tremor severity in patients with Parkinson disease (PD), and to better understand the effects of PD medications in a naturalistic environment.
Using the Welch method, we generated periodograms of accelerometer data and computed signal features to compare patients with varying degrees of PD symptoms.
We introduced and empirically evaluated the tremor intensity parameter (TIP), an accelerometer-based metric to quantify hand tremor severity in PD using smartphones. There was a statistically significant correlation between the TIP and self-assessed Unified Parkinson Disease Rating Scale (UPDRS) II tremor scores (Kendall rank correlation test: z=30.521,
Our work demonstrates the potential use of smartphone inertial sensors as a systematic symptom severity assessment mechanism to monitor PD symptoms and to assess medication effectiveness remotely. Our smartphone-based monitoring app may also be relevant for other conditions where hand tremor is a prevalent symptom.
Parkinson disease (PD) is a neurodegenerative condition that affects patients’ physical and mental health [
Among patients with PD, approximately 75% suffer from rest tremor, around 50% from moderately severe postural tremor [
Dyskinesia is defined as involuntary movement, different from tremor, and is related to the timing and dosage of levodopa medication [
Accelerometer data have been used to assess hand tremor using smartphones [
Bazgir et al [
Although these previous experiments have had positive results on quantifying tremor, we believe that the utility of the findings outside of the laboratory or a health care facility is limited. The practicality of carrying and wearing a glove at all times is up for debate, especially under extreme weather conditions. Accordingly, a smartphone-only solution was first investigated by Woods et al [
In our study, the data were collected in naturalistic settings using a mobile toolkit, the Sentient Tracking of Parkinson’s (STOP) app, for monitoring PD symptoms in daily life. STOP includes a gamified tremor assessment module based on a ball-balancing game that logs the smartphone’s accelerometer, gyroscope, and rotation data. STOP also provides users with a medication intake journal and a daily symptom survey mechanism [
We installed the STOP app into the smartphones of 13 participants diagnosed with PD and collected accelerometer data and medication logs. We used the Welch method to generate the power spectral densities (PSDs) and extracted features from the accelerometer data that we used to investigate the feasibility of hand tremor assessment and medication effect.
STOP is a smartphone app developed for people with PD with four core functionalities: (1) an accelerometer-based ball game for quantifying patients’ hand tremor, (2) a medication journal for logging medication intake times, (3) a daily survey for reporting the overall severity of PD symptoms, and (4) reminder notifications [
To play the ball game, one has to place the smartphone horizontally on the palm of the hand for 10 seconds and try to keep a virtual ball inside a circle at the center of the screen. During the game session, STOP logs data from the accelerometer, linear accelerometer (acceleration without gravity’s influence), gyroscope, and rotation vector sensors [
During a real-world trial of STOP, data were collected from 13 participants with PD, eight females and five males [
Participants from Finland were located around the country, and their consent to participate in the study was given via the application. Participants from the United Kingdom, on the other hand, signed a paper consent form. Following local guidelines, approval from the University of Oulu’s ethical committee was not needed because the risks associated with participating in the study were similar to those of daily smartphone use. In previous publications, we have shared users’ experience during the trial and an analysis of the interview data [
Overview of participants’ characteristics.
Characteristics | Participants (n=11) | ||
Age (years), mean (range) | 64.7 (52-73) | ||
Years since PD diagnosed, mean (range) | 7.1 (2-17) | ||
Number of PD medications, mean (range) | 3 (1-5) | ||
Number of total daily medicationsa, mean (range) | 4.3 (1-7) | ||
UPDRS II scoreb, mean (range) | 11.8 (3-31) | ||
Tremor on UPDRS, mean (range) | 1.2 (0-3) | ||
Deep brain stimulator installed, n (%) | 2 (18%) | ||
Suffer from hand tremor, n (%) | 4 (36%) | ||
Plays with tremor-affected hand, n (%) | 2 (18%) | ||
Suffers from other issues affecting hands, n (%) |
|
||
|
Rigidity | 3 (27%) | |
|
Bradykinesia | 1 (9%) |
aRefers to the number of medication intake times (ie, how many times per day the participant has to take medications, one or several at a time).
bThe Unified Parkinson Disease Rating Scale (UPDRS) II score quantifies the severity level of Parkinson disease (PD) symptoms affecting daily activities (maximum score of 52). The scale for the tremor item on the UPDRS is as follows: 0=no tremor, 1=slight and infrequently present tremor, 2=moderate and bothersome tremor, 3=severe tremor interfering with many activities, and 4=marked tremor interfering with most activities.
