A Pre-Study on Tremor Classiﬁcation During Activities of Daily Living

Linda B

uker

, Sandra Hellmers

and Andreas Hein

Assistance Systems and Medical Device Technology, Department for Health Services Research,

School VI - School of Medicine and Health Sciences, Carl von Ossietzky Universit

at Oldenburg,

Ammerl

ander Heerstraße 114-118, 26129 Oldenburg, Germany

{linda.bueker, sandra.hellmers, andreas.hein}@uol.de

Keywords:

Tremor Classiﬁcation, Inertial Measurement Units, Support Vector Machine, Activities of Daily Living.

Abstract:

Motor impairments, such as tremors, are often measured with speciﬁc tests or rating scales. As these have some

disadvantages, like an inter-rater reliability and a lack of representation of the everyday life, a sensor-based

continuous and objective monitoring of activities of daily living could be a suitable alternative. According

to the literature, the use of inertial measurement units attached to the tremor-dominant arm in combination

with support vector machines or neural networks seem to be promising. However, many approaches have

to be adapted individually. Therefore, we conducted a preliminary study with ten healthy participants, who

were asked to perform conventional and simulated tremor movements during ﬁve different activities related

to eating. These movements were recorded with inertial measurement units. We identiﬁed four different

parameters calculated from the recorded data, that we used to train multiple support vector machines for a

non-individualized approach. The overall median accuracy score was 0.75, which is comparable to the results

reported in the literature. This shows that support vector machines may be a non-individualized approach for

differentiating between tremor and non-tremor movements during activities of daily living.

1 INTRODUCTION

Motor impairment is an important indicator for the

early detection and monitoring of disease progression

in a number of conditions, including Parkinson’s dis-

ease, stroke, or multiple sclerosis. These motor im-

pairments are measured in a medical context, for ex-

ample, to either detecting them or determining the de-

gree of impairment. There are multiple tests for mea-

suring these impairments. The Action Research Arm

Test (ARAT), for example, is a standardized tool for

measuring the arm motor status in individuals who

have experienced a stroke (Yozbatiran et al., 2008).

The Uniﬁed Parkinson’s Disease Rating Scale (UP-

DRS) is frequently used for assessing the severity of

Parkinson’s disease (Goetz et al., 2008).

Even though these tests are used very frequently

and have many advantages, they come with a few dis-

advantages. These tests are unable to reﬂect possible

ﬂuctuations throughout the day, as they are not a con-

tinuous measurement (Heldman et al., 2011). Further-

more, the test results may be biased due to inter-rater

https://orcid.org/0000-0002-6129-0940

https://orcid.org/0000-0002-1686-6752

https://orcid.org/0000-0001-8846-2282

reliability, with the clinician potentially inﬂuencing

the output (Heldman et al., 2011). In addition, it is

uncertain whether the test accurately reﬂects the limi-

tations encountered in everyday life. On the one hand,

the individuals may perform to a higher standard in a

test situation than they would in their everyday lives.

On the other hand, the exercises included in the test

may not fully cover the individually important every-

day movements of the individuals being tested.

As sensorimotor impairments correlate with activ-

ities of daily living (ADLs) (Shamay et al., 2011), par-

ticularly eating, which is one of the most affected ac-

tivities by a tremor (Heldman et al., 2011; Feys et al.,

2004), measurements during ADLs, especially eating,

could counteract some of the aforementioned disad-

vantages. To enable continuous and objective mon-

itoring, measurements could be conducted via sen-

sors. Soran et al. achieved an accuracy of 95.4% in

the detection of tremors using a camera as a sensor

and a support vector machine (SVM) for training pur-

poses (Soran et al., 2012). In the context of everyday

live, cameras could have two potential disadvantages:

ﬁrstly, that they are stationary, and secondly, that they

may violate the anonymity of individuals. In contrast,

inertial measurement units (IMUs) have the advan-

tage of mobility and greater anonymity. As IMUs

Büker, L., Hellmers, S. and Hein, A.

A Pre-Study on Tremor Classiﬁcation During Activities of Daily Living.

