Accelerometer-Based Multi-Domain Feature

Extraction for Parkinsonian Activity Recognition

1st Hong Thien An Nguyen 2nd VAN ANH DAO 3rd Quoc Huy Pham
2025 International Conference on Activity and Behavior Computing (ABC) | 979-8-3315-3437-0/25/$31.00 ©2025 IEEE | DOI: 10.1109/ABC64332.2025.11118560

dept. of Applied Mathematics dept. of Applied Mathematics dept. of Computer Science

Ho Chi Minh City Ho Chi Minh City Ho Chi Minh City
University of Technology- University of Technology- University of Technology-
Vietnam National University Vietnam National University Vietnam National University
Ho Chi Minh City, Vietnam Ho Chi Minh City, Vietnam Ho Chi Minh City, Vietnam
[Link]@[Link] anh.dao1606@[Link] [Link]@[Link]

4th Dang Khoa Nguyen 5th Quynh Nhu Pham 6th Nhat Tan Le
dept. of Applied Mathematics dept. of Applied Mathematics dept. of Biomedical Engineering
Ho Chi Minh City Ho Chi Minh City Ho Chi Minh City
University of Technology- University of Technology- University of Technology-
Vietnam National University Vietnam National University Vietnam National University
Ho Chi Minh City, Vietnam Ho Chi Minh City, Vietnam Ho Chi Minh City, Vietnam
khoa.nguyendangk24@[Link] nhu.pham172@[Link] lenhattan@[Link]

Abstract—This paper details our methodology for the ABC cantly impact quality of life. With rapid population aging, PD
Conference 2025 challenge on human activity recognition using prevalence is expected to rise, increasing the risk of severe
wearable accelerometer data. Research also explores the use of disability and mortality [4]. Diagnosis is often delayed due to
broader wearable data for monitoring and predicting Parkinson’s
disease-related phenomena such as wearing-off [1]. The ap- variable clinical expression and symptom overlap with other
proach combines comprehensive feature extraction from diverse diseases [5]. Moreover, motor fluctuations, such as “wearing
domains (statistical, energy, frequency, nonlinear and tempo- off” [6], complicate treatment. Continuous, accurate monitor-
ral) with Recursive Feature Elimination using Cross-Validation ing via wearable devices (accelerometers, gyroscopes) now
(RFECV) to select a highly relevant feature subset. The selected enables early, objective diagnosis and symptom quantification
features are then used to train multiple classifiers, including
XGBoost, Random Forest, Multi-Layer Perceptron (MLP), and [5] [7]. These devices can differentiate PD patients from
Self-Training Random Forest. Among these, XGBoost achieved healthy controls and monitor medication effects, informing
the best performance and was ultimately used to classify ten tailored treatment strategies. Given the increasing prevalence
different activities, including Parkinson’s-related movements. The and impact of PD, there is a growing need for systems aiding
evaluation results indicate that the XGBoost model achieves in its early and automatic diagnosis [2].
a high overall accuracy and a weighted F1-score of 40% on
the challenge dataset. The contribution of key features, such as The paper is organized as follows: Section II reviews related
spectral centroid and Petrosian fractal dimension, to classification work, Section III outlines the proposed methodology, presents
performance is discussed. Limitations of the current methodol- the classification models for PD-related activities. Section IV
ogy and potential future work, including the incorporation of evaluates model performance. Section V discusses results and
deep learning techniques and multi-modal data integration to limitations, and Section VI concludes with key findings.
enhance robustness and generalizability for Parkinson’s activity
monitoring, are also addressed. II. R ELATED WORKS
Index Terms—Human Activity Recognition, Accelerometer
Data, Wearable Sensors, Machine Learning, Feature Extraction, Several studies have demonstrated the effectiveness of
Time-Series Analysis, Parkinson’s Disease, XGBoost, RFECV, accelerometer data for recognizing Parkinson-related motor
Classification Models symptoms. For instance, Sun et al. (2023) [8] used ac-
celerometer data to classify motor states with high accuracy
I. I NTRODUCTION in detecting PD-related tremors, and similar research has
Parkinson’s disease (PD), the second most common neu- explored accelerometers for identifying various Parkinson’s
rodegenerative disorder [2], is increasing worldwide. It af- symptoms [1]. In our work, Support Vector Machines (SVMs)
fects approximately 0.3% of the general population, rising combined with manually crafted features—selected via the
to 2–3% in individuals over 65 and over 3% in those older LASSO algorithm—achieve over 92% accuracy in five-fold
than 80 [3] [3]. PD is characterized by tremors, rigidity, cross-validation and over 90% in leave-one-out evaluation.
bradykinesia/akinesia, and postural instability, which signifi- Similarly, Rasel Ahmed Bhuiyan et al. (2020) [9] propose

