Context-agnostic machine learning for Parkinson’s disease motor symptom detection using wearable sensors

Abstract

Background

Parkinson’s disease is a rapidly growing neurodegenerative disorder with various motor and non-motor symptoms, affecting millions of people worldwide. These symptoms demonstrate significant medication-related fluctuations and inter-patient variability, highlighting the need for personalized management. Objective longitudinal symptom monitoring through wearable sensors and machine learning can support individualized care. However, to date, most approaches have been tested in lab-constrained environments. This study aims to develop a modular pipeline to automatically detect three cardinal Parkinson’s disease motor symptoms, tremor, bradykinesia, and levodopa-induced dyskinesia in more realistic scenarios.

Methods

The proposed approach was evaluated on three datasets: the Levodopa Response Study and two newly introduced ALAMEDA datasets, containing tri-axial wrist accelerometer data collected with commercial wearable devices during clinical assessments and activities of daily living. For each symptom, separate context-agnostic models were developed using 92 hand-crafted features. Multiple segmentation window lengths and preprocessing techniques, including resampling and dimensionality reduction, alongside various machine learning models, including logistic regression, k-nearest neighbor, multilayer perceptron, support vector machine, decision tree, AdaBoost, and random forest, were explored. Statistical significance between configurations was assessed with the Wilcoxon signed-rank test. Model interpretability was investigated using Shapley additive explanations to identify highly influential predictors and assess their physiological relevance.

Results

In the Levodopa Response Study dataset, tremor, bradykinesia, and dyskinesia detection reached 0.664, 0.636, and 0.443 area under the precision-recall curve, respectively, demonstrating scalability in high-complexity settings and revealing physiologically meaningful patterns. When evaluated on the ALAMEDA datasets, tremor and dyskinesia detection achieved 0.879 and 0.648 area under the precision-recall curve, highlighting strong model and feature generalizability. Across symptoms, longer segmentation windows and random forest classifiers performed better, while synthetic oversampling and principal component analysis showed limited impact.

Conclusions

Automated Parkinson’s disease symptom detection is feasible in more realistic, free-living conditions, with only a slight performance decrease despite substantially increased complexity. With carefully selected features and pipeline components, the objective, unobtrusive monitoring of motor symptoms can support personalized, evidence-based treatment suggestions, eventually improving patients’ quality of life.

Clinical trial number

Not applicable.

Background

Introduction

Parkinson’s disease (PD) is the second most common neurodegenerative disorder after Alzheimer’s disease [1] and one of the most common movement disorders worldwide [2]. Throughout the last decades, it has affected a continuously increasing portion of the global population, especially the elderly [1, 2]. Today, it is estimated that more than 10 million people worldwide suffer from PD [3], compared to 9.4 million in 2020 [4], 8.5 million in 2019 [5], 6.1 million in 2016 and 2.5 in 1990 [6], leading approximately to a 300% increase during the last three decades and marking PD as the fastest expanding neurological disorder [7]. This is also accompanied by a respective increase in disability-adjusted life years (DALYs) and deaths attributed to PD, bearing a considerable socioeconomic burden [6] and applying pressure on national healthcare systems.

PD is a chronic, progressive condition that is pathologically characterized by loss of dopaminergic neurons, abnormal aggregation of alpha-synuclein proteins in Lewy bodies and Lewy neurites in the substantia nigra [2], neuroinflammation and dysfunction of mitochondria, lysosomes, as well as synaptic and vesicle transport problems, resulting in a prominent deficiency of dopamine [8, 9]. Scientists have identified some fundamental factors that contribute to the incidence and prevalence of PD, including aging of the global population, longevity, genetic predisposition, increased industrialization, and environmental factors, such as exposure to specific pollutants, toxins, and pesticides, as well as the decreased smoking rate [7, 9, 10]. However, the exact cause of PD in most diagnosed cases remains unknown [2, 9]. The fact that unlike other diseases, the prevalence rate of PD demonstrates a positive correlation with the socio-demographic index (SDI) and the gross domestic product (GDP) per capita is also particularly concerning [7].