Our data set contained a total of 1856 medication logs (mean of 107 [SD 54.9] logs per participant) and 2213 game sessions (mean of 138 [SD 60.6] sessions per participant). These data were recorded in 13 participants (P01 to P13) in naturalistic conditions. Participants had varying medication regimens.
Game sessions were 10 seconds long. We excluded P03’s sessions because the accelerometer sampling rate of his smartphone was approximately 25 Hz instead of the desired 50 Hz, and we excluded P04’s sessions because they only contained one sensor sample throughout the entire game for unknown technical reasons. P05’s phone had data synchronization issues, so only 1 week of data was collected, and P07 missed the first week of data collection because he had problems installing the application. Despite this, we included P05 and P07 in the analysis, resulting in a total of 11 participants.
Because the data were collected during a trial deployment of the STOP app, there were no participant exclusion criteria related to PD symptom severity. Based on the UPDRS II tremor self-reports, we categorized the participants into five groups: (1) all participants: P01, P02, P05, P06, P07, P08, P09, P10, P11, P12, and P13; (2) no tremor (participants reported no tremor): P02, P09, and P11; (3) tremor (participants reported tremor but in an unspecified location): P06, P07, P08, and P12; (4) hand tremor (participants reported hand tremor and played with the unaffected hand): P01 and P05; and (5) plays with hand tremor (participants reported hand tremor and played with the affected hand): P10 and P13.
We highlighted individual circumstances that might affect STOP’s measurements. P02 reported that his hand rigidity helped him to keep the ball still during game sessions. P09 had poor rotation in his wrists and P11 suffered from rigidity and bradykinesia that make him feel stiff and slow, which might have had a similar effect to that of P02. Finally, P01 was right-handed but used his left hand for playing.
Accelerometer data were recorded as participants played a game for a duration of 10 seconds, henceforth referred to as a “game session.” The accelerometer sampling rate was set to 50 Hz, but the sampling rate varied across different smartphones, as the participants used their own devices for the study. In addition, for some devices, the sampling rate varied within a game session. In the examples in
We identified the closest medication intake record to each game session and labeled the game sessions as “before” or “after” (see
Best case (left) and worst case (right) examples of varying smartphone sampling frequency during a game session.
The timing of medication intake and game sessions. The x-axis shows the time since medication, 0 is the medication intake time and is highlighted with a red vertical line. Each game is associated with the closest medication intake time, either before or after. The y-axis presents the acceleration signal power in logarithmic scale; the sum of power is calculated over the entire spectrum for each game session. Note that the y-axis ranges differ. The first three rows show participants with more than two intakes per day while the last row shows those with only one or two.
PD symptoms can be observed in specific frequency bands: dyskinesia (1-3 Hz), rest tremor (3-6 Hz), postural tremor (6-9 Hz), and kinetic tremor (9-12 Hz). As described in the introduction, tremor can be classified by its activation condition. In our study setup, depending on the user’s posture during a game session, we expected to see differences in the accelerometer signal in rest tremor, postural tremor, and dyskinesia frequencies, which we tried to detect by analyzing this signal in the frequency domain.
We used the Welch method [
From the periodograms, several features were calculated to describe the characteristics of the signal:
area under the curve (AUC): describes the total power of the signal (Hz) [
peak value (PV): represents the maximum value of the PSD;
fundamental frequency (F0): the frequency of maximum power [
central frequency (F50): the central point where the periodogram is divided into two equal parts in PSD [
frequency dispersion (SF50): describes the width of the frequency band around F50 containing 68% of the total power of the signal [
|F50-F0|: the difference between F50 and F0 [
tremor intensity parameter (TIP): calculated as PV divided by SF50. In
We utilized these features to quantify tremor severity during a game session and to detect a difference in medication effects between different game sessions.