DOI: 10.5220/0013307400003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 805-812

ISBN: 978-989-758-731-3; ISSN: 2184-4305

805

can capture the kinematics during ADLs in healthy

adults (Aguirre, 2016), they could be useful sensors

for monitoring. Additionally, IMUs can be used to es-

timate upper limb impairments after a stroke in ADLs

(Oubre and Lee, 2022). Furthermore, Thorp et al.

suggested that IMUs could be a promising approach

for detecting tremor in in-home settings (Thorp et al.,

2018).

There are multiple studies using IMUs or simi-

lar sensors to monitor ADLs: As demonstrated by

Schmidle et al., the analysis of IMU data during

ADLs can lead to conclusions regarding the frailty

status of the elderly (Schmidle et al., 2020). Gulde

et al. used an accelerometer, gyroscope, and pedome-

ter during daily routines of stroke patients, for ob-

taining information of stroke-related laterality (Gulde

et al., 2024). Heldman et al. used an IMU on the in-

dex ﬁnger during the everyday live to detect and clas-

sify tremors and to quantify tremor severity (Heldman

et al., 2011). Nevertheless, it is unclear whether the

placement of the sensor on the index ﬁnger is an op-

timal choice. As the sensor must not interfere with

the performance of the ADLs (Thorp et al., 2018).

Skaramagkas et al. demonstrated that training a SVM

on features derived from accelerometer data to distin-

guish between essential tremor, Parkinson’s tremor,

and no tremor works best on poses and movements

that are similar to ADL (Skaramagkas et al., 2021). In

their literature review, Thorp et al. also concluded that

Parkinson’s disease symptoms can be classiﬁed using

movement and muscle activity sensors with machine

learning. They found that IMU sensors in combina-

tion with neural networks appear to be particularly

promising for this purpose (Thorp et al., 2018).

Many algorithms need to be customized for each

individual, which is a major challenge for clinical use

(Thorp et al., 2018). Therefore, the aim of this pa-

per is to check whether a non-individualized approach

can be used across multiple individuals. This will

be demonstrated through the analysis of tremor/ no

tremor classiﬁcation in ADLs using an IMU sensor

on the potential affected arm. For this purpose, a pre-

study will be conducted with healthy subjects simu-

lating tremor and moving in a conventional manner

while performing various eating activities, testing two

distinct IMU positions.

2 METHODS

2.1 Study Design

To generate a dataset containing both tremor and con-

ventional movements, we conducted a study with 10

healthy and young subjects (7 male, 3 female). The

subjects were seated in front of a table and performed

ﬁve different exercises. Each exercise was initially

performed in a conventional manner and then with a

simulated tremor in the dominant arm. For all sub-

jects, the dominant arm was the right arm. The ﬁve

exercises were as follows:

1. Hold an apple in the hand and bite into it.

2. Grasp a piece of cake with a fork and then eat it.

3. Cut a slice of bread with a knife.

4. Imitate to spread butter on a slice of bread with a

knife.

5. Consume soup with a spoon.

For the tremor condition, participants were instructed

to shake their dominant arm at a consistent frequency

while performing the exercise. Subsequent analysis

conﬁrmed that the maximum frequency of 75% of

the recorded simulated tremors was between 4 and

12 Hz, which is consistent with the typical frequency

range for postural and kinetic tremors (Heldman et al.,

2011).

The subjects were equipped with two inertial mea-

surement units (Move4, Movisens). As stated in

(Thorp et al., 2018), the sensors should be positioned

on the (tremor-)dominant arm. Accordingly, one sen-

sor was attached to the upper arm and another to the

lower arm enabling a comparison of these two posi-

tions. Each IMU recorded data from an accelerometer

and a gyroscope at a frequency of 64 Hz. In addition,

the subjects were ﬁlmed by a camera (Azure Kinect

DK, Microsoft). An illustration of the study setup is

shown in Figure 1.

Figure 1: Illustration of the study setup. The participant was

seated at a table with the eating utensils positioned in front

of them. The camera was ﬁlming the scene from an angle.

The study was approved by the local ethics com-

mittee (ethical vote: Carl von Ossietzky Universit

Oldenburg (Drs.EK/2024/022), and conducted in ac-

cordance with the Declaration of Helsinki.