A wide spectrum of both motor and non-motor symptoms is encompassed in PD. Some non-motor manifestations, such as olfactory dysfunction, sleep disorders, depression and constipation, usually precede PD motor clinical features [11, 12]. Additionally, some other non-motor features, including fatigue, pain, visual hallucinations, anxiety, cognitive impairment, dementia, psychosis, bladder and autonomic dysfunction, are likely to appear as the disease progresses [11–13].

However, to date, according to the updated Movement Disorder Society (MDS) diagnostic criteria, PD diagnosis is based primarily on the presence of at least two cardinal motor symptoms, bradykinesia in combination with either rest tremor or rigidity, or both [14]. Bradykinesia corresponds to the decrease in speed or/and amplitude of movements and serves as an umbrella term for hypokinesia and akinesia [14]. On the other hand, tremor is present in up to 80% of PD patients [8] and is divided into two main categories, rest tremor, which occurs at 3–7 Hz when muscles are relaxed, and action tremor, which occurs at 4–12 Hz during voluntary movements and is further subdivided into postural, kinetic and intention tremor, among others [15, 16]. The presence or absence of tremor also classifies a person diagnosed with PD into the tremor-dominant or the akinesia/rigidity-dominant PD phenotype [15]. Other PD motor symptoms include postural instability, gait impairments and freezing, stiffness, micrographia, and speech problems [17].

Although there is no therapy capable of impeding the progression of PD [8], most motor symptoms can be alleviated by medications that stimulate dopamine receptors or increase dopamine concentration, such as levodopa [9]. However, these medications are also associated with adverse effects, such as motor and non-motor fluctuations related to ON and OFF states, and dyskinesias (involuntary movements) that are likely to occur after years of levodopa administration in advanced stages of the disease [1, 18, 19]. These medication-related complications could be mitigated by adjustments in dosing and scheduling of levodopa intakes [8].

Clinical decision-making and prescribing medications are further hindered by the variability of PD symptoms observed among different PD patients [13], highlighting the urgent need for personalized treatments [8]. Today, cardinal PD motor symptoms, such as tremor, bradykinesia, and levodopa-induced dyskinesia, are reported in widely used disease-specific scales, such as the MDS-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) [14]. Similarly, there are other rating scales dedicated to the assessment of both motor and non-motor symptoms and the estimation of disease progression, such as the Scales for Outcomes in Parkinson’s Disease-Motor (SCOPA-Motor), the Schwab and England Activities of Daily Living (ADLs) Scale, and the Unified Dyskinesia Rating Scale (UDysRS), among others [20]. However, all these scales are associated with a certain level of subjectivity [20]. In the case of self-reported outcomes, patients may omit some important manifestations and exaggerate or underestimate their condition, but even when neurologists assess patients’ clinical symptoms, it is very common to observe inter-rater variability to some extent.

Artificial intelligence (AI) and internet of things (IoT)-enabled solutions can support the objective monitoring and assessment of PD symptoms, as well as the estimation of disease progression. Nowadays, there are various wearable sensors and smart devices that unobtrusively collect large amounts of data and transfer them over a network even in near real time [21]. Many types of sensors have been identified in the literature, supporting PD diagnosis, monitoring and management, such as inertial measurement units (IMUs), including accelerometers, gyroscopes and magnetometers, pressure or force sensors, video cameras and depth sensors, microphones, and other portable physiological monitoring devices [22].

While all these sensors provide valuable information, accelerometers are especially relevant for detecting PD motor symptoms, as they provide objective time-series data capturing movement dynamics. Specifically, PD tremor oscillations appear as repeated patterns in the time domain of acceleration signals and as distinct peaks of power at characteristic frequencies in the frequency domain. Notably, accelerometers have been shown to outperform gyroscopes in detecting PD tremors [23]. Bradykinesia is associated with an overall reduction in acceleration magnitude, with signal energy concentrated at lower frequencies due to slower and smaller movements. In contrast, dyskinesia manifests as irregular, excessive movements characterized by greater variability in acceleration amplitude in the time domain and a wider frequency content.