Mean of the power spectral densities with the 95% CI for the “no tremor” group (P02, P09, and P11). The left column shows the mean of all game sessions, and the right column shows the mean of the power spectral densities for “before” (red) and “after” (blue) games. Note that the y-axis ranges differ. Frequency areas (dyskinesia, rest tremor, postural tremor, and kinetic tremor) are denoted by different column shades of the background.
Mean of the power spectral densities with the 95% CI for the “tremor” group (P06, P07, P08, and P12). The left column shows the mean of all game sessions, and the right column shows the mean of the power spectral densities for “before” (red) and “after” (blue) games. Note that the y-axis ranges differ. Frequency areas (dyskinesia, rest tremor, postural tremor, and kinetic tremor) are denoted by different column shades of the background.
Mean of the power spectral densities (PSDs) with the 95% CI for the “hand tremor” group (P01 and P05) and the “plays with hand tremor” group (P10 and P13). The left column shows the mean of all game sessions, and the right column shows the mean of the power spectral densities for “before” (red) and “after” (blue) games. Note that the y-axis ranges differ. Frequency areas (dyskinesia, rest tremor, postural tremor, and kinetic tremor) are denoted by different column shades of the background.
The game sessions categorized as dyskinesia, rest tremor, postural tremor, or kinetic tremor according to the fundamental frequency are shown as percentages of all game sessions of each participant (the absolute number of game sessions appears in parentheses).
Participant by group | Dyskinesia | Rest tremor | Postural tremor | Kinetic tremor | |
|
|||||
|
P02 | 52%a (56/107) | 7% (7/107) | 41% (44/107) | 0% (0/107) |
|
P09 | 18% (19/104) | 62%a (64/104) | 20% (21/104) | 0% (0/104) |
|
P11 | 64%a (170/265) | 34% (89/265) | 2% (5/265) | 0% (1/265) |
|
|||||
|
P06 | 30% (53/175) | 70%a (122/175) | 0% (0/175) | 0% (0/175) |
|
P07 | 10% (5/51) | 10% (5/51) | 80%a (41/51) | 0% (0/51) |
|
P08 | 30% (52/174) | 12% (21/174) | 58%a (101/174) | 0% (0/174) |
|
P12 | 34% (38/111) | 40%a (44/111) | 26% (29/111) | 0% (0/111) |
|
|||||
|
P01 | 1% (26/167) | 66%a (111/167) | 18% (30/167) | 0% (0/167) |
|
P05 | 41%a (63/152) | 21% (32/152) | 38% (57/152) | 0% (0/152) |
|
|||||
|
P10 | 4% (7/169) | 95%a (160/169) | 1% (2/169) | 0% (0/169) |
|
P13 | 8% (22/282) | 19% (54/282) | 73%a (206/282) | 0% (0/282) |
aThe most prevalent symptom of the participant.
The power spectral density of one game of P10, playing with hand tremor, and one game of P02, with no tremor. The red, vertical line shows the fundamental frequency (F0), the green line shows the central frequency (F50), and the gap between the lines is the difference between F50 and F0 (|F50-F0|). For P10, F0 and F50 are the same frequency, hence, |F50-F0|=0. The blue rectangle shows the SF50 (the frequency band around F50 containing 68% of the total power of the signal). P10 has a high peak value (PV) and a narrow SF50, leading to a high tremor intensity parameter (TIP) of 24.7. The PV of P02 is small (as is the signal power in the PSD in general) and SF50 is wide; hence, he has a low TIP of 0.26. Note that the y-axis ranges in both plots differ.
In this section, we study our two research questions using the PSD features described in the previous section: (1) how feasible is it to characterize tremor using inertial data captured during our smartphone game?, and (2) can the effects of PD medication be detected using the same inertial data captured during game sessions played before and after medication?