Authorized licensed use limited to: American International University Bangladesh. Downloaded on July 30,2026 at 07:43:56 UTC from IEEE Xplore. Restrictions apply.
a feature extraction model using Enveloped Power Spectrum
(EPS) for noise-resistant frequency analysis, Linear Discrimi-
nant Analysis (LDA) for dimensionality reduction, and Multi-
class SVM for classification, achieving 98.67% and 100%
accuracy on the UCI-HAR and DU-MD datasets, respectively.
These studies underscore the reliability of accelerometer data
in capturing motor fluctuations, motivating its use in this
research.
Despite this potential, challenges remain in effectively ex-
tracting and selecting meaningful features from time-series
data. Temporal dependencies, high dimensionality, and over-
lapping motor fluctuations complicate the process, necessitat-
ing advanced signal processing techniques and domain exper- Fig. 1. Proposed pipeline for accelerometer data processing and machine
tise. As not all extracted features contribute equally to per- learning-based classification
formance, selecting the most informative and non-redundant
features is crucial for building efficient models. Moreover,
rows that fall within the defined intervals, ensuring that each
developing a multimodal classification system that generalizes
segment accurately represents a true activity event.
well across diverse populations is complex.
Once the relevant data has been extracted, we employ a
Within the framework of the 7th ABC Challenge (ABC
sliding window segmentation method with a fixed 3-second
Conference [10]), this study introduces a classification pipeline
window and 50% overlap to further partition the continu-
specifically designed for recognizing normal and tremor activ-
ous time-series data into structured units. This segmenta-
ities from sensor data. The objectives are:
tion strategy ensures that each window maintains temporal
• Preprocessing accelerometer data from real-world set- continuity while capturing sufficient movement patterns for
tings, addressing noise, data imbalance, and temporal activity recognition. The 50% overlap is chosen to balance
dependencies; the trade-off between preserving transitions between activities
• Feature Extraction from multiple domains to enhance and maintaining computational efficiency. Each segment is
generalization; then assigned an activity label based on the mode of activ-
• Feature Selection Using RFECV: applying RFECV for ity occurrences within the window, ensuring consistency in
optimal feature selection; classification even when transitions occur between different
• Analysis of Feature Importance: Analyzing feature im- activities.
portance to identify characteristics most indicative of Compared to standard fixed-window segmentation applied
Parkinson’s disease. uniformly across the entire dataset, this timestamp-aligned
segmentation method minimizes the inclusion of irrelevant
III. M ETHODOLOGY data and reduces noise by excluding time periods that do
The pipeline begins by querying timestamps from the not correspond to designated activities [11]. By aligning
training file in the dataset and segmenting the corresponding segments with precise, activity-specific timestamps and further
accelerometer signals. Feature extraction and selection are then refining them through a structured sliding window approach,
performed to isolate the most informative characteristics, fol- our method enhances the accuracy of activity labeling and
lowed by standardization to ensure consistent data scaling. The improves the reliability of downstream feature extraction and
resulting feature set is split into training (70%) and validation classification analyses.
(30%) subsets, and machine learning models are trained and Figure 2 shows how many segmented files were generated
evaluated on these subsets. In the testing phase, the pipeline for each of the ten activity labels. Label 10 has the highest
uses timestamps from a separate test file to segment and number of segments (404), followed by Label 9 (353) and
preprocess the accelerometer signals using the same feature Label 1 (314). In contrast, Label 7 has the fewest segments
extraction and standardization procedures. Finally, the trained (80). Overall, the distribution is uneven, indicating that some
model generates the predicted labels, which form the basis of activities yielded more segmented files than others.
the final challenge submission.
B. Feature Extraction and Selection
A. Querying Timestamps and Segmentation In this stage, a broad set of features is computed from
In this pipeline, we begin by querying the timestamps from the accelerometer signals, covering statistical, frequency-
the training file to isolate intervals of sensor data correspond- domain, energy and power domain, temporal domain and
ing to specific activities precisely. For each record, the start nonlinear domain. Each category targets a different aspect
and end timestamps, along with the associated activity type of movement, such as distribution (statistical), spectral pat-
identifier, are extracted and used to create a unique filename. terns (frequency-domain), or overall intensity (energy and
The raw accelerometer data is then filtered to retain only those power). After extraction, a selection step keeps only the

Authorized licensed use limited to: American International University Bangladesh. Downloaded on July 30,2026 at 07:43:56 UTC from IEEE Xplore. Restrictions apply.
 where, xi , yi and zi represent acceleration values along the
respective axes at the ith time point, while µx , µy and µz
denote their mean values. The number of data points is
given by N. Correlation values range from −1 to 1, when
a correlation of 1 indicates a perfect positive correlation (both
variables increase together), −1 represents a perfect negative
correlation (one increases while the other decreases), and 0
signifies no correlation (there is no relationship between their
changes). High correlation between acceleration axes suggests
that the motion is occurring in a consistent direction or pattern,
such as walking or standing up, whereas low correlation
indicates independent or uncoordinated movements, like hand
shaking or sudden stops.
Collectively, statistical features such as Mean, Minimum,
Fig. 2. Statistical analysis of the number of Segmented files across 10 Maximum, Variance and Correlation play a crucial role in
activities
the analysis and recognition of activities from accelerometer
data. Each feature provides valuable insights into the level of
most informative features—those that best capture activity- fluctuation, stability, or abrupt changes in acceleration over
specific variations—resulting in a refined set that improves time. When effectively combined and utilized, these features
model performance while reducing complexity, as shown by can aid in constructing accurate activity recognition models,
the difference between “Extracted Features” and “Selected helping differentiate between various types of activities.
Features” in the table I. 2) Frequency domain features: Motion data from wearable
1) Statistical features: Statistical features in time-series accelerometers inherently exists in the time domain but can
data analysis are critical for activity recognition tasks, particu- be transformed into the frequency domain through processing
larly when utilizing accelerometer sensor data. These features techniques such as Fast Fourier Transform (FFT) and wavelet
help summarize the overall trends and describe the changes in transforms [14]. These transformations allow the extraction
the signal over time, thereby supporting activity classification. of essential characteristics that enhance the classification of
In this study, we extract and analyze several statistical features human activities, supporting the challenge’s objective of accu-
to enhance activity recognition performance. The mean is used rately distinguishing ten different movement classes. Spectral
to determine the overall acceleration level over a period of centroid, a key frequency-based feature, quantifies the distribu-
time, helping to identify the general trend of an activity and tion of motion signal energy across frequencies and is partic-
distinguishing between static and dynamic activities. Addition- ularly effective in identifying activities involving rapid, high-
ally, the minimum and maximum values capture the range of frequency movements [15]. Higher spectral centroid values are
acceleration fluctuations, providing insights into the amplitude commonly associated with activities such as hand tremors or
of motion variations. Furthermore, variance helps evaluate the full-body shaking, which are especially relevant for detecting
variability of acceleration. High variance in a short amount of Parkinsonian symptoms. Additionally, wavelet-based analysis
time indicates there could that be rapid activities (like shaking) provides a powerful alternative by preserving both frequency
while in a long time there could be walking. Conversely, low and temporal information, enabling improved recognition of
variance represents stable activities such as cool down, sitting transient and periodic motion patterns [16]. Wavelet energy
or relaxing [12]. measures the distribution of signal intensity across frequency
Beyond basic statistical measures, correlation is crucial in bands, making it effective for differentiating dynamic activities
assessing the relationship between acceleration values across like walking from static postures such as sitting or standing
different axes (X, Y, Z) [13]. It quantifies the degree to [17]. Meanwhile, wavelet entropy quantifies the complexity
which movements along these axes are related, as defined by of motion signals, offering a useful metric for identifying
Equations 1, 2, and 3: unpredictable movements that may indicate irregular motor
PN functions. By leveraging these frequency-domain features, the
(xi − µx )(yi − µy ) model can more accurately classify a range of activities,
rxy = qP i=1 (1)
N 2
PN 2
ensuring robust performance in real-world scenarios.
(x
i=1 i − µ x ) (y
i=1 i − µy )
3) Energy and power domain: The Energy-Power Domain
PN focuses on measuring the intensity and variability of signals,
i=1 (xi − µx )(zi − µz ) providing insights into the subject’s level of activity. These
rxz = qP (2)
N PN
− µx )2 i=1 (zi − µz )2 features are particularly valuable in biomedical monitoring
i=1 (xi
applications, such as detecting movement disorders in PD. By
PN capturing variations in movement intensity, they enable the
− µy )(zi − µz )
i=1 (yi
ryz = qP (3) identification of abnormal motor activities, including tremor,
N 2
PN 2
i=1 (yi − µy ) i=1 (zi − µz )
freezing, and impaired movement. Previous studies have high-