In addition, accelerometers are characterized by a lower power consumption compared to gyroscopes, reducing the need for frequent recharging and making them less burdensome for daily use [24]. Moreover, the use of familiar commercial devices with embedded sensors further enhances patient acceptance, although such devices remain underrepresented in the literature for PD symptom detection tasks [22]. Similarly, the exploitation of a minimum set of sensors for passive monitoring, such as one or two wrist-worn devices, can lead to a higher acceptability rate than more complex and obtrusive body sensor networks.

The collected data can undergo several signal processing techniques and train various machine learning (ML) and deep learning (DL) models, aiming to detect PD symptoms, infer their level of severity or predict the progression of the course of the disease [22]. Indicative common ML methods explored for this purpose span across neural network (NN)-based approaches, ensemble learning techniques, and conventional ML methods, such as support vector machines (SVMs), k-nearest neighbors (kNN), decision trees (DTs), logistic regression (LR), etc. [22].

Leveraging these technological advancements, doctors can gain useful insights into hidden patterns of the course of the disease and quantify PD symptoms and their fluctuations in correlation with prescribed medication intakes. This enables more proactive and personalized treatment adjustments, aligning with the precision medicine paradigm and supporting a truly patient-centered approach in PD care [25]. Moreover, passive, unobtrusive symptom monitoring through wearables offers a less burdensome alternative to traditional or electronic symptom diaries, helping engage more patients in multidisciplinary teams for optimal PD management [8]. Such monitoring solutions are also both time- and cost-effective and can be applied under free-living conditions. However, these AI-based approaches are not yet widely used in clinical practice, as most methods have been tested only in lab environments during specific instructed exercises and the respective datasets are relatively small.

Related work

Several studies in the literature leverage acceleration signals from wearable sensors to detect PD tremor, bradykinesia, and dyskinesia, as well as to estimate their severity using ML and DL methods.

Tremor detection

Firstly, regarding tremor detection, Ibrahim et al. [26] proposed a multilayer perceptron (MLP) classifier trained on features extracted manually from hand-mounted accelerometers during specific upper-limb exercises, obtaining 92.9% accuracy, 98.6% precision, and 86.7% recall. Furthermore, Sun et al. [27] explored various ML models to detect tremor among 30 PD patients, using time- and frequency-domain hand-crafted features extracted from a wrist-worn tri-axial accelerometer and gyroscope during UPDRS and Fahn-Tolosa-Marin scale activities. SVM outperformed the other models with 92.26% F1-score in 5-fold cross-validation.

Rodriguez et al. [28] also employed an SVM to estimate tremor severity on a 3-point scale, derived from MDS-UPDRS tremor ratings of a 24 PD patient cohort. Accelerometer and gyroscope signals were recorded from wrist- and ankle-mounted IMUs, filtered, and subjected to wavelet-based time–frequency analysis to extract various time-, frequency-, and non-linear features. Using a single random train–test split, the proposed model achieved 88% accuracy and 90% sensitivity. Similarly, Wu et al. [29] quantified rest and postural tremor with 95.30% and 91.18% accuracy, respectively, using an NN trained on both temporal and spectral features after applying principal component analysis (PCA).

Yuan et al. [30] used wrist-worn accelerometers to detect rest tremor and quantify its severity based on MDS-UPDRS III item scores. They trained a random forest (RF) classifier on time- and frequency-domain features and applied the synthetic minority oversampling technique (SMOTE), obtaining an average accuracy of 89.70% with leave-one-subject-out (LOSO) cross-validation. Similarly, Sigcha et al. [31] proposed a context classifier to detect the resting phases followed by a multi-task convolutional neural network (CNN) to classify rest tremor labels based on 290 temporal and spectral features extracted from a smartwatch’s tri-axial accelerometer, achieving 86.1% sensitivity, 86.1% specificity, and 0.936 area under the receiver operating characteristic curve (AUROCC).