2.2 Software Used

All data processing (data preprocessing, parameter

calculation, and analysis) was done on the same com-

HEALTHINF 2025 - 18th International Conference on Health Informatics

806

puter using Python (version 3.12.7). The most impor-

tant used packages were numpy (version 1.26.4), pan-

das (version 2.2.2), scikit-learn (version 1.5.1), and

matplotlib (version 3.9.2).

2.3 Data Preprocessing

Prior to data analysis, it was necessary to preprocess

the data. First, each exercise in the video data was

visually labeled. This was necessary in order to an-

alyze each exercise separately. As the start time was

known from both the videos and the IMU data, the

labeled time spans could be transferred to the IMU

data. However, the IMU data and the video data were

not recorded by the same computer, and the computer

clocks were not synchronized. Consequently, the time

spans must be shifted by seven seconds. To verify the

correct assignment of time spans, each subject’s IMU

data was visually compared to the video data.

The IMU data was then loaded into Python, with

each exercise correctly assigned to the data. Both the

accelerometer and gyroscope sensors provide data for

three axes separately. Therefore, the Euclidean norm

was calculated to obtain the total acceleration and an-

gular velocity. Additionally, the offset observed in

each sensor (along with the 1 g of gravity) was sub-

tracted from the data. These offsets were determined

prior to analysis.

2.4 Parameter Calculation

In order to detect a difference between a conventional

manner and a tremor, a number of parameters de-

rived from the accelerometer and gyroscope data were

calculated. There are multiple parameters found in

the literature, that are used for analyzing activities

of daily living. These include parameters to detect

stroke-related laterality (Gulde et al., 2024), to mea-

sure the frailty status (Schmidle et al., 2020) or to get

an upper limb functional status (Nam et al., 2022).

As it was unclear if these parameters could be used to

distinguish between tremors and conventional move-

ments, we decided to test multiple found parameters.

A list of the calculated parameters is provided in the

appendix (Table 3).

To identify the most important parameters, an

SVM with radial basis function as the kernel was

trained multiple times, with one participant left out

as test set for each SVM on the data of the upper arm.

The most important features were determined for each

run using the permutation importance method, that

determines the contribution of each feature to the per-

formance of the SVM. Subsequently, the four most

prominent features were identiﬁed across all runs.

The four parameters that were used in the following

analysis for both sensor positions are ”Peak Ratio Ac-

celeration”, ”Relative Activity Acceleration”, ”Peak

Ratio Angular Rate”, ”Number Peaks Angular Rate”.

The ”Peak Ratio Acceleration” (see (Schmidle

et al., 2020)) was deﬁned as the ratio between the

number of peaks having a prominence of 0.2

higher and the total number of peaks found in the ac-

celeration data. This parameter was calculated with

the scikit-learn method ﬁnd peaks:

PeakRatioAcc =



{i | prominence(acc

) ≥ 0.2

}



{i | i is a peak in acc}

(1)

The ”Relative Activity Acceleration” (see (Schmi-

dle et al., 2020)) was deﬁned as the number of accel-

eration data points, where the absolute data points are

greater than 0.2

in relation to the total number of

data points:

RelativeActivityAcc =



{z ∈ acc | |z| > 0.2

}



|acc|

(2)

The ”Peak Ratio Angular Rate” (see (Gulde et al.,

2024)) represented the signal-to-noise ratio of the an-

gular velocity signal. It was calculated by the number

of peaks having a minimum prominence of 0.17

◦

di-

vided by the total number of peaks of the angular ve-

locity data. This parameter was also calculated using

the scikit-learn method ﬁnd peaks:

PeakRatioGyro =



{i | prominence(gyro

) ≥ 0.17

◦

}



{i | i is a peak in gyro}

(3)

The ”Number Peaks Angular Rate” was deﬁned

as the ratio of all peaks in angular rate data having a

minimum height of 1.05

◦

and all data points. This

calculation used the scikit-learn method ﬁnd peaks:

NumberPeaksGyro =



{i | height(gyro

) ≥ 1.05

◦

}



|gyro|

(4)

The used thresholds for each parameter were pri-

marily derived from existing literature. After identi-

fying these four parameters, we attempted to optimize

the thresholds using Bayesian optimization. As the

results of the SVMs using the optimized thresholds

are comparable, we employed the aforementioned

thresholds from the literature for the subsequent anal-

ysis.