Authorized licensed use limited to: American International University Bangladesh. Downloaded on July 30,2026 at 07:43:56 UTC from IEEE Xplore. Restrictions apply.
 TABLE I
E XTRACTED AND S ELECTED F EATURES TABLE

Domain Quantity (Extracted) Extracted Features Quantity (Selected) Selected Features

Statistical 36 Mean, median, standard deviation, 27 Mean, median, minimum, maximum,
minimum, maximum, variance, mean distance, kurtosis, mean absolute
absolute deviation, median absolute deviation, median absolute deviation (X,
deviation, kurtosis, skewness, distance, Y, Z); Standard deviation (X, Y); Variance
ECDF slope (X, Y, Z) (Y)
Frequency- 33 Spectral centroid, spectral spread, 15 Wavelet energy, spectral entropy, spectral
domain spectral flatness, spectral entropy, centroid, spectral spread, spectral flatness
median frequency, dominant frequency, (X, Y, Z)
fundamental frequency, FFT magnitude,
wavelet energy, zero-crossing rate, peak
frequency (X, Y, Z)
Energy 16 RMS, absolute energy, average power, 14 Total energy, SMA; Absolute energy,
and Power relative log energy (X, Y, Z) average power, RMS, relative log energy
(X, Y, Z)
Nonlinear 24 Higuchi fractal dimension, detrended 14 Entropy (Y); Petrosian fractal dimension,
fluctuation analysis, permutation entropy, lyapunov exponent, teager-kaiser energy,
entropy, petrosian fractal dimension, permutation entropy (X, Y, Z); Hurst
lyapunov exponent, teager-kaiser energy exponent
(X, Y, Z)
Temporal 24 Signal slope changes, jerk feature, 12 AUC, jerk feature (X, Y, Z); Vector
total distance traveled, autoregressive magnitude, magnitude; Correlation
coefficients, AUC, hurst exponent (X, between axes (XY, XZ, and YZ); Signal
Y, Z); Vector magnitude, magnitude; slope changes (X)
Correlation between axes (XY, XZ, and
YZ)

lighted the significance of the domain. For instance, Luft of low-powered devices used in practical deployments, PFD
et al. (2019) [18] employed Power Spectral Density (PSD)- presents a viable option for efficient and effective activity
based methods to detect tremor and tremor intermittency recognition.
in movement disorders, demonstrating the relevance of the By integrating these nonlinear dynamical features with
Energy-Power Domain in distinguishing Parkinsonian move- traditional frequency-based methods, the model enhances its
ment states. capability to distinguish between the ten activity classes in the
4) Nonlinear domain: Nonlinear dynamical features pro- challenge, ensuring accurate and reliable recognition of human
vide additional insights into movement complexity, aiding movements.
in the classification of different activities. Fractal-based ap- 5) Temporal domain: Temporal features play a crucial role
proaches, such as box-counting dimension, measure the self- in analyzing motion signals by capturing changes over time,
similarity of signals across scales [19]. This is particularly use- making them indispensable for human activity recognition.
ful in activity recognition, as activities with varying levels of Since accelerometer signals exhibit strong temporal dependen-
motion complexity—such as walking versus sitting—produce cies, extracting these features provides valuable insights into
accelerometer signals with different fractal dimensions. By movement dynamics and behavioral patterns.
capturing these variations, box-counting dimension effectively One of the fundamental temporal features is the Area Under
differentiates between distinct movement patterns. the Curve (AUC), which quantifies the cumulative magnitude
Another valuable nonlinear measure is the Lyapunov ex- of a signal over time [23]. This measure reflects the total
ponent, which quantifies the divergence or convergence of intensity of movement, making it useful for distinguishing
nearby trajectories in a dynamical system [20]. This metric between high-energy activities, such as walking, and more
is particularly relevant for human motion analysis, as it helps stationary behaviors, like sitting.
assess gait stability and detect irregular movement patterns. Another key temporal feature is Run-Length Encoding
A high Lyapunov exponent suggests chaotic movement, often (RLE), which captures the persistence of specific states by
seen in dynamic activities, while lower values indicate more measuring the duration and frequency of repeated values in a
stable and predictable motion [21]. Such characteristics are time series [24]. This feature helps identify prolonged motion
crucial for identifying neurological disorders that manifest or inactivity, which is particularly relevant for recognizing
through gait instability. structured activities such as continuous walking or standing
Furthermore, the Petrosian Fractal Dimension (PFD) pro- still.
vides a computationally efficient measure of signal complexity Additionally, the jerk feature, which quantifies the rate
by evaluating the number of sign changes in a time series of change of acceleration, provides insights into movement
[22]. Its ability to quickly assess signal irregularity makes it transitions [25]. This feature is particularly useful in detecting
particularly suited for real-time applications, such as wearable abrupt changes, such as sudden starts, stops, or shifts in
activity monitoring and fall detection. Given the constraints velocity, allowing for better differentiation between dynamic