Finally, in [23], the authors compared accelerometer and gyroscope data from commercial smartwatches for detecting and quantifying rest tremor based on the MDS-UPDRS 3.17 item. After filtering and segmenting the signals, they manually extracted time- and frequency-domain features to train various classifiers, such as LR, kNN, DT, adaptive boosting (AdaBoost), extreme gradient boosting (XGBoost), and RF. Accelerometer data proved more effective, with the RF model outperforming the others, achieving 95% accuracy in rest tremor detection and 96.8% accuracy in rest tremor severity quantification.

Bradykinesia detection

Babayam et al. [32] identified the most informative temporal, spectral, and statistical features for detecting PD bradykinesia using limb-worn sensors. Habets et al. [33] dynamically detected bradykinesia using a single wrist-worn accelerometer placed on the most-affected side. They extracted 103 temporal and spectral features in 60 s windows from 20 PD patients during ADLs. They compared individually with group trained SVM and RF classifiers, achieving similar performance, 0.65–0.70 area under the precision-recall curve (AUPRC), across all cases.

In addition, Habets et al. [34] developed an open-source tool to estimate bradykinesia severity according to MDS-UPDRS 3.4 item scores, using an index-mounted accelerometer. They proposed a two-phase approach: first, detecting tapping activity blocks and single-trial taps, and subsequently, estimating MDS-UPDRS 3.4 item scores using an RF classifier trained on kinematic features. The proposed tool demonstrated a positive correlation with ground-truth data in 5 out of 7 validation subjects.

Dyskinesia detection

Regarding dyskinesia detection, Belgiovine et al. [35] trained an SVM classifier on 168 manually extracted features from acceleration and angular velocity signals, collected with a smartwatch worn on the wrist and hip for upper- and lower-limb assessment, respectively. Across 8 motor tasks, the proposed model achieved 70% accuracy and 70% macro F1-score for upper-limb assessment, and 75% accuracy and 77% macro F1-score for lower-limb assessment.

Additionally, Lee et al. [36] estimated the severity of dyskinesia according to the Rush Dyskinesia Rating Scale (RDRS) during 12 motor tasks based on multiple Shimmer accelerometers mounted on both upper and lower limbs. They employed a cost-sensitive meta-classifier with an RF as base classifier, achieving a root mean squared error (RMSE) ranging between 0.41 and 0.67 with LOSO cross-validation. Similarly, Grigoriou et al. [37] proposed an SVM model to predict Clinical Dyskinesia Rating Scale (CDRS) scores based on a single accelerometer, achieving 0.829 mean Pearson’s correlation coefficient across 25 PD patients.

Moreover, Pfister et al. [38] differentiated dyskinesia, ON and OFF states based on data collected with a wrist-worn accelerometer in unscripted free-living environments. They used a CNN, obtaining 65.4% 3-class balanced accuracy with LOSO cross-validation. Finally, Thomas et al. [39] estimated a treatment response index, where −3 corresponds to the OFF state, 0 corresponds to the ON state, and +3 corresponds to dyskinesia. They extracted various time- and frequency- domain features from tri-axial accelerometer and gyroscope data, applied PCA and finally, trained an SVM model, achieving 0.73 RMSE and 0.82 correlation coefficient with 10-fold cross-validation.

Multiple motor symptom detection

Several studies have also addressed the detection of multiple PD motor symptoms, including bradykinesia, tremor, and dyskinesia. First, Lang et al. [40] proposed a multi-layer Gaussian process to differentiate all these three symptoms and estimate their severity levels. To this end, they exploited gyroscope and accelerometer data from a wrist-worn band, and applied discrete wavelet transform (DWT) to extract several temporal and spectral features, achieving a mean accuracy of 78.50%-88.77% for tremor, 43.38%-67.44% for bradykinesia, and 37.13%-77.48% for dyskinesia.

Similarly, Crowe et al. [41] trained ML models to detect tremor, bradykinesia, and dyskinesia using data collected from a wrist- and an ankle-worn IMU in a supervised lab environment, and then evaluated them in free-living conditions. The results showed balanced accuracies of 83, 75, and 81% in the lab environment, using CNN, gradient boosted trees, and long short-term memory (LSTM) models, respectively, and 63, 63, and 67% in free-living environments.

