The role of signal preprocessing on the discriminability of canonical time-series characteristics and classification among individuals with and without Parkinson’s disease during serious game interaction

1 Introduction

The global prevalence of Parkinson’s disease (PD) has been rising rapidly in recent decades (Tolosa et al., 2021; Bloem et al., 2021). Currently, more than 6 million people live with the disease, representing a 2.5-fold increase within a single generation. Although PD is rare in individuals under 40, it affects up to 3% of the population over 80 (Tanner and Ostrem, 2024). PD is a chronic, progressive disorder of the central nervous system and is now recognized as a leading cause of neurological disability worldwide (Tanner and Ostrem, 2024; Tolosa et al., 2021; Bloem et al., 2021). It is a multifactorial condition influenced by both genetic and non-genetic factors, including environmental exposures and inheritance. As the disease progresses, non-motor symptoms become increasingly prominent. Core motor symptoms include tremor, rigidity, bradykinesia, and gait abnormalities (Farhah, 2024; Tanner and Ostrem, 2024).

The MDS–Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) is the standard clinical tool for assessing motor and non-motor PD symptoms. Over the past decade, numerous studies have attempted to quantify PD symptoms objectively, yet no device has achieved widespread clinical adoption. To address this gap, our team developed RehaBEElitation (Cardoso Mendes et al., 2023), a serious game patented (BR512021001975-0) by the National Institute of Industrial Property. Its source code is openly available, enabling broad use in research and clinical environments. The game records hand motor data as patients interact with it, supported by a wearable glove equipped with inertial sensors and conductive threads (Mendes et al., 2021; Rosa et al., 2021). The intuitive narrative of RehaBEElitation aligns with players’ mental models, enhancing engagement, interaction, and acceptance (Mendes et al., 2022). The embedded sensors in the glove provide high-resolution data capable of quantifying tremor and other motor symptoms more reliably than visual observation, supporting integration with computational approaches such as machine learning to improve monitoring, treatment evaluation, and research (Patel et al., 2009; Espay et al., 2016).

Time series, such as those recorded by RehaBEElitation, are complex due to their amplitude, frequency, noise, transient fluctuations, and overlapping dynamic patterns, making analysis difficult. To assess PD symptoms from signals, advanced processing steps are needed, such as time series decomposition to isolate relevant components (Mendes et al., 2022). Having tools to simplify this processing would be a major advance because it is methodologically complex and computationally expensive. Analysis and methodological efficiency could improve if representative and discriminative motor characteristics could be estimated from the captured signals.

Classification of time series data is difficult due to its diversity that may be related to the presence of trends, frequency and magnitude variability, irregularities, discontinuities, local patterns and global patterns. These characteristics led to the development of several approaches to extract features and classify time-series (Tan et al., 2025). Some methods are based on similarity measures (e.g., Proximity Forest (Lucas et al., 2019) and Proximity Forest 2.0 (Tan et al., 2025)). Some combine different ways to compare time-series, for instance, similarity, repeated patterns and the occurrence of values in specific time intervals (e.g., TS-CHIEF - Time Series Combination of Heterogeneous and Integrated Embedding Forest (Shifaz et al., 2020)). Some methods use different scaling techniques to detect and learn fast and slow patterns automatically (e.g., InceptionTime (Ismail Fawaz et al., 2020)). Some use linear models (logistic regression) and multiple symbolic representations of time series for classification (e.g., Mr-SEQL (Le Nguyen et al., 2019)). Some approaches are focused on dealing with the presence of noise in the time-series (e.g., Bag of Symbolic Fourier Approximation Symbols (BOSS) (Schäfer, 2014), and Scalable Dictionary Classifiers for Time Series Classification (cBOSS) (Middlehurst et al., 2019)).

According to Dempster et al. (2020), the methods discussed above suffer from major limitations, including high computational complexity, limited scalability, and substantial computational cost. These challenges motivated the development of the RandOm Convolutional KErnel Transform (ROCKET), which applies random convolutional kernels to time series and uses the resulting features to train a linear classifier. Building on this idea, (Dempster et al., 2021), introduced MINIROCKET, a reformulation of ROCKET that is up to 75 times faster on large datasets and is almost deterministic, while maintaining essentially the same level of accuracy.

Lubba et al. (2019) introduced catch22 (CAnonical Time-series CHaracteristics), a curated set of 22 high-performing time-series features selected from more than 7000 features available in hctsa (Highly Comparative Time-Series Analysis (Fulcher et al., 2013)). These features were chosen based on their classification performance across 93 real-world time-series classification tasks. By summarizing complex signals into a compact and interpretable set of descriptors, feature-based representations provide a substantial reduction in dimensionality while retaining essential information. In a recent study, the catch22 feature set was applied to EEG data from PD patients during cycling exercises of varying intensity, revealing dynamic neural patterns associated with motor symptoms like tremor and bradykinesia (Alizadeh et al., 2025).

The present study aims to evaluate the use of the catch22 feature set in extracting features from time series of movement recorded by inertial sensors during interaction with a serious game. The extraction of the catch22 feature set from the signals recorded during interaction with the serious game may enable the objective quantification of motor behaviour descriptors, complementing traditional clinical assessments. The aim is to investigate the potential of this approach in discriminating between healthy individuals and patients with PD under different clinical conditions (on and off medication states), in order to verify its utility as a complementary measure to traditional clinical assessments. The expectation is that the use of the catch22 feature set can simplify signal processing while simultaneously offering greater sensitivity in identifying subtle motor changes, contributing to the advancement of objective methods for disease assessment.

To evaluate the applicability of the catch22 feature set in characterising motion time series obtained from inertial sensors during interaction with a serious game, an exploratory approach was adopted, consisting of multiple preprocessing steps followed by the extraction of the 22 canonical time-series features. Different pipelines—that is, combinations of preprocessing procedures and classification models—were compared through participant-stratified cross-validation, using nested cross-validation for hyperparameter tuning and rigorous procedures to avoid data leakage, ensuring that all transformations were estimated exclusively on the training sets.

In light of recent advancements in time-series classification for Parkinson’s disease, such as the application of RandOm Convolutional KErnel Transform (ROCKET) (Donié et al., 2025; Dempster et al., 2020) and deep learning architectures, there is a growing need to establish how traditional signal conditioning affects these high-dimensional methods. This study not only evaluates the impact of preprocessing on the catch22 feature set but also provides a comparative analysis with the ROCKET framework. By utilizing a Ridge Classifier with cross-validation (RidgeClassifierCV), we aim to position our wavelet-based preprocessing pipeline within the current landscape of state-of-the-art PD monitoring tools, ensuring a robust evaluation against established benchmarks in the literature (Daneault et al., 2021; Khan et al., 2025).

In this study, we address the following research questions: (RQ1) How do different signal conditioning strategies (raw, interpolation, wavelet filtering, and combined interpolation and wavelet) affect the discriminability of compact canonical time-series representations (catch22) for PD-related classification? (RQ2) Are the observed effects consistent across binary (CG vs. PDG) and multi-class (CG vs. PDG-ON vs. PDG-OFF) settings under participant-stratified evaluation? Our main contributions are: (C1) a controlled comparison of four preprocessing pipelines tailored to inertial signals collected during a serious game and (C2) an evaluation across 9 ML learners under nested group-wise cross-validation.