Authorized licensed use limited to: American International University Bangladesh. Downloaded on July 30,2026 at 07:43:56 UTC from IEEE Xplore. Restrictions apply.
and static activities. It also enables a more thorough evaluation of the model’s
By incorporating these temporal features, the model en- predictive power.
hances its ability to classify activities effectively, ensuring
D. Classification Models
more accurate recognition of movement patterns and improv-
ing overall system performance. To classify the 10 different activities based on accelerometer
6) Recursive Feature Elimination with Cross-Validation: data, four machine learning models were employed: Random
Recursive Feature Elimination with Cross-Validation Forest (RF), Multi-layer Perceptron (MLP), Extreme Gradient
(RFECV) is a wrapper method for feature selection that Boosting (XGBoost), and a set of semi-supervised learning
identifies the optimal subset of features for a machine learning models, including Self-Training Random Forest (SelfTraining
model [26]. It combines Recursive Feature Elimination (RFE), RF), Label Propagation, and Label Spreading. Each model
which iteratively removes the least important features based on was configured with tuned hyperparameters to optimize clas-
the estimator’s internal measures, with Cross-Validation (CV), sification performance. Furthermore, these models including
which rigorously evaluates model performance at each step XGBoost and Random Forest have been successfully applied
to ensure generalizability [26]. By systematically eliminating in other Parkinson’s disease detection studies using different
features and assessing performance via CV, RFECV selects types of data [28].
the subset yielding the highest cross-validation score (e.g., • Extreme Gradient Boosting (XGBoost) is a powerful en-
accuracy or F1-score), thereby improving model performance, semble learning model based on boosting, where multiple
reducing complexity, and potentially accelerating computation decision trees are trained sequentially to correct the errors
[27]. of previous trees [28]. This allows XGBoost to capture
7) Feature set and model parameter Optimization: To opti- both linear and non-linear relationships within the data,
mize the model’s performance, we applied Recursive Feature making it highly effective for complex datasets such as
Elimination with Cross-Validation (RFECV). This technique accelerometer readings. Furthermore, its ability to handle
iteratively removes the least important features while evaluat- imbalanced datasets is beneficial for this study, where
ing model performance through cross-validation, ensuring that different activities have varying sample distributions. The
only the most relevant features are retained. Additionally, we model was configured with 100 estimators, a maximum
conducted a systematic search to determine the most suitable tree depth of 6, a learning rate of 0.1, and a multi-
number of estimators, exploring different values (5, 10, and class softmax objective function to handle the 10 activ-
100). By testing these configurations, we identified the optimal ity classes. To optimize performance, Recursive Feature
balance between model complexity and accuracy. Elimination with Cross-Validation (RFECV) was applied,
ensuring that the model focuses on the most relevant
C. Validation features, improving accuracy and efficiency.
• Random Forest (RF) is an ensemble learning algorithm
To ensure the reliability and generalizability of the models, that constructs multiple decision trees using random sub-
two primary validation strategies were employed: sets of the data and features, improving generalization
1) Train-Test Split: The dataset was divided into two and reducing overfitting. It is particularly effective for
subsets: a training set and a test set, using a 70/30 ratio, high-dimensional data and serves as a strong baseline
where 70% of the data was utilized for model training, and model for comparison with more advanced approaches.
30% was reserved for testing and evaluation. This split ratio The RF model in this study was configured with 100
strikes a balance between having enough data for training the trees and a maximum depth of 6, maintaining a balance
classifiers while also ensuring a fair evaluation of the model’s between complexity and computational efficiency. Its
performance on unseen data. It helps to verify that the models inherent ability to capture diverse patterns in data makes
are trained on a diverse set of examples, and the evaluation it well-suited for activity recognition.
reflects their ability to generalize to new, previously unseen • Multi-layer Perceptron (MLP) is a neural network de-
activities. signed to model complex, non-linear relationships in data.
2) Cross-Validation: To further enhance model perfor- The MLP model used in this study consists of two hidden
mance and mitigate the risk of overfitting, cross-validation was layers with 82 and 32 neurons, respectively, utilizing
applied during training. Specifically, k-fold cross-validation ReLU activation functions and the Adam optimizer. The
with k = 5 was employed. In this technique, the training data model was trained for up to 500 iterations to ensure
is partitioned into 5 subsets. The model is trained on 4 subsets convergence. Unlike tree-based methods, MLP captures
and tested on the remaining subset. This process is repeated higher-order dependencies through multiple layers, mak-
5 times, with each subset serving as the test set once. The ing it particularly effective for recognizing subtle varia-
final performance metrics are then averaged over the 5 runs, tions in time-series accelerometer data.
providing a more reliable estimate of the model’s performance. • Semi-supervised learning techniques were employed to
This approach helps ensure that the models are evaluated on leverage both labeled and unlabeled data, enhancing
multiple subsets of the data, enhancing their reliability and classification performance. Label Propagation and Label
reducing the likelihood of overfitting to any specific partition. Spreading were implemented with radial basis function