2.5 Analysis

The aim of this paper is to classify the calculated pa-

rameters into two distinct categories: conventional

A Pre-Study on Tremor Classiﬁcation During Activities of Daily Living

807

movements and tremor movements. Initially, all pa-

rameters were visually analyzed to identify a poten-

tial threshold that might separate data from tremor

and conventional movements. This was conducted for

each calculated parameter and for all used sensors. In

addition, we analyzed the distribution of various pa-

rameters for each individual participant and for each

exercise.

In a second step, we applied machine learning for

the purpose of classiﬁcation. Given the limited size

of the dataset, we have decided not to use neural net-

works. In the existing literature, SVMs were used fre-

quently (Soran et al., 2012; Skaramagkas et al., 2021).

Therefore, we trained multiple SVMs. In addition,

we attempted to train multiple Random Forest Clas-

siﬁers. However, the results were slightly worse than

those obtained from the trained SVMs, and thus we

concentrated our analysis on the SVMs. For training

the multiple SVMs, the identiﬁed four most important

parameters (see Section 2.4) were used. The SVMs

were trained using scikit-learn SVC. For each sensor,

we trained SVMs, with four different kernels: linear,

radial basis function (rbf), polynomial (poly), and sig-

moid. All other SVC parameters were set to their

default values. For each SVM, cross-validation was

employed for training, using a StratiﬁedKFold with

ten splits and shufﬂe, and with accuracy serving as

the scoring metric as the accuracy was also given as

results in the literature (Soran et al., 2012; Skaram-

agkas et al., 2021). For each kernel, multiple SVMs

were trained with distinct train and test sets. The

data of two participants was consistently designated

as the testing set, while the remaining participants’

data served as the training set. This process was re-

peated for every possible combination of participants.

The performance of the SVMs was evaluated based

on the accuracy, recall, precision, and F1 scores ob-

tained from all runs of each used kernel.

All analysis steps were done on both sensors, up-

per and lower arm, to compare these two positions.

3 RESULTS

Due to the malfunction of the camera during the video

recording of one participant, which made it impossi-

ble to recognize the different exercises, the data from

nine participants could be evaluated. In addition, one

participant did not perform the cutting exercise (exer-

cise 3) with a simulated tremor. Therefore, the dataset

contained 45 conventional exercise performances and

44 simulated tremor exercise performances.

As described in Section 2.5, we analyzed all

parameter values visually to determine whether the

tremor and conventional movements could be distin-

guished. Figure 2 shows the four parameters from

Equations 1 to 4 that were used for training the SVM,

calculated for the right upper arm. The parameter val-

ues for all nine participants and all ﬁve exercises (re-

sulting in 45 and 44 values for conventional move-

ment and simulated tremor, respectively) are shown.

No clear threshold could be identiﬁed for the individ-

ual parameters to separate conventional and simulated

tremor data. This behavior was observed consistently

across all other parameters and sensor positions.

Figure 2: The parameter values for all four used parameters,

divided for conventional manner and simulated tremor. The

plotted data represent the values for all participants and all

exercises. The data are presented for the upper arm.

Furthermore, the distribution of two out of the

four used parameters for each exercise and each par-

ticipant, calculated on the upper arm data, is shown

in Figure 3 and Figure 4, respectively. Figure 3

shows that some conventional performances (repre-

sented by a cross) of exercises eating an apple (1)

and cutting bread (3) were distributed more in the

cluster representing the tremor performances (repre-

sented by a circle). Figure 4a shows that some perfor-

mances of participant 0 with a simulated tremor were

closely aligned with the cluster representing conven-

tional movements (depicted by blue circles on the

left). In addition, the conventional movements of par-

ticipant 5 (illustrated by brown crosses in Figure 4b)

tended towards higher parameter values than the other

participants.