2 Methods

2.1 RehaBEElitation serious game and interface device

RehaBEElitation is a bee-themed serious game developed within the Brazilian Federal Agency for Support and Evaluation of Graduate Education (CAPES) and French Committee for the Evaluation of University Cooperation with Brazil (COFECUB), project (23038.003097/2019-07), a Brazil–France cooperation program that supports joint research projects and academic exchange between universities of both countries, to support monitoring and rehabilitation of people with PD. Designed by a multidisciplinary team, the game immerses users in a 3D environment where they control a bee through movements aligned with the MDS-UPDRS Part III motor assessment, such as hand flexion/extension, wrist adduction/abduction, finger tapping, and forearm pronation/supination. Structured in four phases (Figure 1), the game gamifies therapeutic exercises to increase patient motivation and engagement. For a detailed description of each phase, see Cardoso Mendes et al. (2023).

Using inertial sensors incorporated to a glove that promotes communication between the real movements and the game (Figure 2), it also records motor performance, enabling the identification of clinical markers related to bradykinesia, tremor, and rigidity. This combination of entertainment, monitoring, and therapy provides an innovative tool for supporting PD management.

2.2 Experimental setup

This study was approved by the Human Research Ethics Committee of the Federal University of Uberlândia, Minas Gerais, Brazil. Each eligible participant agreed to participate in the study by signing a consent form. It involved fifteen individuals diagnosed with PD, who formed the group of individuals with PD (PDG), and fifteen healthy individuals, who composed the control group (CG). Participants in both groups were matched by age and sex to ensure comparability.

Individuals with PD interacted with the game in both on and off medication states, allowing for the evaluation of potential differences in motor function. Prior to gameplay, participants in the experimental group were assessed with the MDS-UPDRS Part III by a trained physiotherapist, ensuring consistent evaluation across conditions.

Participants were seated on a chair approximately 1 m away from the TV screen displaying the serious game. They wore the glove device on the dominant hand in the case of healthy individuals or on the hand most affected by the disease in the case of individuals with PD. The limb was supported on a forearm rest to facilitate the required movements. This setup ensured that the interaction with the game could be assessed comfortably and reliably.

2.3 Data processing and analysis workflow

All signal processing and data analysis steps were performed using R (version 4.5.0) (R Core Team, 2025), a programming language and an open source for data visualisation and statistical analysis, in RStudio environment (version 2025.9.0.387).

Our workflow began with the collection of triaxial accelerometer signals from an inertial sensor embedded in the glove worn by participants while playing the serious game, RehaBEElitation. All acquired signals were then carefully visualized to verify their integrity and identify potential issues. To comprehensively assess the capacity of the features to manage signals exhibiting various types of signal processing issues, we employed four distinct strategies. The first strategy used the raw, unprocessed signals. The second strategy introduced a signal preprocessing step to address discontinuities by using piecewise cubic Hermite interpolation. The third strategy focused on filtering the signals to remove trends and voluntary activity using wavelet decomposition. Finally, the fourth strategy combined the second and third approaches, applying both interpolation and wavelet decomposition.

For each of these four strategies, the extracted features were used as inputs for nine machine learning algorithms. We first assessed hyperparameter tuning for each model using an internal 5-fold cross-validation. Subsequently, the tuned models were trained and tested using an external 5-fold cross-validation, ensuring that observations from the same individual were not present in both the training and test sets. We evaluated the models on two separate classification tasks: a binary task to distinguish between healthy individuals (CG) and patients with PD (PDG), and a more challenging multi-class task to differentiate between healthy individuals (CG) and patients under on (PDG-ON) and off (PDG-OFF) medication.

2.3.1 Signal visualization

All recorded signals were carefully visualised in order to verify signal integrity and identify potential trends, noise and discontinuities that may require specific preprocessing steps. In total, 405 (15 individuals per groups × 3 groups × 3 game phases per individual × 3 axes) signals were visually inspected, revealing trends attributable to the influence of gravity and voluntary movements performed during the game (i.e., voluntary activity) on the accelerometer measurements, as well as occasional discontinuities caused by sample loss during signal acquisition. Thus, this step was very important in order to support subsequent analyses and outlining the proposed preprocessing approaches.

During the signal visualization stage, we discovered that the second phase of the game contained discontinuities lasting over one second, which were caused by a temporary sensor disconnection during the game play. Due to the significant nature of these gaps in signals from this phase, we made the decision to exclude all signals corresponding to this phase from our processing and analysis. This ensured that our subsequent analyses were based on reliable signals with minimal interference from discontinuities, allowing for a more consistent assessment of how the 22 features addressed this processing issue and thereby enhancing the integrity of the results.

2.3.2 Signal preprocessing

In order to verify the relevance of different signal conditioning steps combined with the performance of catch22 features as a predictor to discriminate between healthy individuals and patients with PD under on and off-medication states, four signal preprocessing strategies were adopted. For each strategy, distinct steps were implemented to account for the manner in which catch22 feature set address specific signal processing challenges, thereby enabling the identification of the most appropriate conditioning approach for feature extraction and data modeling.

2.3.2.1 First strategy

To evaluate the performance of canonical time-series features in discriminating between groups without any signal conditioning before feature extraction, the first strategy did not include any signal preprocessing steps. Therefore, our initial processing strategy aimed to assess the discriminative power of features from the catch22 set without any prior signal preprocessing. Consequently, all features were extracted directly from each individual’s raw accelerometer signals.

2.3.2.2 Second strategy

The second strategy focused on addressing signal discontinuities resulting from sample loss during signal acquisition before feature extraction. In fact, sample loss during signal recording is a typical issue in biomedical signal processing and sensor-based studies, often resulting from noise, hardware limitations, transmission errors, among others factors (El-Sheimy et al., 2008). Missing samples can influence downstream analyses, including feature extraction and data modelling. To reduce the influence of sample loss, preprocessing strategies like interpolation are commonly adopted to estimate missing values and reconstruct the signal. Selecting an appropriate interpolation method ensures that the reconstructed signal maintains its characteristics, such as trends and waveform shape (Najarian and Splinter, 2005).

Thus, to address discontinuities identified both through visual inspection (e.g., identification of abrupt changes and sudden jumps in the signals) and by analyzing the uniformity of temporal resolution between consecutive samples over the entire signal (i.e., the distribution of the first discrete derivative of the sample times), each signal was resampled from 50 Hz to 200 Hz, using Piecewise Cubic Hermite Interpolating Polynomial (PCHIP). This interpolation method was adopted to enhance the waveform representation of the signals. Unlike standard cubic splines, this method preserves monotonicity, preventing overshoot and ensuring that the interpolated values remain within the range of the original signal.