Antonini et al. [42] also proposed the PDMonitor system, which incorporates multiple IMUs, including wrist-worn devices, for comprehensive monitoring of PD motor symptoms. The system first classifies the user activity and then, based on whether that activity permits reliable symptom assessment, proceeds to tremor, bradykinesia, dyskinesia, and other motor manifestation estimations. In a clinical evaluation involving 65 PD patients, the PDMonitor achieved sensitivities of 0.84, 0.84, and 0.82 for wrist tremor, arm bradykinesia, and dyskinesia, respectively, when compared against clinical scale scores as ground truth.

Complementarily, Gutowski et al. [43] estimated the severity of tremor, bradykinesia, and dyskinesia in PD patients using accelerometer and gyroscope data from the MYO armband and a mobile phone positioned in the hand. The raw inertial signals were filtered, and both temporal and spectral features were extracted from the signal magnitudes. These features were combined with patient- and examination-specific metadata to train SVM, RF, and XGBoost regression models. When evaluated in a cohort of 356 PD patients, the models achieved values of 0.293–0.616 for tremor, 0.094–0.408 for bradykinesia, and 0.084–0.385 for dyskinesia severity estimation.

Additionally, Wagner et al. [44] detected bradykinesia during gross upper-limb movements, dyskinesia during resting and walking, and quantified rest tremor. They applied DWT to wrist acceleration signals and extracted several features in time and frequency domains to train an SVM model, obtaining 0.70 AUROCC for bradykinesia, and 0.75 and 0.92 AUROCC for resting and walking dyskinesia, respectively.

Moreover, Shawen et al. [45] detected both bradykinesia and tremor by training an RF classifier on hand-crafted features extracted from upper-limb accelerometer and gyroscope signals, achieving 0.61–0.69 and 0.68–0.79 AUROCC, respectively. The same problem was addressed in [24], where the authors detected rest tremor and estimated bradykinesia severity based on acceleration signals collected from the wrist of the most affected side during two in-clinic visits. The best hand-crafted features were selected to train an RF classifier for tremor detection, achieving 83% accuracy, 86% precision, 86% recall, and 83% F1-score, and a longitudinal mixed-effects model with Akaike information criterion for bradykinesia quantification, obtaining 2.01 RMSE.

Previous studies on the Levodopa Response study dataset

Finally, several previous studies have used the Levodopa Response Study dataset [46, 47], which is also leveraged in the current work and is described in detail in Sect. 2.1.1, to detect and quantify PD motor symptoms.

First, AlMahadin et al. [48] manually extracted various temporal and spectral features from smartwatch and smart bracelet accelerometer data from the dataset to quantify tremor severity. They proposed RF classifiers for estimating tremor severity during ON and OFF medication states separately, obtaining up to 80 and 77% accuracy, respectively.

Channa et al. [49] estimated rest tremor severity on a 5-point scale based on features extracted in both time and frequency domains from an additional set of Shimmer3 accelerometers worn by a subset of patients [47]. After applying PCA and random oversampling, the proposed XGBoost classifier achieved 99% accuracy, 98%-100% precision, 97%-100% sensitivity, 95%-100% specificity, and 98%-100% F1-score. In a related study [50], they estimated tremor and bradykinesia severity on the most-affected hand, applying continuous wavelet transform (CWT) to raw Shimmer3 acceleration signals and using a CNN-based transfer learning–AlexNet network, achieving 0.36–0.97 sensitivity for tremor quantification and 0.77–0.90 sensitivity for bradykinesia quantification.

In contrast, Wiederhold et al. [51] used the Levodopa Response Study dataset in combination with another public dataset to address exclusively dyskinesia detection based on the MDS-UPDRS 3.6 item. They applied PCA on the raw acceleration signal and explored both hand-crafted and automatically extracted features from time series. Out of all the models tested, the RF classifier achieved the highest performance with 68% F1-score and 88% accuracy with automatically extracted features, and 63% F1-score and 88% accuracy with the hand-crafted features. However, performance dropped when these models were evaluated on another dataset, reaching up to 48–53% F1-score and 54–56% accuracy, demonstrating limited generalization ability.