We found a significant correlation between self-reported UPDRS II tremor severity scores (0 to 4) and the TIP (Kendall rank correlation test: z=30.521,
We then compared the groups across all features (see
All features were significantly different between the “no tremor” and “plays with hand tremor” groups. Additionally, all features except for SF50 showed a significant difference between the “no tremor” and “tremor” groups and between the “no tremor” and “hand tremor” groups. SF50 describes the width of the frequency band around F50 containing 68% of the total power of the signal. This suggests that when the tremor was located in a body part other than the hand holding the device, the power of the signal was spread in a wider frequency range, resembling the case with no tremor. However, the “no tremor” group differed significantly from the groups with tremors.
Features between the “tremor” and “hand tremor” groups were significantly different only in the AUC for the dyskinesia, postural tremor, and kinetic tremor frequency ranges. Hence, we can say that the effect of tremors on the accelerometer signal in these groups was mainly similar. In contrast, the comparison of the “plays with hand tremor” group with the “tremor” group and the “hand tremor“ group showed significant differences in all features except F0 and F50. The tremor effect was similar in frequency, but the magnitude of the tremors was different when the tremor hand was used for playing.
Self-reported tremor severity scores using the Unified Parkinson Disease Rating Scale (UPDRS) II, item 16.
|
|
Distribution of tremor intensity parameter | |||||
Participant | UPDRS II item 16 scorea | Minimum | 1st quartile | Median | Mean | 3rd quartile | Maximum |
P02 | 0 | 0.05 | 0.17 | 0.33 | 0.97 | 0.66 | 22.00 |
P09 | 0 | 0.04 | 0.33 | 0.56 | 1.09 | 1.18 | 19.10 |
P11 | 0 | 0.01 | 0.03 | 0.04 | 0.05 | 0.06 | 0.31 |
P05 | 1 | 0.03 | 0.07 | 0.10 | 0.18 | 0.16 | 5.02 |
P07 | 1 | 0.15 | 0.45 | 1.40 | 5.49 | 2.65 | 81.24 |
P08 | 1 | 0.04 | 0.12 | 0.21 | 0.39 | 0.38 | 14.57 |
P12 | 1 | 0.07 | 0.71 | 2.05 | 9.47 | 8.81 | 142.83 |
P01 | 2 | 0.03 | 0.20 | 0.84 | 7.41 | 4.04 | 177.01 |
P10 | 2 | 0.06 | 3.32 | 15.93 | 59.26 | 59.65 | 769.10 |
P13 | 2 | 0.15 | 0.92 | 1.69 | 6.40 | 3.79 | 610.81 |
P06 | 3 | 0.31 | 4.97 | 16.54 | 35.48 | 38.12 | 321.28 |
a0=no tremor, 1=slight and infrequently present tremor, 2=moderate and bothersome tremor, 3=severe tremor interfering with many activities, and 4=marked tremor interfering with most activities.
Comparison of tremor groups using area under the curve for each frequency range: dyskinesia (1-3 Hz), rest tremor (3-6 Hz), postural tremor (6-9 Hz), and kinetic tremor (9-12 Hz).
Comparison of groups in terms of the following features: peak value, fundamental frequency (F0), central frequency (F50), frequency dispersion (SF50), difference between F50 and F0 (|F50-F0|), and tremor intensity parameter.