PCHIP, defined in Equation 1, constructs a piecewise cubic interpolant for each sub-interval [ x k , x k + 1 ].where h 00 ( t ) , h 10 ( t ) , h 01 ( t ) and h 11 ( t ) denote the Hermite basis functions, p 0 = y k and p 1 = y k + 1 correspond to the interpolated values at the endpoints of the sub-interval, and m 0 and m 1 represent the corresponding endpoint derivatives.

p t = h 00 t p 0 + h 10 t x k + 1 − x k m 0 + h 01 t p 1 + h 11 t x k + 1 − x k m 1 , (1)

The interpolant P ( x ) is constructed piecewise such that, for each interval, it satisfies P ( x k ) = y k and P ( x k + 1 ) = y k + 1, with P ( x ) and P ′ ( x ) continuous on the entire domain. However, the second derivative, P ″ ( x ), is not necessarily continuous at the knots { x j } , ∀ j ∈ { 0,1 , … , n }, which may result in discontinuities in curvature across sub-intervals.

Thus, the interpolant is shape-preserving, since the slopes at x j are chosen to preserve shape and respect monotonicity (Fritsch and Carlson, 1980). This property enables the method to handle missing samples while preserving signal trends and enhancing the waveform representation, as illustrated in Figure 3.

2.3.2.3 Third strategy

The third strategy’s goal was to evaluate the performance of the features in discriminating between groups after the removal of trends and the reduction of the influence of voluntary activity on signals. To achieve this, the signals were decomposed into four levels using the Daubechies orthonormal wavelet with a compact support of length 8 (L = 8).

This study employed a signal processing methodology using wavelet decomposition, similar to a recent study on the effect of vibrotactile stimulation on essential tremor (Cabral et al., 2024). The analysis of biomedical signals, especially those related to involuntary movements like tremors, often faces significant challenges, including the presence of artifacts, such as the influence of voluntary activity while playing the serious game, and linear and non-linear trends that can mask the intrinsic characteristics of the phenomenon of interest. Wavelet decomposition is a robust tool for mitigating these issues, as it allows for multi-resolution analysis that separates signal components into different frequency bands.

2.3.2.3.1 Wavelet decomposition methodology

Wavelet decomposition was used as an important step in signal preprocessing. The primary goal was to eliminate trends commonly found in inertial sensor signals, and to reduce the influence of voluntary activity on the signals. This process isolates the signal components and allows components that are most representative of involuntary activity be obtained. The choice of wavelet decomposition is justified by its ability to analyze non-stationary biomedical signals across different time and frequency scales simultaneously.

2.3.2.3.2 Wavelet selection and decomposition levels

For the analysis, the authors chose to decompose the signals into four components using the Daubechies orthonormal wavelet with a compact support of length 8 (db8). The Daubechies wavelet family is widely recognized for its orthogonality and compact support properties, which are beneficial for analyzing signals with transient characteristics and for accurate signal reconstruction. The db8 wavelet provides a good balance between the smoothness of the scaling function and its localization in both time and frequency, making it suitable for detecting subtle variations in biomedical signals.

The original signal was decomposed into four detail components (d 1, d 2, d 3, and d 4) and one approximation component ( s 4 ). In a multi-level wavelet decomposition, detail components represent high-frequency information at different scales, while the approximation component retains the remaining low-frequency information. A detailed description of each of these components and their significance is provided below:

• Detail Components (d 1 to d 4): Each detail component captures the high-frequency content of the signal within a specific frequency band. Component d 1 corresponds to the highest frequencies, and subsequent components (d 2, d 3, and d 4) correspond to progressively lower frequency bands. In the context of the analyzed signals, these components may contain both the signal from involuntary activities, voluntary movements assessed while playing the serious game, and high-frequency noise.
• Approximation Component ( s 4 ): This component represents the low-frequency oscillations of the signal (trends). In many biomedical signal processing contexts, baseline trends, slow voluntary movements, and other slow variations are captured by this approximation component.

2.3.2.3.3 Filtering and signal reconstruction strategy

The filtering strategy used in the study was based on the selective identification of wavelet components corresponding to noise, voluntary movements, or undesirable trends. Once identified, these components were excluded, and the original signal was reconstructed from the remaining components. This method allows for adaptive and precise filtering, preserving the essential characteristics of the signal of interest.

2.3.2.3.4 Trend and voluntary game movements removal

Inertial signals, such as those obtained from accelerometers, are often susceptible to linear and non-linear trends. These trends can be attributed to various factors, such as drift, misalignment of the input axis, temperature, noise, small movements unrelated to the limb being investigated, or even artifacts from the signal acquisition process (e.g., cable movement). The presence of these trends can distort the analysis of the phenomenon of interest, making their removal a critical step in preprocessing.

In the study, trends related to the influence of gravitational acceleration and voluntary movements made along the axis were associated with low-frequency signal oscillations and, consequently, were captured by the approximation component ( s 4 ). Additionally, undesired components related to faster voluntary movements and hesitations during the game were better described by the detail components d 4 and d 3 for signals from the control group and by the detail component d 4 for signals from the group of individuals with PD. Therefore, the strategy for removing these trends consisted of reconstructing the original signal by excluding the approximation component ( s 4 ) for both groups and the detail components (d 3 and d 4), depending on the group being analyzed. This means the filtered signal was obtained by the algebraic sum of the detail components (d 1 to d 3), resulting in a signal free from the slow variations that characterize trends and influence of voluntary movements.

For feature extraction, the signals resulting from the third preprocessing strategy corresponded to the reconstruction by means of the algebraic sum of components d 1 and d 2 for signals from the control group and the sum of components d 1 to d 3 for signals from individuals with PD. Figures 4, 5 illustrate this process, showing two original signals with trends (one from the control group and one from the PD group, respectively) and their respective reconstructed signals after removing the components.

2.3.2.4 Fourth strategy

The fourth strategy corresponded to a combination of the second and third preprocessing approaches, i.e., discontinuities were addressed, trends were removed and the influence of voluntary activity was reduced. Thus, all signals were resampled to 200 Hz and subsequently filtered, using wavelet decomposition (Figure 6).

2.4 Feature extraction

From signals resulting from each preprocessing strategy, the set of 22 Canonical Time-Series Characteristics (catch22) were extracted. Each signal was represented by 22 features, as detailed in Table 1. This compact feature set attempts to characterize time series with minimal redundancy by capturing linear and nonlinear autocorrelation, successive differences, value distributions and outliers, and fluctuation scale dynamics.

TABLE 1. Description of extracted features from signals.

To facilitate interpretation, these features were originally designed to be grouped into categories according to the underlying properties of the time series they describe. Specifically, the feature set captures distributional properties, extreme events, linear and nonlinear autocorrelation, symbolic patterns, forecasting behaviour, incremental differences, and self-affine scaling, providing a comprehensive summary of the time series (Lubba et al., 2019).

All available features were derived from the Highly Comparative Time Series Analysis (hctsa) toolbox, which originally introduced over 7,000 candidate features to characterize time series. However, the features in catch22 were selected through a systematic evaluation process with the aim of maintaining high classification performance while reducing redundancy and computational complexity (Lubba et al., 2019).

Essentially, catch22 feature set transforms a time series into a 22-dimensional feature space. The goal is to obtain a representation that is computationally efficient, interpretable, and well suited for tasks such as classification, clustering, and anomaly detection. According to previous studies, this process is approximately a thousand times faster than the original hctsa toolbox.