Moreover, Gutowski [52] trained CNNs to automatically extract features from the raw smartwatch and smart bracelet accelerometer data of the dataset. They combined these features with prior information on the hand each device is worn and on the task performed to detect bradykinesia, dyskinesia, and tremor, and to estimate tremor severity for each task of the 18 in the dataset separately. Performance was evaluated on a single train-test split and varied substantially across tasks: 0.237–0.957 AUPRC and 27.5%-88.4% accuracy for bradykinesia detection, 0.027–0.512 AUPRC and 50.6%-85.6% accuracy for dyskinesia detection, 0.189–0.855 AUPRC and 62.7%-85.5% accuracy for tremor detection, and 0.530–0.998 AUPRC and 34.9%-84% accuracy for tremor severity classification. These results highlight the varying difficulty of detecting specific PD symptoms depending on the task performed.

Khan et al. [53] proposed a two-stage approach: first detecting anomalies (any of the three symptoms), then classifying tremor, bradykinesia, and dyskinesia in the most-affected hand. They compared an RF classifier trained on temporal, spectral, and wavelet features with a CNN-recurrent neural network (RNN) using acceleration magnitude from the smart bracelet. The CNN-RNN achieved stronger peak performance (up to 93.03% recall for anomaly detection and 91.54% recall for symptom classification), while the RF showed more consistent high performance across configurations.

Donie et al. [54] differentiated from other approaches and instead of extracting features manually or automatically via deep neural networks (DNNs), they reframed the problem as a time series classification task. They used InceptionTime and random convolutional kernel transform (ROCKET) models to estimate tremor severity and detect bradykinesia and dyskinesia on the most-affected hand based on smart bracelet acceleration signals. Using InceptionTime, tremor severity was estimated with up to 65.4% accuracy, 46.1% balanced accuracy, 41.2% F1-score, and 0.728 AUROCC, while bradykinesia was detected with up to 75.2% accuracy, 75.5% balanced accuracy, 63.1% F1-score, and 0.841 AUROCC with a single train-test split. Dyskinesia detection with ROCKET reached up to 93.2% accuracy and 0.709 AUROCC, but obtained only 53.1% balanced accuracy and 11.7% F1-score, highlighting the inherent challenge of detecting dyskinesia during ADLs.

Elziaat et al. [55] extended this work [54] by reimplementing their methodology and benchmarking it against a variational autoencoder (VAE), focusing exclusively on bradykinesia detection and using a subset of patients wearing the smart bracelet on the right (most-affected) hand. Results confirmed the superiority of InceptionTime and ROCKET architectures, obtaining 0.66 and 0.70 balanced accuracy, respectively, with a 5-fold cross-validation.

In the end, the PD Digital Biomarker Dialogue for Reverse Engineering Assessment and Methods (DREAM) Challenge [56] which included three sub-challenges corresponding to the prediction of active limb tremor severity, resting limb dyskinesia, and active limb bradykinesia used this dataset. Participants had at their disposal accelerometer data from wrist-worn devices and other meta-data. Although this dataset contains accelerometer recordings from 18 motor tasks, including both clinical assessments and ADLs, only a limited subset of these was used in the challenge: 6 upper-limb tasks for tremor severity estimation, 5 upper-limb tasks for bradykinesia detection, and 2 MDS-UPDRS III upper-limb tasks for dyskinesia detection.

The challenge received several submissions proposing various feature extraction methods, including neural networks, unsupervised learning, and hand-tuned signal processing, such as statistical-, spectral-, and energy-based methods. All submitted feature sets were evaluated with a soft-voting ensemble classifier that combined an RF, an LR, and an SVM classifier. All winning methods were based on conventional signal processing approaches, achieving 0.75 AUPRC for tremor severity estimation, 0.48 AUPRC for dyskinesia detection, and 0.95 AUPRC for bradykinesia detection based on a single train-test split.