Furthermore, multiple SVMs were trained for four

different kernels and different train and test sets. Ta-

ble 1 shows the minimum, maximum, median, and

standard deviation of accuracy, recall, precision, and

F1 scores for all four kernels, calculated on all splits

HEALTHINF 2025 - 18th International Conference on Health Informatics

808

Figure 3: Scatter plot of two out of the four used param-

eters. The color represents the ﬁve different exercises eat-

ing an apple (1), eating with a fork (2), cutting a bread (3),

spreading butter (4), and eating soup (5). A cross represents

a performance in a conventional manner while a circle rep-

resents a performance with a simulated tremor. Red circles

highlight performances of exercises 1 and 3 with conven-

tional movements. The data are presented for the upper arm.

of the train and test sets. These values are for the sen-

sor attached to the upper arm. The median accuracy

was 0.75 for all kernels, while the median F1 score

ranged from 0.72 (sigmoid) to 0.79 (poly). In com-

parison, the metrics for the lower arm are shown in

Table 2. Here, the median accuracy ranged from 0.60

(poly) to 0.65 (linear and rbf), while the median F1

score ranged from 0.63 (rbf and poly) to 0.67 (linear

and sigmoid).

4 DISCUSSION

Upon visual analysis of the parameter values, we

could not identify thresholds that would differenti-

ate the values between conventional movements and

movements with a simulated tremor. Therefore, there

does not seem to be a clear distinction between the

two behaviors across all participants. The distribution

of the individual participants across the parameters in

Figure 4 also demonstrated that there are differences

between individuals. These observations align with

those of Thorp et al., indicating that most algorithms

require individual thresholds for individual persons

(Thorp et al., 2018).

The trained SVMs showed that a distinction be-

tween conventional movements and tremor yielded

to better outcome when a sensor was positioned at

the upper arm in comparison to the lower arm. This

(a) Red circles highlight performances of participant 0 with

a simulated tremor.

(b) Red circles highlight performances of participant 5 with

conventional movements.

Figure 4: Two scatter plots, each showing two out of the

four used parameters. The color represents the nine differ-

ent participants. A cross represents a performance in a con-

ventional manner while a circle represents a performance

with a simulated tremor. The data are presented for the up-

per arm.

is particularly interesting, as the videos indicate that

the participants tended to present more noticeable

tremors in their lower arms than in their upper arms.

Given the considerable discrepancy in scores be-

tween the different splits of train and test sets, we

have conducted a more detailed analysis of the test

sets with accuracy scores below or equal to 0.6. It is

striking that participant 5 had primarily recall scores

A Pre-Study on Tremor Classiﬁcation During Activities of Daily Living

809

Table 1: The minimum (Min), maximum (Max), median,

and standard deviation (Std) of the accuracy, recall, preci-

sion, and F1 scores for all test sets of the trained support

vector machines separated by the four different kernel types

linear, rbf, poly, and sigmoid. The data are presented for the

upper arm.