2.5 Data modeling

The objective of data modeling was to perform a supervised classification using the catch22 feature set obtained from each preprocessing strategy as predictors, distinguishing both between individuals with and without Parkinson’s disease (i.e., CG and PDG), as well as between healthy individuals and individuals diagnosed with PD under the on and off-medication states (i.e., CG, PDG-ON and PDG-OFF).

In total, nine algorithms were used to classify the data, employing the mlr3 ecosystem that provides a unified interface in R for model training, resampling, hyperparameter tuning, and performance evaluation (Lang et al., 2019). In this study, the models used were Random Forest (RF), Extreme Gradient Boosting, Single Layer Neural Network, Classification Tree, Support Vector Machine, Generalized Linear Models (GLM) with Elastic Net Regularization, k-Nearest-Neighbor, Naive Bayes, and Multinomial Log-Linear Model.

For the data modeling, mlr3 was used because it provides a modern, extensible architecture that supports reproducible machine learning workflows. Compared to older frameworks such as caret or the original mlr, mlr3 offers superior modularity through object-oriented programming (R6-based abstractions), allowing clear separation of tasks, learners, resampling strategies, and evaluation metrics. This design not only ensures transparency and comparability across experiments but also facilitates the integration of complex modeling pipelines. The extended mlr3 ecosystem (e.g., mlr3tuning, mlr3pipelines) further supports hyperparameter optimization and workflow automation, making it a more robust and flexible choice than traditional, monolithic alternatives.

For each model, hyperparameter tuning was assessed, considering a search space as described in Table 2. The search within the defined space was conducted using Grid Search with a resolution equal to five. The performance of each hyperparameter combination was evaluated by internal cross-validation with five partitions (5-fold cross-validation), using balanced accuracy as the performance metric and respecting the data’s grouping structure so that observations from the same individual were always allocated to the same fold.

TABLE 2. Description of the search space for the used models.

After hyperparameter tuning, each model was trained and tested using an external 5-fold cross-validation, also respecting the data’s grouping structure to prevent individual observations from overlapping between the training and test sets. Finally, aggregated performance metrics were calculated to obtain overall estimates for each model in binary (between CG and PDG) and multiclass (between CG, PDG-ON and PDG-OFF) classifications. To evaluate both general performance and the ability to handle class imbalances, different metrics were estimated, including accuracy (ACC), balanced accuracy (BACC), sensitivity (recall), precision, and F1-score. For binary classification, performance was evaluated using accuracy, balanced accuracy, sensitivity, precision, and F1-score. For multiclass classification, overall performance was assessed using accuracy, balanced accuracy, logarithmic loss (Log loss), and classification error (CE), while per-class evaluation included balanced accuracy, sensitivity, precision, and F1-score.

2.6 Comparative benchmarking

To evaluate catch22 with different classifiers, the Random Convolutional Kernel Transform (ROCKET) combined with a Ridge classifier, a state-of-the-art approach for time-series representation and classification, was also used as a benchmark.

In contrast to the catch22 features, ROCKET produces a high-dimensional representation of the time series, capturing temporal patterns through convolutional transformations. This feature extraction method involves generating random kernels, convolving each time series with the kernels, and summarizing the convolutional responses to obtain features. The kernels used in ROCKET vary in length, weights, bias, dilation, and padding, allowing patterns to be detect at multiple temporal scales. For each kernel, the convolutional responses are summarized using two statistics: the maximum value and the proportion of positive values (Dempster et al., 2019).

Unlike the extraction of the 22 features, the high-dimensional features obtained using ROCKET were derived from windowed segments of the signals. Due to variability in signal duration across participants and game phases, each preprocessed signal was segmented using nonoverlapping rectangular windows of 10 s prior to feature extraction. As a result, multiple fixed-length time series were obtained per individual for each game phase. For instance, a signal lasting 302 s was represented by 30 distinct 10-s signal segments.

For feature extraction, the set of fixed-length time series was divided into training and test sets using a 5-fold outer cross-validation scheme, with all time series from a participant assigned exclusively to one set to prevent data leakage. Within each fold, ROCKET was applied using 1,000 randomly generated kernels. The transformation was fitted solely on the training data and then applied to both sets, converting the original time series into a high-dimensional feature space based on two summary statistics: the maximum value and the proportion of positive values. Thus, after applying the ROCKET transformation, both the training and test sets were represented in a 2,000-dimensional feature space.

Classification was performed using the RidgeClassifier, a linear model with L 2 regularization, chosen for its computational efficiency and robustness in high-dimensional settings (Dempster et al., 2019). The regularization parameter was optimized via grid search over 10 logarithmically spaced values ranging from 1 0 − 3 to 1 0 3. This optimization utilized a 5-fold internal cross-validation while maintaining participant-level stratification, with balanced accuracy as the optimization metric.

The performance of both binary and multiclass classification was evaluated using the same metrics applied to catch22 with the other classifiers. The entire pipeline was implemented in Python (v3.12.12) using a parallelized architecture to optimize computational throughput across 44 logical cores.

2.7 Data and code availability

The database supporting the findings of this study are not publicly available due to ethical and privacy considerations. However, processed data (e.g., the extracted catch22 and ROCKET features) can be obtained from the corresponding author upon reasonable request. Additionally, the scripts used for data preprocessing, feature extraction, and classification can be obtained from the authors upon request.

3 Results

To provide a comprehensive overview of the impact of signal conditioning on Parkinson’s disease classification, we first present a synthesis of the performance metrics across the four evaluated preprocessing strategies (Tables 3, 4). This summary highlights the comparative effectiveness of each approach in both binary and multi-class scenarios, serving as a guide for the detailed analysis of model behaviors and feature discriminability presented in the subsequent subsections. As observed, the transition from raw data to sophisticated wavelet-based filtering marks a pivotal shift in the system’s ability to accurately distinguish between clinical states.

TABLE 3. Best balanced accuracy (BACC) results for the binary classification task (CG vs. PDG) across preprocessing strategies.

TABLE 4. Best balanced accuracy (BACC) results for the multi-class classification task (CG/ON/OFF) across preprocessing strategies.

In Tables 3, 4 results are presented as balanced accuracy for the best-performing machine learning model in each scenario. Note the substantial performance gain achieved through wavelet-based decomposition (Strategies 3 and 4) compared to raw and interpolated signals.

3.1 First strategy

After analyzing the performance metrics for the first data processing strategy, which used raw, unprocessed signals, a clear pattern emerges, as shown in Table 5; Figure 7. The GLM with Elastic Net Regularization model stands out as the top performer, achieving a balanced accuracy of 0.71, the highest among all models. It also demonstrated excellent sensitivity (0.60), precision (0.62), and a relevant F1-score (0.51), indicating a good balance between correctly identifying positive cases and avoiding false positives. The Multinomial Log-Linear Model and Extreme Gradient Boosting also showed relevant results, with balanced accuracies of 0.65 and 0.61, respectively.

TABLE 5. Performance metrics for each machine learning model using features from the first data processing strategy (raw signals).