Authorized licensed use limited to: American International University Bangladesh. Downloaded on July 30,2026 at 07:43:56 UTC from IEEE Xplore. Restrictions apply.
 (RBF) and k-nearest neighbors (KNN) kernels, with TABLE II
hyperparameters tuned to gamma = 0.2 and k = 7, P ERFORMANCE COMPARISON OF FOUR CLASSIFIERS ( IN PERCENT )
respectively. These methods iteratively assign labels to
XGBoost RF MLP SelfTraining RF
unlabeled data, improving classifier generalization. Fur-
thermore, Self-Training Random Forest (SelfTraining RF) Accuracy 40 34 38 42
was employed, using a base Random Forest classifier Macro F1 36 27 34 37
with 100 trees, a maximum depth of 20, and a minimum Weighted F1 40 32 38 41
samples split of 5. A confidence threshold of 0.75 was F1 score class 1 45 44 46 53
applied to incorporate high-confidence pseudo-labeled F1 score class 2 37 29 40 38
samples into training, mitigating the scarcity of labeled
F1 score class 3 22 21 22 26
data in real-world applications.
F1 score class 4 38 26 31 30
These models were trained and validated to assess their
F1 score class 5 30 3 23 35
effectiveness in classifying activities based on accelerometer
data. The following section presents the classification results F1 score class 6 34 24 35 42
and a comparative analysis of their performance. F1 score class 7 9 0 10 10
F1 score class 8 42 32 40 38
E. Test result generation
F1 score class 9 47 43 43 42
F1 score class 10 54 50 49 53

B. Criteria Analysis
1) Accuracy: The SelfTraining RF model achieves the
highest accuracy at 42%, followed closely by XGBoost at
40%. While Random Forest and MLP have lower accuracies
Fig. 3. Data processing pipeline
of 34% and 38%, respectively, they still show strengths in
classifying certain individual classes.
The proposed pipeline processes accelerometer sensor data
by aligning test file timestamps with the corresponding sensor 2) F1 Score: The F1 score for each class show clear differ-
readings, applying segmentation, feature extraction, and stan- ences across the models. XGBoost performs exceptionally well
dardization, and then generating preliminary activity labels. in most individual F1 score, especially for classes with distinct
The figure above illustrates this workflow, showing how each motion characteristics, such as class 1 (Sitting Position) and
component is integrated to produce the final predictions. class 10 (Standing Position). The model achieves the highest
In the first stage, the sensor data is divided into fixed 3- F1 score for these classes, with 54% for class 10 and 45%
second time windows, with each window grouping together all for class 1, indicating its ability to accurately classify actions
rows in the test file whose timestamps fall within that interval. with distinctive movement patterns.
For each segment, the classification component assigns an
initial activity label based on the extracted features. This C. Model Analysis
approach ensures consistency between the data processing 1) SelfTraining RF: Although SelfTraining RF achieves
steps used in both the development and testing phases. the highest overall accuracy (42%), it does not consistently
Finally, to produce the final labels, the predicted labels outperform XGBoost across all classes. For some classes, like
within each time window are aggregated by selecting the class 3 and class 5, the F1 scores are lower compared to
mode of those labels. This mode label is then assigned to XGBoost, even though SelfTraining RF performs better in
every row in the corresponding interval. By consolidating the certain other classes. This suggests that while SelfTraining
dominant activity within each segment, the pipeline enhances RF can leverage unlabeled data, it does not always outperform
the reliability of the final submission while maintaining a other models.
straightforward and interpretable classification process. 2) MLP: MLP shows stable but mediocre performance,
with F1 scores that do not stand out in any class. This
IV. R ESULTS AND A NALYSIS
could be due to MLP’s difficulty in learning subtle patterns
A. Result in movement data, particularly when dealing with complex
The classification results presented in Table II and the action variations. MLP struggled to classify classes like class
corresponding confusion matrices illustrate the performance 2 (Standing Slow Walk) and class 4 (Walking), with F1 scores
of four classifiers in activity recognition: Extreme Gradient of only 29% and 31%, respectively.
Boosting (XGBoost), Random Forest, Multi-layer Perceptron 3) Random Forest: Random Forest performs the least well,
(MLP) and Semi-supervised Self-Training Random Forest with an accuracy of 34%, the lowest among the four models.
(SelfTraining RF). Although it shows some strong performance in certain classes

Authorized licensed use limited to: American International University Bangladesh. Downloaded on July 30,2026 at 07:43:56 UTC from IEEE Xplore. Restrictions apply.
 Fig. 4. Confusion Matrices for Semi-Supervised, Random Forest, MLP, and XGBoost Models in Activity Recognition