Contributions of the current approach

The current study focuses on the automated detection of three principal PD motor symptoms, tremor, bradykinesia, and dyskinesia, among subjects diagnosed with PD, a task less explored in the literature compared to PD diagnosis and classification of PD patients and healthy subjects [22]. The motivation is to support the transition from lab-constrained assessments to real-world, free-living settings, ultimately enabling evidence-based, personalized PD care. This goal encompasses multiple aspects, including the use of patient-friendly, highly acceptable wearable devices for data collection, the generalizability and robust detection of symptoms across diverse tasks and settings, and the interpretability of model predictions.

To ensure real-world applicability, the proposed approach relies on a minimal set of wrist-worn accelerometers embedded in commercial smartwatches and smart bracelets, suitable for daily use. These sensors have proven particularly effective for detecting PD motor symptoms, even outperforming gyroscopes in some cases [23], while being more user-friendly, less burdensome [24], and generally more acceptable to patients.

The proposed method builds upon prior work and extends the preceding contributions to the PD Digital Biomarker DREAM Challenge [56], and specifically its three sub-challenges corresponding to the prediction of active limb tremor severity, resting limb dyskinesia, and active limb bradykinesia, leveraging the Levodopa Response Study dataset [46, 47]. Unlike existing approaches that focus on the most-affected limb [50, 53–55] or specific tasks or require prior knowledge of the activity context [51, 52, 56], the current study addresses a broader and more complex problem. It detects tremor, bradykinesia, and dyskinesia in both upper limbs across all 18 motor tasks in the dataset, without relying on task, limb, medication state or other contextual information as input. To the best of our knowledge, this is the first study to explore context-agnostic PD symptom detection using this dataset, marking an important step toward efficient PD motor symptom detection under free-living conditions.

To achieve this, a separate context-agnostic classifier is developed for each symptom, eliminating the need for prior activity recognition or limb identification. Both instructed exercise from the MDS-UPDRS III protocol and ADLs are processed the same way, regardless of whether data originate from the most- or least-affected, active, or resting upper limb. In the case of tremor detection, this unified approach allows the same classifier to identify both rest and action tremors. By avoiding context-dependent inputs and reducing system complexity, the proposed framework minimizes potential error propagation typical of multi-stage pipelines, enhances robustness, and requires less input data, facilitating its adoption in real-world scenarios.

This is a considerably challenging problem. Symptoms expressed on the least-affected side are often subtler and therefore harder to identify, while certain voluntary movements can obscure or mimic pathological ones. For example, dyskinesia during active motion can overlap with intentional movements [44], while bradykinesia, which is characterized by decreased velocity, cannot be detected in static postures, such as standing or lying [40]. However, addressing these challenges is essential to enable efficient PD motor symptom monitoring under free-living conditions. The present study also explores the trade-off between model performance and complexity. Despite the significant increase in complexity, when benchmarked against the winning solutions of the PD Digital Biomarker DREAM Challenge, the proposed approach resulted in only a slight to moderate reduction in performance, demonstrating high scalability.

Additionally, while the PD Digital Biomarker DREAM Challenge assessed only the impact of different feature sets using a single ensemble classifier, the current study conducts a more extensive evaluation across multiple configurations, including different segmentation window lengths, resampling, and dimensionality reduction techniques, as well as different ML models. To ensure a fair and rigorous comparison, an evaluation strategy, accounting for data dependencies and class imbalance, is employed along with a wide range of carefully selected evaluation metrics. Furthermore, statistical analysis is performed to assess the robustness and reliability of the results.

A comprehensive set of hand-crafted, physiologically meaningful features in both time and frequency domains is extracted. Shapley additive explanations-based (SHAP) feature importance analysis is conducted to identify the most influential predictors for each symptom and to quantify their contribution to model decisions, providing transparent, physiologically grounded insights, supporting the identification of potential digital biomarkers, and ensuring interpretability, unlike automatically extracted features from DNNs, which are often treated as “black boxes” due to their complexity. The proposed solutions are also lightweight and thus, well-suited for deployment in resource-constrained devices.