However, a more rigorous analysis of these results reveals that a balanced accuracy around 0.60 indicates a limited ability to effectively discriminate between the groups. These scores suggest that, while some models performed better than others, the lack of signal conditioning severely hampered their performance. The raw signals contain significant noise, voluntary movements and confounding variables that prevent even the more advanced algorithms from achieving a higher level of confidence. This is particularly evident in the poor performance of simpler models like Random Forest (0.56 balanced accuracy) and Support Vector Machine (0.55 balanced accuracy). Their low sensitivity and F1-scores indicate a clear limited ability to correctly identify individuals with PD, highlighting the critical need for proper signal preprocessing. Overall, the results show that, despite the efforts of some algorithms, the raw accelerometer signals present a significant challenge, making this strategy largely ineffective for a robust classification task.

The first processing strategy, which used raw, unprocessed signals, also provided valuable insights into the multi-class classification problem—differentiating between healthy individuals (CG), Parkinson’s patients on medication (PDG-ON), and patients off medication (PDG-OFF). The overall performance metrics for this more complex task were significantly lower, as shown in Tables 6, 7. This highlights the inherent difficulty of distinguishing between three groups using noisy, raw data. The Multinomial Log-Linear Model was the top performer, achieving a balanced accuracy of 0.48, which, while the best, still indicates a limited ability to classify all three groups accurately.

TABLE 6. Overall performance metrics for each machine learning model in the multi-class classification of the control group (CG), Parkinson’s disease group on medication (PDG-ON), and Parkinson’s disease group off medication (PDG-OFF).

TABLE 7. Performance metrics by class for each machine learning model in the multi-class classification of the control group (CG), Parkinson’s disease group on medication (PDG-ON), and Parkinson’s disease group off medication (PDG-OFF).

A more detailed examination of Table 7 reveals a more nuanced picture of how each model exhibited limitations with this task. For instance, the Random Forest and Single Layer Neural Network models had particularly difficulty with the PDG-ON group, with sensitivities of only 0.19 and 0.24, respectively. This suggests that the features extracted from raw signals were not sufficient to reliably identify individuals who were on medication. In contrast, models like GLM with Elastic Net Regularization and the Multinomial Log-Linear Model showed a better, more balanced performance across all three groups, correctly classifying the healthy group with a high sensitivity (0.64 and 0.60, respectively). The varied performance across classes underscores the complexity of this task and demonstrates the limitations of using raw data for fine-grained distinctions between patient groups.

3.2 Second strategy

The second data processing strategy, which involved resampling signals from 50 Hz to 200 Hz using piecewise cubic Hermite interpolation to address missing samples, produced mixed results. As shown in Table 8; Figure 8, the overall performance metrics showed some improvements but also some declines compared to the first strategy (using raw signals).

TABLE 8. Performance metrics for each machine learning model using features from the second data processing strategy (resampled signals).

The Extreme Gradient Boosting model delivered the best overall performance, with a balanced accuracy of 0.62. This indicates that it was the most effective model in correctly classifying both the control and Parkinson’s groups, even with the data having undergone interpolation. The k-Nearest-Neighbor and GLM with Elastic Net Regularization models also performed reasonably well, with balanced accuracies of 0.61 and 0.63 respectively. However, the models that previously performed well, like the Multinomial Log-Linear Model, saw a decrease in balanced accuracy from 0.65 to 0.57, suggesting that interpolation did not help this model’s ability to discriminate between the groups.

The most notable result is the performance of the Naive Bayes model, which had a very high sensitivity of 0.74 but a very low precision of 0.36. This indicates that while the model was excellent at correctly identifying individuals with PD (true positives), it did so at the cost of incorrectly classifying many healthy individuals as having the disease (false positives). This imbalance is further reflected in its low F1-score of 0.44, which penalizes such a disparity between precision and recall. Conversely, the Support Vector Machine model showed the lowest performance, with a balanced accuracy of just 0.51 and very low sensitivity (0.19), making it largely ineffective for this classification task.

The results from this strategy suggest that simple interpolation is not sufficient. While it can enhance the signals and improve performance for some models like Extreme Gradient Boosting and k-Nearest-Neighbor, it can also disrupt the underlying patterns in a way that harms the performance of other models, such as the Multinomial Log-Linear Model and Naive Bayes. This underscores the need for more sophisticated signal processing techniques that can better capture the nuances of the data.

The second strategy, which involved interpolating signals to address discontinuities, also presented an opportunity to explore the multi-class classification problem (CG vs. PDG-ON vs. PDG-OFF). However, the results, shown in Tables 9, 10, reveal that simple interpolation alone was not enough to produce robust classifications. The overall performance was low, with the Multinomial Log-Linear Model again leading the pack with a balanced accuracy of just 0.46, followed closely by Extreme Gradient Boosting with a balanced accuracy of 0.43. These scores highlight the significant challenge of classifying multiple groups with data that, despite being preprocessed, still contains substantial underlying noise, presence of voluntary activity and complex patterns.

TABLE 9. Overall performance metrics for each machine learning model in the multi-class classification of the control group (CG), Parkinson’s disease group on medication (PDG-ON), and Parkinson’s disease group off medication (PDG-OFF).

TABLE 10. Performance metrics by class for each machine learning model in the multi-class classification of the control group (CG), Parkinson’s disease group on medication (PDG-ON), and Parkinson’s disease group off medication (PDG-OFF).

A closer look at the per-class metrics shows where the models exhibited their greatest limitations. The sensitivities for the PDG-ON and PDG-OFF groups were particularly low across almost all models. For instance, the Single Layer Neural Network had a sensitivity of only 0.14 for the PDG-ON group and 0.22 for the PDG-OFF group. Similarly, Support Vector Machine and Naive Bayes showed very low sensitivities for these patient groups. The Multinomial Log-Linear Model demonstrated a slightly better ability to classify healthy individuals (CG) with a sensitivity of 0.62, but its performance dropped sharply for the patient groups. This consistent pattern of low sensitivity indicates that interpolation, while addressing signal gaps, fails to make the fine-grained distinctions necessary to reliably separate these specific patient populations.

3.3 Third strategy

The third processing strategy, which involved wavelet decomposition to remove trends and reduce the influence of voluntary activity, yielded a significant improvement in model performance. As shown in Table 11; Figure 9, all models achieved a balanced accuracy of 0.92 or higher, indicating that this preprocessing step was highly effective.

TABLE 11. Performance metrics for each machine learning model using features from the third data processing strategy (wavelet-decomposed signals).

The Extreme Gradient Boosting and Support Vector Machine models were the top performers, both achieving a balanced accuracy of 0.96. Their high scores across all metrics—including accuracy, sensitivity, and recall—demonstrate their robust ability to accurately classify individuals with and without PD after the signals had been properly conditioned. The Random Forest and Multinomial Log-Linear Model also achieved remarkable results, with balanced accuracies of 0.94, representing a substantial increase from their performance on the raw and interpolated data.

Even models that had difficulty with previous strategies, such as the Single Layer Neural Network and k-Nearest-Neighbor, showed a remarkable improvement, with balanced accuracies of 0.94 and 0.92, respectively. The Naive Bayes model also improved significantly, with a balanced accuracy of 0.93 and a high sensitivity of 0.95. This shows that the wavelet decomposition effectively isolated the most relevant features of the involuntary movements, leading to a much clearer distinction between the two groups. The overall performance of all models suggests that the wavelet decomposition was a crucial and highly effective step in mitigating the effects of noise and voluntary activity.