(such as class 9 and class 10), it does not achieve the same standardized, which affected the quality of training data. Addi-
level of precision as XGBoost in other classes. tionally, the data collected were from real-world environments,
4) XGBoost: XGBoost achieves 40% accuracy, excelling or ”in the wild,” making it more susceptible to inconsistencies,
in standing (54% F1) and sitting (45% F1). It struggles with such as missed data and noise. The presence of environmental
jumping (22% F1) and lying down (9% F1). It has a 36% factors and variations in user behavior also contributed to the
Macro F1 and 40% Weighted F1, outperforming Random complexity of distinguishing between activities accurately. To
Forest and MLP, while SelfTraining RF reaches 42% accuracy. improve the accuracy and generalizability of the model, we
XGBoost demonstrates superior performance in accurately propose augmenting the dataset by incorporating additional
classifying activities with distinct movement patterns, achiev- data sources, such as videos, images, and other health pa-
ing high F1 scores in many classes, especially those that rameters. By integrating multimodal data, the model could
are more challenging to classify, such as class 10 (Standing achieve a more comprehensive understanding of the subject’s
Position). This model has proven effective at handling the motor activity, leading to better predictions and enhanced
complex features of Parkinsonian movement data. generalizability. This approach would also help mitigate issues
Although SelfTraining RF achieved the highest overall related to missed or inconsistent data by providing richer
accuracy (42%), it lacks consistency compared to XGBoost in contextual information.
all classes. Some classes were misclassified, while XGBoost Among many methods for this project, these include Semi-
maintained higher accuracy across more classes. supervised and XGBoost. XGBoost is favored over semi-
Other models like Random Forest and MLP showed limita- supervised methods for accelerometer data due to the frequent
tions in accurately classifying activities, especially when subtle availability of sufficient labeled data, allowing XGBoost’s
motion changes were involved, further supporting XGBoost as high-accuracy supervised learning. Semi-supervised methods,
the optimal model for this task. while useful with limited labels, suffer from instability due
For these reasons, XGBoost is the best choice for this to confirmation bias and sensitivity to unlabeled data quality.
problem, thanks to its ability to classify accurately and con- XGBoost’s efficiency and robustness, coupled with effective
sistently across multiple classes, especially in a complex data feature engineering, make it a preferred choice.
environment requiring fine-grained.
B. Stacking model comparison
V. D ISCUSSION
A stacking ensemble comprising Random Forest, XGBoost,
A. Preference of XGboosts over Semi-supervised method semi-supervised Random Forest, and MLP achieved a slightly
One of the key limitations lies in the data acquisition pro- higher F1-score (43%) compared to the standalone XGBoost
cess. The labels provided by the organizers were insufficiently model. While ensemble methods are known to improve pre-

Authorized licensed use limited to: American International University Bangladesh. Downloaded on July 30,2026 at 07:43:56 UTC from IEEE Xplore. Restrictions apply.
dictive performance by leveraging the strengths of diverse STOP/frozen, full body shaking, rotate then return back (9),
base learners, they often introduce increased computational show higher spectral centroid values due to high-frequency
complexity and reduced interpretability. In contrast, XGBoost components. As shaking motions typically occur vertically,
offers a more efficient and interpretable solution, with faster the z-axis effectively captures these fluctuations. Moreover,
inference and better suitability for deployment in practical the average power feature—representing energy content along
scenarios. Although stacking provides valuable insights and the y-axis—differentiates high-intensity movements like Walk
indicates potential performance gains, XGBoost remains the (LEFT → Right → Left) (8) from lower-energy activities
preferred model due to its balance of accuracy, efficiency, and such as Cooldown – sitting/relax (7). Since walking involves
real-world applicability. continuous vertical displacement, energy variations along the
y-axis are more pronounced, highlighting its role in movement
classification.
The Petrosian fractal dimension of the y-axis quantifies
signal complexity and helps identify irregular, chaotic move-
ments. In (Sideway) both hands SHAKING (sitting) (5), high
values indicate erratic, involuntary hand tremors, primarily
captured through vertical oscillations, underscoring its im-
portance for detecting nonlinearity in tremor-related actions.
Additionally, the area under the curve (AUC) of y-axis ac-
celeration encapsulates movement magnitude over time. This
feature effectively differentiates dynamic activities, such as
Slow walk (SHAKING hands/body, tiny step, head forward)
(10), from static postures like (Sideway) Sit stand (4), with
continuous movements producing higher AUC values.
Each feature was carefully selected for its ability to capture
specific movement characteristics. The mean vertical accelera-
tion and its AUC capture vertical displacement, distinguishing
static from dynamic activities. The spectral centroid of z-axis
acceleration effectively detects high-frequency fluctuations re-
lated to tremors and shaking, while the correlation between
Fig. 5. Confusion Matrices for Stacking model (from Semi-Supervised, RF,
MLP, XGBoost) in Activity Recognition
x- and y-axis accelerations reveals inter-axis dependencies
indicative of coordinated limb movements. Furthermore, the
Petrosian fractal dimension quantifies signal complexity and
C. Discussions on some feature analysis
characterizes irregular movements, and the average power
To enhance action recognition, we analyze six representative of vertical acceleration differentiates high-intensity actions
features extracted from accelerometer data across different from low-energy postures. Collectively, these multi-domain
domains. Figure 6’s boxplots display the distribution of these features substantiate our diversified feature selection strategy
features over 10 activity labels, providing insights into their for optimizing recognition accuracy.
classification capabilities.
Notably, the mean acceleration along the y-axis is essential D. Discussion on some features
for distinguishing activities with significant vertical displace-
ment. For example, static postures such as (FACING camera) Based on the confusion matrices and F1 scores from the
Sit and stand (1) exhibit stable y-axis means, while dynamic four models (XGBoost, Random Forest, MLP, and Semi-
actions like Walk & STOP/frozen, full body shaking, rotate supervised Self-Training RF), several observations emerge
then return back (9) show greater variation, confirming its rel- regarding activity classification.
evance in capturing posture transitions. The correlation feature 1) Easy-to-Recognize Activities: Activities with distinct,
quantifies the dependency between x- and y-axis movements, clear patterns—such as Activity 1 (Sit and Stand), Activity
making it particularly useful for identifying coordinated hand 9 (Walk & STOP/frozen, full body shaking, rotate then return
motions. In (FACING camera) both hands SHAKING (sitting back), and Activity 10 (Slow walk with shaking hands/body,
position) (2), high correlation values indicate synchronized tiny step, head forward)—are more easily recognized. For
movement between the hands and torso. Conversely, lower cor- example, Activity 1 shows rapid acceleration changes during
relations in activities like Slow walk (SHAKING hands/body, transitions, Activity 9 exhibits clear stop-and-shake fluctua-
tiny step, head forward) (10) suggest more independent motion tions, and Activity 10 maintains steady acceleration. These
between limbs and the body. contrasts in dynamic versus stable patterns provide strong
Derived from the z-axis, the spectral centroid feature learning signals, suggesting that further refinements in feature
measures the distribution of frequency components. Activ- extraction and model algorithms can enhance classification
ities with tremors and rapid oscillations, such as Walk & performance.