SVM kernel

linear rbf poly sigmoid

Accuracy

Min 0.58 0.58 0.63 0.55

Max 0.95 0.95 0.90 0.90

Median 0.75 0.75 0.75 0.75

Std 0.09 0.10 0.06 0.08

Recall

Min 0.40 0.40 0.70 0.30

Max 1.00 1.00 1.00 1.00

Median 0.80 0.80 1.00 0.70

Std 0.18 0.19 0.08 0.19

Precision

Min 0.54 0.55 0.56 0.57

Max 1.00 1.00 0.9 1.00

Median 0.75 0.78 0.67 0.75

Std 0.11 0.12 0.07 0.11

Min 0.50 0.50 0.72 0.40

Max 0.95 0.95 0.90 0.90

Median 0.76 0.75 0.79 0.72

Std 0.11 0.12 0.04 0.11

of 0.75 or 1.0 and precision scores of 0.5, indicat-

ing that numerous conventional movements were de-

tected as simulated tremors. Upon inspection of the

video, it became evident that this participant seemed

to experience involuntary tremors on occasional ba-

sis. This is observable between the exercises during

periods of rest. It is possible that this participant also

experienced tremors while performing the exercises

in a conventional manner, which could have resulted

in the detection as a simulated tremor. In contrast,

participant 8 often showed recall and precision scores

of 0.6, with three correct and two incorrect exercises

of each type of movement. This behavior cannot be

explained at ﬁrst glance, as the videos indicate clear

differences between the participants conventional and

tremor movements. Participant 0 is also noticeable,

with a recall score of only 0.2 and a precision score of

1.0 for the majority of the exercises, as all exercises

were usually classiﬁed as “conventional”. In com-

parison to other participants, this participant seems

to simulate a slightly lighter tremor, which could be

the source of these results. These scores align with

the results presented in Figure 4. This shows that, on

the one hand, a slight natural tremor movement can

be detected as tremor while on the other hand, slight

simulated tremor movements can be classiﬁed as con-

ventional movements. This needs to get investigated

further in the future.

The overall results of the trained SVMs on the

Table 2: The minimum (Min), maximum (Max), median,

and standard deviation (Std) of the accuracy, recall, preci-

sion, and F1 scores for all test sets of the trained support

vector machines separated by the four different kernel types

linear, rbf, poly, and sigmoid. The data are presented for the

lower arm.

SVM kernel

linear rbf poly sigmoid

Accuracy

Min 0.37 0.42 0.47 0.45

Max 0.90 0.85 0.90 0.85

Median 0.65 0.65 0.60 0.62

Std 0.11 0.11 0.10 0.11

Recall

Min 0.20 0.10 0.00 0.20

Max 1.00 1.00 1.00 1.00

Median 0.70 0.60 0.63 0.78

Std 0.24 0.22 0.27 0.19

Precision

Min 0.40 0.43 0.00 0.40

Max 1.00 1.00 1.00 0.89

Median 0.64 0.68 0.63 0.58

Std 0.13 0.14 0.18 0.12

Min 0.29 0.17 0.00 0.27

Max 0.91 0.86 0.90 0.86

Median 0.67 0.63 0.63 0.67

Std 0.15 0.16 0.18 0.13

upper arm with a median accuracy of 0.75 for ev-

ery kernel indicate that it may be feasible to detect

a tremor during ADLs with a non-individualized ap-

proach. Skaramagkas et al. show success rates be-

tween 56.9% and 96.5%, with an average of 75.66%

for data recorded at the forearm (Skaramagkas et al.,

2021). These ﬁndings are comparable to our re-

sults. It should be noted that a distinction was made

between three classes (essential tremor, Parkinson’s

tremor, and no tremor) and not between two classes,

as is the case in our study. It is also important to note

that a separate classiﬁer was trained for each exercise,

rather than a single classiﬁer for different movements

combined. As shown in Figure 3, the distributions

of parameters may vary in different activities, poten-

tially inﬂuencing the results. Our results were not as

good as those reported by Soran et al. (accuracy of

95.4%) (Soran et al., 2012), although it is important

to note that they used a camera, which is a differ-

ent sensor system than our used IMUs. The study’s

ﬁndings indicate the presence of differences among

the participants, primarily due to the absence of a

clear threshold for distinguishing between the move-

ments and the presence of variability in the precision

and recall scores of the individual participants. Nev-

ertheless, a non-individualized approach was imple-

mented, enabling the differentiation of tremor from

conventional movements with a median accuracy of

0.75.

HEALTHINF 2025 - 18th International Conference on Health Informatics

810

It should be noted that the data set is relatively

small, with only nine participants and ﬁve exercises.

Additionally, the participants did not experience an

actual tremor, but rather simulated one. Nevertheless,

the overall results appear promising, suggesting that

the study should be repeated with a larger sample size

in the future. This should include participants with

and without actual tremor performing different exer-

cises to verify that machine learning algorithms like

SVM can effectively differentiate between individu-

als with and without tremor in ADLs. With a larger

data set, it is also possible to test whether neural net-

works, such as LSTM, deliver better results than an

SVM. Additionally, a more detailed evaluation of the

other calculated parameters could be conducted in the

future, as it is possible that the permutation feature

method may discard relevant parameters if they ap-

pear to correlate with other parameters.