The results clearly demonstrate that properly preprocessing the signals is essential for achieving high classification performance. The significant increase in balanced accuracy across all models confirms that the wavelet decomposition strategy successfully mitigated the effects of noise, trends, and voluntary activity in the signals, making the features more discriminative for the classification task.

The third strategy, which used wavelet decomposition to remove trends and voluntary activity, produced much better results for the multi-class classification (CG vs. PDG-ON vs. PDG-OFF). As shown in Table 12, the overall balanced accuracy for all models improved significantly, with the Random Forest model leading at 0.67. This demonstrates that filtering the signal was an essential step for handling the complexity of a three-group classification.

TABLE 12. Overall performance metrics for each machine learning model in the multi-class classification of the control group (CG), Parkinson’s disease group on medication (PDG-ON), and Parkinson’s disease group off medication (PDG-OFF) for the third processing strategy.

Looking at the metrics by class, Table 13 reveals how this improvement translated to individual group performance. For the healthy group (CG), most models showed near-optimal classification, with models like Random Forest, XGBoost, and Naive Bayes achieving a sensitivity of 0.96, 0.96, and 0.94 respectively. However, the patient groups (PDG-OFF and PDG-ON) were still a considerable challenge. For instance, the Naive Bayes model had a high sensitivity of 0.74 for the PDG-ON group, but only 0.30 for the PDG-OFF group, indicating it underperformed in identifying patients who were off medication. In contrast, models like Random Forest and XGBoost showed a more balanced, although still modest, performance on the patient groups, with sensitivities ranging from 0.43 to 0.53. This highlights the difficulty in separating the two patient sub-groups, even with a strong preprocessing step.

TABLE 13. Performance metrics by class for each machine learning model in the multi-class classification of the control group (CG), Parkinson’s disease group on medication (PDG-ON), and Parkinson’s disease group off medication for the third processing strategy.

3.4 Fourth strategy

The fourth data processing strategy combined the second and third approaches—interpolation to handle discontinuities followed by wavelet decomposition to remove trends and voluntary activity. This two-step preprocessing yielded the most robust and highly specific results for classification. The provided figures and tables demonstrate the performance of this strategy, both in a binary classification (GC vs. PDG) and a more complex multi-class classification (CG vs. PDG-ON vs. PDG-OFF).

In the binary classification scenario (CG vs. PDG), the results are remarkable. As shown in Table 14, nearly all models achieved a perfect score across all metrics, with an accuracy, balanced accuracy, and recall of 1.00. The Multinomial Log-Linear Model was the only one that did not achieve a perfect score in all metrics, with a precision of 0.97 and a recall of 1.00. This is a significant improvement over all previous strategies, demonstrating that the combined approach successfully created a feature set with almost perfect separability between the two groups. The boxplots in Figure 10 also visually support this, showing a much clearer separation between the distributions of the features for the two groups compared to the previous strategies.

TABLE 14. Performance metrics for each machine learning model in the binary classification of the control group (CG) and the Parkinson’s disease group (PDG).

However, the real test of this strategy’s effectiveness is in the multi-class classification, where the goal was to differentiate between healthy individuals (CG), Parkinson’s patients on medication (PDG-ON), and Parkinson’s patients off medication (PDG-OFF). This task is much more challenging, and the results, while still good, reflect this complexity.

As detailed in Table 15, the Random Forest model performed best overall, with a balanced accuracy of 0.76. The Extreme Gradient Boosting and Multinomial Log-Linear Model also performed well, with balanced accuracies of 0.71 and 0.69, respectively. Table 16 provides a deeper look, showing the strengths and weaknesses of each model for each class. For instance, the Naive Bayes model shows a high sensitivity of 0.80 for the PDG-ON group, but a low sensitivity of 0.20 for the PDG-OFF group, indicating it underperforms when identifying patients who are off medication.

TABLE 15. Overall performance metrics for each model in the multi-class classification of CG, PDG-OFF, and PDG-ON.

TABLE 16. Performance metrics for each class (CG, PDG-OFF, PDG-ON) in the multi-class classification for each model.

In contrast, the Random Forest and XGBoost models demonstrate a more balanced performance. For the PDG-OFF group, Random Forest achieved a sensitivity of 0.53, while XGBoost reached 0.53. For the PDG-ON group, Random Forest had a sensitivity of 0.56, while XGBoost reached 0.61. This shows that while all models faced challenges with the multi-class problem, the ensemble methods were better at handling the nuances and differences between the two Parkinson’s groups.

Ultimately, the combined preprocessing strategy proved to be extremely effective. It successfully enabled nearly perfect binary classification and significantly improved the performance on the more complex multi-class problem, confirming its superiority over the other three strategies.

3.5 Comparative performance: catch22 vs. ROCKET benchmark

To contextualize the performance of the proposed pipeline, a comparative benchmarking analysis was conducted against the ROCKET-RidgeClassifierCV framework. As summarized in Tables 17–19, this framework achieved a balanced accuracy of 100.0% in the binary classification and 69.0% in the multi-class classification under fourth signal processing strategy (i.e., combining interpolation and wavelet decomposition for signal conditioning before feature extraction and classification). Under the same processing strategy, the catch22 combined with different classifiers yielded 100.0%, while catch22-RF achieved 76.0%, respectively.

TABLE 17. Performance metrics for each strategy in the binary classification (CG vs. PDG) using ROCKET-RidgeClassifierCV.

TABLE 18. Overall performance metrics for each strategy in the multi-class classification of CG, PDG-OFF, and PDG-ON using ROCKET-RidgeClassifierCV.

TABLE 19. Performance metrics for each class in the multi-class classification using ROCKET-RidgeClassifierCV.

These results suggest that, while ROCKET provides a high-dimensional representation, the use of the compact catch22 feature set may offer improved discriminative performance in more complex multi-class scenarios when combined with the most suitable signal processing strategy. In practice, the proposed pipeline relies on a markedly lower number of features (22 features) than the 2,000 features produced by ROCKET using 1,000 kernels, which may represent an advantage in terms of computational efficiency, particularly in real-time serious game applications.

4 Discussion

4.1 The role of serious games in data collection

There is currently no cure for Parkinson’s disease. Pharmacological treatments focus on mitigating symptoms, while rehabilitative therapy aims to improve motor function and prevent cognitive decline. Physical rehabilitation is a vital part of managing PD, as it can significantly enhance patients’ quality of life and functionality. Traditionally, this method involves a multidisciplinary team of physical and occupational therapists, but these methods often face challenges like poor adherence and limited access to specialized services (Quirosa-Galán et al., 2024).

Recently, virtual reality (VR) and video games have become promising rehabilitative tools. VR creates interactive, customizable environments, and video games use playful challenges to boost patient motivation and engagement. These technologies encourage active participation by providing sensory feedback. VR, in particular, has become more prominent in rehabilitation due to its low cost and versatility, allowing rehabilitation exercises to be performed at home. In immersive VR, the user is completely integrated into the virtual world, using glasses or gloves (Sevcenko and Lindgren, 2022; Quirosa-Galán et al., 2024).