Authorized licensed use limited to: American International University Bangladesh. Downloaded on July 30,2026 at 07:43:56 UTC from IEEE Xplore. Restrictions apply.
 Fig. 6. Boxplot Analysis of Accelerometer Features Across 10 Activity Classes.
(a) Scatter Box Plot of y mean by Activity.
(b) Scatter Box Plot of correlation xy by Activity.
(c) Scatter Box Plot of spectral centroid z by Activity.
(d) Scatter Box Plot of average power y by Activity.
(e) Scatter Box Plot of petrosian f d y by Activity.
(f) Scatter Box Plot of auc y by Activity.

2) Easily Misclassified Activities: Activities such as Ac- Stand” with “Stand up with shaking” produce similar
tivity 2 (Both hands SHAKING – sitting) and Activity 5 accelerometer readings. The models struggle to distin-
(Sideway Both hands SHAKING – sitting) are challenging due guish subtle differences in movement direction or shaking
to similar motion patterns. Despite differing body orientations, intensity, leading to lower F1 scores.
the hand-shaking movements yield comparable accelerometer • Imbalanced Data: Additionally, imbalanced class distri-
signatures, with similar frequency and amplitude, making it butions can increase misclassification rates, particularly
difficult for the model to differentiate between them. Similarly, for activities with similar motion characteristics, thereby
Activity 3 (Stand up from chair with shaking) is prone to affecting overall model.
misclassification because its shaking motion overlaps with
those of similar activities.
3) Reasons for Easy Recognition or Misclassification: In summary, activities with clear, sharp transitions in move-
• Easily Recognized Activities: Activities such as “Sit ment are easier for models to classify, whereas those with
and Stand,” “Walk & Stop,” and “Slow Walk” exhibit similar dynamic patterns or subtle differences tend to be
strong, distinct acceleration patterns and clear changes in misclassified. Analyzing both misclassified and easily rec-
movement direction, which the models readily capture, ognized activities allows us to refine our algorithms and
resulting in high F1 scores. improve feature extraction methods, ultimately leading to more
• Easily Misclassified Activities: In contrast, activities like effective models. Furthermore, addressing the imbalance in the
“Both hands shaking” in different positions or “Sit & dataset is essential to enhance overall classification accuracy.