5 CONCLUSION

In conclusion, we trained a SVM as a non-

individualized approach to distinguish between a

tremor and conventional movements during ADLs

with a median accuracy of 0.75. Therefore, in addi-

tion to the tests and rating scores used to quantify im-

pairments, data could be recorded in everyday life to

identify possible ﬂuctuations throughout the day, gen-

erate more objective measurements, and enhance the

recognition of actual effects on everyday life. How-

ever, this requires further conﬁrmation through a more

detailed study.

ACKNOWLEDGEMENTS

We thank Matthis Heese for conducting the study.

This work was supported by the Research Training

Group (RTG) 2783, funded by the German Research

Foundation (DFG) - Project ID 456732630.

REFERENCES

Aguirre, A. (2016). Evaluating upper-extremity

(dys)function using measurement unit technol-

ogy and its applications to resource-constrained

settings. In 2016 IEEE Global Humanitarian

Technology Conference (GHTC), pages 640–645.

Feys, P., Romberg, A., Ruutiainen, J., and Ketelaer, P.

(2004). Interference of upper limb tremor on daily life

activities in people with multiple sclerosis. Occupa-

tional Therapy In Health Care, 17(3-4):81–95. PMID:

23941223.

Goetz, C. G., Tilley, B. C., Shaftman, S. R., Stebbins, G. T.,

Fahn, S., Martinez-Martin, P., Poewe, W., Sampaio,

C., Stern, M. B., Dodel, R., et al. (2008). Movement

disorder society-sponsored revision of the uniﬁed

parkinson’s disease rating scale (mds-updrs): scale

presentation and clinimetric testing results. Movement

disorders: ofﬁcial journal of the Movement Disorder

Society, 23(15):2129–2170.

Gulde, P., Vojta, H., Schmidle, S., Rieckmann, P., and

Hermsd

orfer, J. (2024). Outside the laboratory assess-

ment of upper limb laterality in patients with stroke: A

cross-sectional study. Stroke, 55(1):146–155.

Heldman, D. A., Jankovic, J., Vaillancourt, D. E., Prodoehl,

J., Elble, R. J., and Giuffrida, J. P. (2011). Essential

tremor quantiﬁcation during activities of daily living.

Parkinsonism & Related Disorders, 17(7):537–542.

Nam, H. S., Lee, W. H., Seo, H. G., Smuck, M. W., and

Kim, S. (2022). Evaluation of motion segment size

as a new sensor-based functional outcome measure in

stroke rehabilitation. Journal of International Med-

ical Research, 50(9):03000605221122750. PMID:

36129970.

Oubre, B. and Lee, S. I. (2022). Estimating post-

stroke upper-limb impairment from four activities of

daily living using a single wrist-worn inertial sen-

sor. In 2022 IEEE-EMBS International Conference

on Biomedical and Health Informatics (BHI), pages

01–04.

Schmidle, S., Gulde, P., Jansen, B., Herdegen, S., and

Hermsd

orfer, J. (2020). Frailty assessment in daily

living (frail) - assessment of adl performance of frail

elderly with imus. In Stephanidis, C., Antona, M.,

and Ntoa, S., editors, HCI International 2020 – Late

Breaking Posters, pages 92–101, Cham. Springer In-

ternational Publishing.

Shamay, N. S., William, T. W., Patrick, K. W., Philip, T. T.,

and Jefferry, W. C. (2011). Sensorimotor impairments

of paretic upper limb correlates with activities of daily

living in subjects with chronic stroke. South African

journal of physiotherapy, 67:9–16.

Skaramagkas, V., Andrikopoulos, G., Kefalopoulou, Z., and

Polychronopoulos, P. (2021). A study on the essen-

tial and parkinson’s arm tremor classiﬁcation. Signals,

2(2):201–224.

Song, X., Van De Ven, S. S., Liu, L., Wouda, F. J., Wang,

H., and Shull, P. B. (2022). Activities of daily living-

based rehabilitation system for arm and hand motor

function retraining after stroke. IEEE Transactions

on Neural Systems and Rehabilitation Engineering,

30:621–631.

Soran, B., Hwang, J.-N., Lee, S.-I., and Shapiro, L. (2012).

Tremor detection using motion ﬁltering and svm. In

Proceedings of the 21st International Conference on

Pattern Recognition (ICPR2012), pages 178–181.