Serious games are increasingly used as innovative tools in healthcare to boost patient engagement and adherence to therapy by integrating entertaining, interactive elements into treatment regimens. Designed to address specific medical needs, these games offer personalized therapeutic activities tailored to individual patient requirements. By leveraging technology, serious games facilitate remote monitoring and management, allowing for continuous, non-invasive data collection and sophisticated analysis. This approach enables healthcare providers to track patient progress in real time, support clinical decision-making, and adjust treatment plans as needed. Overall, serious games aim to sustain patients’ health, enhance their quality of life, and provide motivating, enjoyable experiences that encourage ongoing participation in therapy (Tarousi et al., 2021).

The use of digital games represents an important differential in the objective assessment of patients with PD. During the interaction, the movements can be recorded in real-time and stored in a database, which allows for subsequent quantitative analyses with greater precision compared to traditional clinical assessments, often influenced by the examiner’s subjectivity. The developed game has already been duly registered with the National Institute of Industrial Property (INPI), under number BR512021001975-0.

4.2 Preprocessing as a key to accurate classification

PD is characterized by motor signs such as bradykinesia, tremor, rigidity, and postural instability, which can be quantified by the signals obtained during the execution of the movements proposed in the game. The processing of these signals can enable the identification of movement patterns and the measurement of characteristics related to disease progression. In this context, the catch22 feature set were used in an attempt to quantify different measures of time series resulting from the game. The establishment of standardized, compact and interpretable metrics facilitates an objective representation of signals for disease analysis. However, the findings of this study demonstrates that signal preprocessing can play a decisive role in the development of machine learning-based systems for PD classification when employing this feature set, in contrast to the simplicity originality proposed by Lubba et al. (2019).

The results from the four signal processing strategies highlight the role of signal preprocessing in discriminating between individuals with and without PD, using the catch22 features. Our initial approach, using raw accelerometer signals, demonstrated the baseline discriminative power of the catch22 features. The modest performance, with balanced accuracies generally ranging between 0.55 and 0.71, suggests that the set of estimated features is not able to appropriately describe signals affected by common signal processing issues, such as discontinuities resulting from loss sample during signal acquisition, as well as complex signals such as those resulting from the game interaction, in which there is an influence of voluntary movements necessary to control the bee in the game environment and involuntary movements (e.g., the manifestation of tremors and postural instability during voluntary movements). Models like Random Forest and Support Vector Machine particularly exhibited limited performance, underscoring their sensitivity to estimated features of unprocessed signals. This finding indicates that reliance on a powerful algorithm alone, in the absence of appropriate signal conditioning, is insufficient to address a complex classification task based on time-series properties. Thus, characterizing the raw signals of the different groups with the catch22 features can be challenging and still not allow the description of intrinsic characteristics of both involuntary and voluntary movements of the individuals.

The second strategy, which addressed signal discontinuities through interpolation, provided distinct results among the models. Although some models like Extreme Gradient Boosting and k-Nearest-Neighbor exhibited a slight performance boost, others experienced reduced performance relative to the initial strategy. This difference may be attributed to the sensitivity of catch22 features to discontinuities in signals. As a result, signal interpolation may have removed patterns arising from these discontinuities that were being used by some classifiers to differentiate groups, thereby impairing classification performance in certain models. This hypothesis is raised because some features of catch22 depend on the distribution shape, temporal dependence, and occurrence of extreme events. Interestingly, among the models, the Naive Bayes model demonstrated the highest sensitivity, but with very low precision. Therefore, even though signal interpolation improves signal representation, it alone is insufficient to discriminate complex signals, such as those from gaming, using catch22 features.

The third strategy, which employed wavelet decomposition to remove trends and reduce voluntary activity, demonstrated to be the most important preprocessing step for distinguishing between signals from individuals with and without PD. This preprocessing method led to a consistent improvement in classification performance. All models exhibited a balanced accuracy consistently exceeding 0.92, indicating their capacity to differentiate the groups. In particular, the Extreme Gradient Boosting and Support Vector Machine models were the top performers, with a balanced accuracy of 0.96. The improved model performance observed can be explained by the isolation of involuntary activity components of the signals, using wavelet decomposition. This is because these components typically differ between individuals with and without the disease, leading to greater separation between the group distributions for the extracted features (Figure 9). Thus, this finding underscores the importance of obtaining components that better differentiate signals, allowing for better characterization of signals through 22 features.

In the accelerometer signals, wavelet decomposition enabled the discrimination of distinct signal components, the removal of trends, and the reduction of the influence of voluntary movements. The choice of the db8 wavelet and the filtering strategy based on the exclusion of specific components (d 3, d 4, and s 4 for trends) were fundamental for isolating relevant information from the accelerometer signal. This rigorous preprocessing is crucial for ensuring the validity and accuracy of subsequent time and frequency domain analyses, significantly contributing to the understanding of the effects of involuntary movements during serious game play. In addition, the ability to isolate the signal of interest from inherent artifacts and noise in biomedical data underscores the value of wavelet decomposition as an indispensable tool in signal processing research.

Finally, the fourth strategy, which combined interpolation and wavelet decomposition, achieved the best performance across the evaluated approaches. For the simple binary classification (CG vs. PDG), the models achieved near-perfect performance, with balanced accuracies of 1.00. These results suggest that a multi-step preprocessing pipeline plays an important role in appropriately conditioning this type of data for the extraction of the 22 features and binary classification. While the multi-class classification (CG vs. PDG-ON vs. PDG-OFF) proved more challenging, the results were still the best among all strategies. The Random Forest and Extreme Gradient Boosting models performed best, demonstrating that even with the added complexity of distinguishing between different medication states, a set of features extracted from properly conditioned signals allows for better classification. The differences in performance across the multi-class groups highlight the subtle variations in movement patterns for patients in their on- and off-medication states, which the models were able to partially capture. This may reflect several factors, including the potential beneficial or adverse effects of medication on the motor symptoms of individuals.

The results from multi-class classification underscore the challenges of distinguishing healthy individuals from individuals with the disease on and off medication and further emphasize the potential importance of signal preprocessing. The models based on features extracted from signals obtained using the first two preprocessing strategies consistently exhibited poor performance, underscoring the difficulty of classifying the differences using features extracted from raw or interpolated signals. In these scenarios, most models had difficulty distinguishing between the groups of individuals with PD (PDG-ON and PDG-OFF), as evidenced by their low sensitivities. This suggests that trends, voluntary movements, artifacts, and overlapping features in the unprocessed signals can mask group-specific motor differences, limiting classification consistency using the feature set. However, the application of wavelet decomposition in the third strategy led to an improvement in model performance. Although the performance metrics remained low, the models generally demonstrated improved discrimination among the three groups, with Random Forest and XGBoost showing the most balanced performance. This reinforces our primary conclusion: adequate signal conditioning is not just a beneficial step but a necessity for unlocking the subtle, discriminative patterns in biomedical data, especially when attempting to differentiate between clinically relevant patient sub-groups, using features like catch22.