Authorized licensed use limited to: American International University Bangladesh. Downloaded on July 30,2026 at 07:43:56 UTC from IEEE Xplore. Restrictions apply.
E. Limitations and Challenges [8] M. Sun, W. Jung, K. Koltermann, G. Zhou, A. Watson, G. Blackwell,
N. Helm, and L. Cloud, “Parkinson’s disease action tremor detection
While the pipeline demonstrates effective performance, sev- with supervised-learning models,” Proceedings of the 2023 International
eral challenges persist: Data inconsistencies, including missing Conference on AI and Mobile Services (AIMS 2023), pp. 1–8, 2023, doi:
10.1145/3580252.3586977.
values, noise, and user behavior variations, may impact result [9] R. A. Bhuiyan, N. Ahmed, M. Amiruzzaman, and M. R. Islam, “A robust
reliability. Additionally, imbalanced class distributions intro- feature extraction model for human activity characterization using 3-axis
duce potential model bias, while the high dimensionality of accelerometer and gyroscope data,” Sensors, vol. 20, p. 6990, oct 2020,
doi: 10.3390/s20236990.
extracted features increases computational complexity, neces- [10] S. Hossain, T. Hossain, T. Tazin, M. I. Mamun, C. Garcia, T. Hujioka,
sitating optimization strategies. S. Inoue, and M. A. R. Ahad, “Summary of the normal versus unusual
activity detection challenge in parkinson’s disease using multimodal
VI. C ONCLUSION data,” IJABC, vol. 2025, no. 2, 2025, doi: 10.60401/ijabc.103.
[11] O. D. Lara and M. A. Labrador, “A survey on human activity recognition
A. Key Contributions and Implications using wearable sensors,” IEEE Communications Surveys Tutorials,
vol. 15, pp. 1192–1209, 2013, doi: 10.1109/SURV.2012.110112.00192.
This study presents an integrated framework for human [12] G. Casella and R. L. Berger, Statistical Inference. Duxbury Press, 2002.
activity recognition using wearable accelerometer data, with [13] J. E. Freund and R. E. Walpole, Mathematical Statistics with Applica-
tions, 7th ed. Pearson Prentice Hall, 2006.
a focus on Parkinsonian movement analysis. The proposed [14] S. Mallat, A Wavelet Tour of Signal Processing, 3rd ed. Academic
pipeline combines multi-domain feature extraction, RFECV- Press, 2009.
based feature selection, and the evaluation of multiple clas- [15] J. G. Proakis and D. G. Manolakis, Digital Signal Processing: Princi-
ples, Algorithms, and Applications, 5th ed. Pearson, 2021.
sifiers. Among these, XGBoost was selected for its effi- [16] I. Guyon and A. Elisseeff, “An introduction to variable and feature
ciency and interpretability in distinguishing 10 distinct activity selection,” Journal of Machine Learning Research, vol. 3, no. Mar, pp.
classes. These findings underscore the importance of combin- 1157–1182, 2003.
[17] E. Gómez-Luna, D. E. Cuadros-Orta, J. E. Candelo-Becerra, and J. C.
ing carefully engineered features with an appropriate model Vasquez, “The development of a novel transient signal analysis: A
choice to address the challenges of time-series data in practical wavelet transform approach,” Computation, vol. 12, no. 9, p. 178, 2024,
wearable monitoring applications. doi: 10.3390/computation12090178.
[18] F. Luft, S. Sharifi, W. Mugge, A. C. Schouten, L. J. Bour, A.-F. van
Rootselaar, P. H. Veltink, and T. Heida, “A power spectral density-
B. Future Work based method to detect tremor and tremor intermittency in movement
Future research will explore the integration of additional disorders,” Sensors, vol. 19, no. 9, 2019.
[19] M. Bouda, J. S. Caplan, and J. E. Saiers, “Box-counting dimension revis-
sensor modalities and advanced learning techniques to further ited: Presenting an efficient method of minimizing quantization error and
enhance classification performance. Expanding the dataset and an assessment of the self-similarity of structural root systems,” Frontiers
incorporating deep learning methods may improve the sys- in Plant Science, vol. 7, p. 149, 2016, doi: 10.3389/fpls.2016.00149.
[20] S. Mitra, M. A. Riley, and M. T. Turvey, “Chaos in human rhythmic
tem’s ability to capture subtle movement variations. These de- movement,” Journal of Motor Behavior, vol. 29, no. 3, pp. 195–198,
velopments aim to refine the proposed framework and extend 1997, doi: 10.1080/00222899709600834.
its applicability, ultimately contributing to more accurate and [21] B. Nasseroleslami, C. J. Hasson, and D. Sternad, “Rhythmic manip-
ulation of objects with complex dynamics: Predictability over chaos,”
accessible solutions for activity recognition and Parkinson’s PLOS Computational Biology, vol. 10, no. 10, p. e1003900, 2014, doi:
disease monitoring. 10.1371/[Link].1003900.
[22] S. Sharanya and S. P. Arjunan, “Fractal dimension techniques for anal-
R EFERENCES ysis of cardiac autonomic neuropathy (can),” Biomedical Engineering:
Applications, Basis and Communications, vol. 35, no. 3, p. 2350003,
[1] M. I. Hosen and M. B. Islam, “Forecasting wearing-off in parkinson’s 2023, doi: 10.4015/S1016237223500035.
disease: An ensemble learning approach using wearable data,” in Activ- [23] D. Anguita, A. Ghio, L. Oneto, X. Parra, and J. L. Reyes-Ortiz, “A public
ity, Behavior, and Healthcare Computing. CRC Press, pp. 324–334. domain dataset for human activity recognition using smartphones,”
[2] F. Latifoğlu, S. Penekli, F. Orhanbulucu, and M. E. Chowdhury, “A in Proceedings of the 21st European Symposium on Artificial Neural
novel approach for parkinson’s disease detection using vold-kalman Networks, Computational Intelligence and Machine Learning (ESANN
order filtering and machine learning algorithms,” Neural Computing and 2013), 2013, pp. 437–442.
Applications, vol. 36, no. 16, pp. 9297–9311, 2024. [24] J. L. Reyes-Ortiz, A. Samà, X. Parra, and C. Ronao, “Human activity
[3] O.-B. Tysnes and A. Storstein, “Epidemiology of parkinson’s disease,” recognition on smartphones using wearable sensors,” Journal of Ambient
Journal of neural transmission (Vienna, Austria : 1996), vol. 124, no. 8, Intelligence and Smart Environments, vol. 08, 2012.
p. 901—905, August 2017, doi: 10.1007/s00702-017-1686-y. [25] J. M. Vazquez, M. R. Garcia, , and J. M. Quevedo, “Jerk-based
[4] R. Balestrino and A. H. Schapira, “Parkinson disease,” European Journal human activity recognition using a tri-axial accelerometer,” International
of Neurology, vol. 27, pp. 27–42, 2020, special Issue on Pervasive Journal of Computational Intelligence Systems, vol. 10, 2017.
Healthcare. [26] P. Misra and A. S. Yadav, “Improving the classification accuracy
[5] E. Rovini, C. Maremmani, and F. Cavallo, “How wearable sen- using recursive feature elimination with cross-validation,” Int. J. Emerg.
sors can support parkinson’s disease diagnosis and treatment: A sys- Technol, vol. 11, no. 3, pp. 659–665, 2020.
tematic review,” Frontiers in Neuroscience, vol. 11, oct 2017, doi: [27] J. Sung, S. Han, H. Park, S. Hwang, S. J. Lee, J. W. Park, and I. Youn,
10.3389/fnins.2017.00555. “Classification of stroke severity using clinically relevant symmetric gait
[6] K. L. Chou, M. Stacy, T. Simuni, J. Miyasaki, W. H. Oertel, K. Sethi, features based on recursive feature elimination with cross-validation,”
H. H. Fernandez, and F. Stocchi, “The spectrum of ”off” in parkinson’s Ieee Access, vol. 10, pp. 119 437–119 447, 2022.
disease: What have we learned over 40 years?” vol. 51, pp. 9–16, june [28] V. Dentamaro, D. Impedovo, L. Musti, G. Pirlo, and P. Taurisano,
2018, doi: 10.1016/[Link].2018.02.001. “Enhancing early parkinson’s disease detection through multimodal deep
[7] A. Weiss, S. Sharifi, M. Plotnik, J. P. P. van Vugt, N. Giladi, and J. M. learning and explainable ai: insights from the ppmi database,” Scientific
Hausdorff, “Toward automated, at-home assessment of mobility among Reports, vol. 14, no. 1, p. 20941, 2024.
patients with parkinson disease, using a body-worn accelerometer,”
Neurorehabilitation and Neural Repair, vol. 25, pp. 810–818, 2011, doi:
10.1177/1545968311424869.

Authorized licensed use limited to: American International University Bangladesh. Downloaded on July 30,2026 at 07:43:56 UTC from IEEE Xplore. Restrictions apply.