Thorp, J. E., Adamczyk, P. G., Ploeg, H.-L., and Pickett,

K. A. (2018). Monitoring motor symptoms during ac-

tivities of daily living in individuals with parkinson’s

disease. Frontiers in Neurology, 9.

Yozbatiran, N., Der-Yeghiaian, L., and Cramer, S. C.

(2008). A standardized approach to performing the ac-

tion research arm test. Neurorehabilitation and Neural

Repair, 22(1):78–90. PMID: 17704352.

A Pre-Study on Tremor Classiﬁcation During Activities of Daily Living

811

APPENDIX

Table 3: All calculated parameters.

Parameter Name Explanation of parameter

Mean Acceleration Mean of acceleration; found in (Aguirre, 2016)

Maximal Jerk Maximum of jerk; found in (Aguirre, 2016)

Mean Jerk Mean of jerk

Mean Absolute Jerk Mean of absolute jerk values

Mean Amplitude Deviation Ac-

celeration (MAD)

The mean of the distances to the mean acceleration; based on (Gulde

et al., 2024)

90th Percentile of Mean Ampli-

tude Deviation

the 90th percentile of the distances to the mean acceleration; based on

(Gulde et al., 2024)

Physical Activity Level The ratio of the time with an absolute distance to the mean acceleration

greater 0.1 g and the total time; based on (Gulde et al., 2024)

Maximum Angular Velocity The maximum of angular velocity; based on (Gulde et al., 2024)

Mean Angular Velocity The mean of angular velocity; based on (Gulde et al., 2024)

Standard Deviation Angular Ve-

locity

The standard deviation of angular velocity; based on (Gulde et al., 2024)

Peak Ratio Angular Rate The ratio of the number of peaks in angular velocity with a prominence

of 0.17

◦

and the number of all peaks in angular velocity; based on

(Gulde et al., 2024)

Number Angular Velocity Peaks

per 360°

The number of peaks in angular velocity per 360°; based on (Gulde et al.,

2024)

Mean Height Angular Velocity

Peaks

The mean of the heights of all peaks in angular velocity; based on (Gulde

et al., 2024)

Mean Standard Deviation Angu-

lar Velocity Peaks

The standard deviation of the heights of all peaks in angular velocity;

based on (Gulde et al., 2024)

Standard Deviation Acceleration Standard deviation of the acceleration data; found in (Song et al., 2022)

Relative Activity Acceleration ”Period of time in which the absolute acceleration signal exceeded 0.2

related to [duration]” (Schmidle et al., 2020)

Peak Standard Deviation Acceler-

ation

”Standard deviation of all acceleration peaks (maxima) in

” (Schmidle

et al., 2020)

Peaks Per Second Acceleration ”Number of acceleration peaks per second” (Schmidle et al., 2020)

Peak Ratio Acceleration ”Ratio between the number of acceleration peaks with a minimum

prominence of 0.2

and the total number of acceleration peaks”

(Schmidle et al., 2020)

Mean Peak Acceleration ”Mean of acceleration peaks” (Schmidle et al., 2020)

Signal to Noise Ratio Accelera-

tion

”Ratio of the sum of the frequency spectrum [of the acceleration data]

by a fast Fourier transformation from 0.01 to 3 Hz and from 0.01 to 50

Hz” (Schmidle et al., 2020)

Frequency Spectrum 0-3 Hz Sum of the frequency spectrum of the acceleration data from 0.01 to 3

Frequency Spectrum 3-50 Hz Sum of the frequency spectrum of the acceleration data from 3 to 50 Hz

Mean Variance Acceleration Mean of the variance of every second of the acceleration data

Root Mean Square Acceleration Root mean square of acceleration data

Root Mean Square Jerk Root mean square of jerk data

Number Peaks Acceleration Ratio of all peaks in acceleration data having a minimum hight of 1.05 g

and all data points

Mean Variance Angular Rate Mean of the variance of every second of the angular rate

Amplitude Angular Rate Difference of maximum and minimum of angular rate

Number Peaks Angular Rate Ratio of all peaks in angular rate data having a minimum hight of 1.05

◦

and all data points

HEALTHINF 2025 - 18th International Conference on Health Informatics

812