The
|
|
No tremor vs tremor | No tremor vs hand tremor | No tremor vs plays with hand tremor | Tremor vs hand tremor | Tremor vs plays with hand tremor | Hand tremor vs plays with hand tremor | ||||||
AUC |
|
|
|
|
|
|
|||||||
|
Dyskinesia | <.001 | <.001 | <.001 | <.001 | .03 | <.001 | ||||||
|
Rest tremor | <.001 | <.001 | <.001 | .89 | <.001 | <.001 | ||||||
|
Postural tremor | <.001 | <.001 | <.001 | .02 | <.001 | <.001 | ||||||
|
Kinetic tremor | <.001 | <.001 | <.001 | <.001 | .03 | <.001 | ||||||
Peak value | <.001 | <.001 | <.001 | .49 | <.001 | <.001 | |||||||
F0 | <.001 | <.001 | <.001 | .62 | .23 | .30 | |||||||
F50 | <.001 | <.001 | <.001 | .43 | .15 | .35 | |||||||
SF50 | .71 | .54 | <.001 | .39 | <.001 | <.001 | |||||||
|F50 - F0| | <.001 | <.001 | <.001 | .18 | <.001 | <.001 | |||||||
TIP | <.001 | <.001 | <.001 | .78 | <.001 | <.001 |
We investigated the effect of medication intake on the accelerometer signal characteristics. PD medication is often targeted to alleviate motor symptoms; thus, it could have affected participants’ motor performance during their game sessions. To explore this possibility, we compared the “before” and “after” game sessions of each individual.
In
In the “tremor” group (
In the “hand tremor” and “plays with hand tremor” groups (
In
Change in means of features as percentages between “before” and “after” medication game sessions. The negative values represent a lower mean in “after” game sessions than in “before” game sessions, while positive values represent the opposite.
Participant by group | AUCa, dyskinesia | AUC, rest tremor | AUC, postural tremor | AUC, kinetic tremor | PVb | F0c | F50d | SF50e | |F50-F0|f | TIPg | |
|
|||||||||||
|
P02 | –51h | –51i | –53h | –56h | –53h | 10 | 9 | 11h | 24 | –3h |
|
P09 | 52 | 1 | –23 | 30 | 1 | –13 | –14 | –23h | –28 | 21 |
|
P11 | –29h | –8 | 4 | 4 | –25h | 7 | 10h | 10h | 48 | –35h |
|
|||||||||||
|
P06 | 14 | 21 | 21 | 25 | 22 | 3 | –4 | 5 | 31 | 4 |
|
P07 | 45 | 148 | 8 | 102 | 4 | 10 | 2 | 1 | 2 | –8 |
|
P08 | 41 | 28 | 142 | 45 | 139 | 1 | 0 | 3 | 4 | 108 |
|
P12 | –62 | –40 | –48 | –30 | –59 | –14 | –9 | 0 | –4 | –69 |
|
|||||||||||
|
P01 | –52 | 9 | 51 | 104 | 2 | 7 | 9 | –5 | –24 | –3 |
|
P05 | –26 | –52 | –26 | –7 | –45 | 13 | –3 | 1 | 23 | –51 |
|
|
||||||||||
|
P10 | 1 | –23 | –52 | –42 | –21 | –1 | –2 | 7 | 61 | –3 |
|
P13 | –63 | –53 | –41 | –34 | –51 | –3 | 1 | 1 | –27 | –65 |
aAUC: area under the curve.
bPV: peak value.
cF0: fundamental frequency.
dF50: central frequency.
eSF50: frequency dispersion.
f|F50-F0|: difference between F50 and F0.
gTIP: tremor intensity parameter.
hDifference is statistically significant at
iDifference is statistically significant at
In this paper, we show that it is feasible to detect and characterize PD hand tremor severity using accelerometer data captured during game play. Further, we investigated the medication effect on the accelerometer signal, demonstrating a statistically significant difference in the accelerometer data characteristics of the game sessions played before and after medication intake by participants with rigidity and bradykinesia.
First, how feasible is it to characterize hand tremor using inertial data captured during our smartphone game? To this end, we introduced the TIP for characterizing hand tremor severity, as computed using accelerometer data. We show that TIP is significantly correlated with the tremor score (item 16) on the UPDRS II [
Inspired by previous work in hand tremor analysis using accelerometer data [
Second, can the effects of PD medication be detected using the same inertial data captured during game sessions played before and after medication intake? In other words, we explored whether or not medication-induced changes in motor symptoms could be measured using frequency-domain features extracted from accelerometer data. We classified the games played into two groups: “before” and “after” medication intake. For participants suffering from rigidity and bradykinesia, we found a statistically significant difference in particular signal characterizing features (
As this study was not a laboratory-controlled experiment, the way participants played the game could have affected the accelerometer data signal. For example, if the participant’s hand was extended, such a position might have induced postural tremor, or if the arm was resting on their lap, rest tremor might have become dominant.