4.3 Benchmarking against state-of-the-art time-series transforms

Our findings regarding the impact of signal conditioning align with and extend recent advancements in the field. While Donié et al. (2025) demonstrated the high performance of ROCKET and InceptionTime architectures for symptom estimation, the present comparative benchmarking suggests that carefully designed preprocessing strategies, such as wavelet decomposition, can reduce the performance gap between compact feature sets and high-dimensional feature space. Notably, while ROCKET-RidgeClassifierCV achieved perfect performance in the binary scenario, the catch22-RF approach showed higher balanced accuracy in the more challenging multi-class task (76.0% vs. 69.0%) when signals were preprocessed using interpolation and wavelet decomposing.

This suggests that for complex clinical states–such as distinguishing between on and off medication–targeted signal conditioning may be more effective at isolating subtle physiological markers than generic kernel-based transformations. Furthermore, our approach addresses the challenges of movement disorder detection highlighted by Khan et al. (2025) by effectively mitigating voluntary movement artifacts. By achieving superior discriminability with a substantially lower feature dimensionality (22 vs. 2,000 features), our pipeline provides a computationally efficient and interpretable alternative to deep learning models, making it highly suitable for real-time, game-based clinical monitoring.

4.4 Beyond the scope: limitations and future perspectives

While this study demonstrates the potential of the proposed pipeline to discriminate between individuals with and without Parkinson’s disease, certain limitations must be acknowledged. First, the findings were based on a specific cohort with a limited number of participants, which may limit the generalizability of the findings to other populations, as well as fail to capture all the inter-individual variability observed between individuals with and without the disease.

Second, the results were derived on short-term signals recorded during game-based tasks, in which participants were tasked with accomplishing specific goals within a time limit per game phase. This time-constrained design may limit the capture of longer-term motor signs. While it enables precise differentiation of individuals during short, task-specific assessments, offering focused information on motor performance under well-defined monitoring or rehabilitation tasks, this time limitation may restrict the observation of long-term motor patterns.

Third, although the pipeline can discriminate between individuals with and without the disease based solely on the involuntary activity of individuals during interaction with the game, it is limited by the absence of processing steps aimed at specifically extracting remarkable motor signs of the disease, such as bradykinesia, dyskinesias, and motor fluctuations. This may restrict the generalization of the proposed method to alternative experimental settings, requiring additional signal processing steps to effectively handle cases with more severe disease progression.

Fourth, ROCKET combined with the RidgeClassifierCV may exhibit sensitivity to the window size used for signal segmentation, potentially affecting feature representation and classification performance. In this study, the window size was chosen to standardize the time series and provide signal segments long enough to capture meaningful activity (e.g., avoiding periods consisting solely of rest, which could reduce group differences) and remaining computationally feasible. However, the window size can influence both the patterns captured and the computational cost of the convolutions.

This is because smaller windows produce shorter time series, which can enhance the detection of rapid variations or short-lived events in signals, but may also generate noisier or less stable convolutional responses. In contrast, larger windows yield longer time series, which are more suitable for capturing long-term patterns or gradual changes, yet may obscure brief events, attenuate rapid fluctuations, and increase the computational cost of the convolutions.

Finally, another limitation of this study is the lack of a formal evaluation of the computational cost associated with each feature extraction and modelling method in order to verify the practical use of each method. In this study, ROCKET combined with RidgeClassifierCV approach required more processing time and computational resources compared to catch22 and the other classifiers, necessitating the use of a dedicated high-performance environment to enable a parallelized architecture.

In this context, a potential direction for future studies is expanding the sample size and including more diverse populations could improve the generalizability of the findings. Future studies applying the proposed pipeline to PD-related datasets involving continuous and long-term monitoring of motor signs, such as the dataset recently introduced by Daneault et al. (2021), may help expand its applicability and reveal additional signal processing steps. In addition, future efforts should focus on exploring alternative window sizes to better inform subsequent analyses and support the identification of an ideal window duration for this type of analysis, while also providing a more comprehensive assessment of the computational cost of the proposed methods. This could lead to the development of new technologies, effectively incorporating machine learning, for better characterization, monitoring, and rehabilitation of Parkinson’s disease, while prioritizing cost-effectiveness and decision reliability.

5 Conclusion

Beyond their role in data collection, serious games hold immense promise for patient care itself. Their interactive and engaging nature can significantly improve patient adherence to rehabilitation programs, transforming what might feel like a chore into a motivating activity. Crucially, these games provide a means for continuous, objective symptom monitoring from the comfort of a patient’s home. This rich, real-world data complements traditional clinical assessments, offering healthcare providers a more complete picture of a patient’s condition over time. Ultimately, integrating serious games into clinical practice represents a significant step towards more personalized, accessible, and patient-centered care for individuals with PD.

Based on all the evidence presented and discussed, it is clear that the choice of signal preprocessing methodology is fundamental to achieving successful classification of biomedical movements. Our results show that using raw signals without any conditioning led to limited performance, as the models were unable to effectively cope with noise and voluntary activity. Similarly, interpolation alone proved insufficient.

The real breakthrough came with the application of wavelet decomposition. This technique not only effectively removed trends and reduced the influence of voluntary movements but also enabled machine learning models to achieve a remarkably high level of accuracy and precision. The combination of interpolation and wavelet decomposition, in particular, proved to be the most robust and complete strategy, leading to near-optimal classification.

Therefore, our study confirms that the quality of the input data is key to model performance. A careful, multi-step signal processing pipeline, such as the combination of interpolation and wavelet decomposition, is not just an optional step but a necessity for extracting meaningful information and ensuring the validity of results in biomedical signal research.

The multi-class classification analysis reinforced the importance of signal preprocessing. While models had difficulty differentiating between healthy individuals, and patients on and off medication using raw or interpolated data, the application of wavelet decomposition significantly improved performance. This demonstrates that sophisticated signal conditioning is not just a beneficial step but a necessity for unlocking the subtle, discriminative patterns needed to classify clinically relevant patient sub-groups.

Finally, the comparative analysis with the ROCKET-RidgeClassifierCV framework underscores the robustness of our methodology. Both catch22 combined with different classifiers and ROCKET-RidgeClassifierCV achieved their best performance with preprocessed signals, where sample losses were addressed (i.e., the interpolation step) and involuntary activity was extracted (i.e., the wavelet decomposition step) before feature extraction and classification. While ROCKET-RidgeClassifierCV achieved a balanced accuracy of 69.0% in the multi-class task, catch22-RF reached 76.0%, suggesting that sophisticated signal conditioning combined with targeted feature extraction can outperform high-dimensional transform methods in distinguishing complex clinical states. These results establish our approach as a viable, high-performance, and computationally efficient alternative for real-time Parkinson’s disease monitoring during serious game interaction.

In conclusion, our study confirms that the choice of data processing strategy is as important as the selection of the machine learning model. A single-step approach is insufficient for handling the complexities of biomedical data. The combination of interpolation to address discontinuities and wavelet decomposition to remove noise and trends emerged as the most effective method, providing a powerful and generalizable framework for analyzing accelerometer signals in the context of neurological conditions.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Ramesh Kandimalla, Indian Institute of Chemical Technology (CSIR), India

Reviewed by: Björn Krüger, University of Bonn, Germany

Mallikarjun Suligavi, BGS MCH, India

Ashwini Pandith N., S R Patil Medical College and Research Centre, India