It should be noted that the F0 in tremor frequencies does not indicate tremor (
With the STOP app, tremor analysis is limited to tremor severity in participants’ hands, measured using their own smartphone as the instrument. Conversely, in assessments by health professionals, tools such as the UPDRS can be used to evaluate other tremor characteristics, such as amplitude in the legs, jaw, and neck. In our study, the fragmentation of the smartphone device base already caused minor issues, and this can only be expected to exacerbate in the future. To this end, measures to also track and account for the exact device make and model should be added to the approach.
Levodopa treatment is prescribed to alleviate motor symptoms, and although we know the medication intake time, we often ignore the symptoms in particular participants that the medication was prescribed for; hence, the analysis of the medication intake effect is preliminary (it is unclear whether or not the medication was meant to reduce tremor severity). Additionally, the time difference between game sessions and medication intake (“before” and “after” game sessions) varied, as did the magnitude of the changes induced by the medication, which were recorded by the STOP app. This time difference should be taken into account in future studies.
The participant sample size was admittedly small. However, this was compensated for by the high number of individual contributions in the form of game sessions. Further, our data analysis focused on results that generalize sufficiently well for the purposes of this paper: investigating the role of accelerometer data in differentiating between different symptoms and the effects of medication.
Further research is needed to assess the internal and external validity of the TIP, as our results suggest it has the potential to quantify tremor severity. Previous studies [
Changes in our accelerometer features were inconsistent between “before” and “after” medication game sessions. This suggests that we could explore personalized tremor classification models. The effect of hand tremor is visible in the accelerometer signal, but we did not find a statistically significant effect of medication. In addition, it is necessary to focus on particular medication types and PD symptoms to explore the difference between “before” and “after” medication game sessions using accelerometer data in more homogeneous conditions.
Given the availability and sensing capabilities of smartphones, we envision that tools such as the STOP app can support the care and monitoring of PD as well as enable frequent, or even continuous, measuring of medication effects in naturalistic conditions. Even though real-life assessments pose a challenge for data quality due to differences in sensing devices and conditions, standalone smartphone solutions can have a lower burden, thus increasing engagement. For clinicians, a richer picture of symptom severity enabled by sensor data could enable them to better understand people’s conditions and prescribe tailored medications.
In summary, it is feasible to detect and quantify the severity of hand tremor using accelerometer data collected with modern, off-the-shelf smartphones. We replicated and validated previously reported features derived from accelerometer data collected in real-world settings. To this end, we presented the TIP, a metric that could support further research into unobtrusive tremor assessment with smartphones but requires further internal and external validation. Additionally, we identified a statistically significant difference between the game sessions before and after medication intake among participants with rigidity and bradykinesia, and concluded that detecting the effects of PD medication is possible but further research is warranted.
Participants' details.
Technical details of the linear interpolation of the accelerometer signal and the Welch method.
Wilcoxon rank sum test details of group comparisons.
area under the curve
fundamental frequency
central frequency
difference between F50 and F0
Parkinson disease
power spectral density
peak value
frequency dispersion
Sentient Tracking of Parkinson
tremor intensity parameter
Unified Parkinson Disease Rating Scale
This work is partially funded by the Academy of Finland (Grants 313224-STOP, 316253-SENSATE, 320089-SENSATE, 318927-6Genesis Flagship, and 318930-GenZ), and by personal research grants awarded by the Finnish Parkinson Foundation, the Tauno Tönning Foundation, the Jenny and Antti Wihuri Foundation, the Nokia Foundation, and the Emil Aaltonen Foundation.
The author Florian Wolling has been supported by the University of Siegen and the German Academic Exchange Service (DAAD), which enabled his research visit at the University of Oulu, Finland, Biomimetics and Intelligent Systems Group.
None declared.