Finally, the best-performing approaches for tremor and dyskinesia detection are evaluated on two additional datasets, the ALAMEDA PD tremor [57] and dyskinesia [58] datasets, achieving high generalizability for both models and highly influential features. Notably, these datasets are first introduced in the current paper and are publicly available, containing data from multiple clinical assessments.

Article structure

The remainder of this article is structured as follows: Sect. 2 describes the datasets and the methodology adopted, including data preprocessing steps, ML models, the evaluation strategy and metrics, as well as the statistical and feature importance analysis conducted. Section 3 presents the results obtained, providing a performance assessment across different preprocessing techniques, ML models, datasets, and other parameters, as well as insights into the most influential features per symptom. Section 4 discusses the implications of the findings, the limitations of the current study along with some potential future research directions. Finally, Sect. 5 offers the main conclusions drawn, highlighting the importance of the current study.

Methods

Description of datasets

Levodopa Response study dataset

In the current article, we use the Levodopa Response Study dataset [46, 47], which is available from the Michael J. Fox Foundation, to evaluate the proposed methodology. In this study, 31 PD patients were initially enrolled, with 28 PD patients ultimately contributing to the dataset. Participants were recruited from two hospitals, the Spaulding Rehabilitation Hospital in Boston and the Mount Sinai Hospital in New York.

Data were collected throughout a 4-day monitoring period. Only data from the first and fourth day, collected during in-clinic visits, were used in our experiments. These data are accompanied by clinician-provided labels indicating the presence or absence of symptoms and their severity on a 5-point scale. During both visits, patients performed a battery of 18 motor tasks, which contains both MDS-UPDRS III assessment exercises (e.g., finger-to-nose and rapid alternating hand movements) and ADLs (e.g., walking, drawing, typing, folding a towel, etc.), simulating activities in a free-living environment. This battery of motor tasks was repeated 6–8 times and each repetition lasted approximately 20 minutes, followed by a 30-minute break.

On the first day, patients visited the clinic while on their medication, while on the fourth day, patients returned to the clinic in an off-medication state and their medication dose was administered to them after the first repetition of the battery of motor tasks was completed. In both cases, the duration of the entire assessment was long enough to capture symptom fluctuations, as medication effect changed over time.

In our experiments, we exploit only the subset of data that corresponds to the use of a minimum set of sensors for upper-limb motor analysis [46] in accordance with the PD Digital Biomarker DREAM Challenge [56]. This subset contains raw tri-axial accelerometer data collected from both wrists with a smart bracelet (GENEActiv) and a smartwatch (Pebble) with a sampling rate of 50 Hz. The GENEActiv bracelet was worn on the wrist of the most affected side, while the Pebble smartwatch was worn on the wrist of the least affected side. The placement of the devices on the subjects’ bodies during monitoring is shown in Fig. 1.

Fig. 1.

Placement of the wearable sensors used for data collection

Although the Levodopa Response Study dataset includes additional sensors, such as Shimmer3 sensors mounted on the forearms, shanks, and lower back, as well as a smartphone accelerometer placed in a fanny pack at the waist, these are not used in the current study. While such sensors can provide valuable insights during short-term clinical assessments, they are less suitable for long-term, real-world monitoring due to the increased burden they place on patients. Wrist-worn accelerometers embedded in commercial devices, such as bracelets and smartwatches, are more acceptable for daily use and efficiently support upper-limb motor analysis relevant to PD symptom detection. Therefore, they offer a more sustainable and patient-friendly solution, supporting the transition from clinical settings to free-living environments and laying the foundation for personalized PD care.

Moreover, even though the dataset contains demographic information and other metadata from patient clinic visits, these data are not incorporated as supplementary input to train our ML models. While including this information may enhance the models’ performance, the objective of the current approach is to explore the predictive power of sensorial data in detecting PD motor symptoms. Therefore, we focus exclusively on extracting features from the acceleration signals.

ALAMEDA datasets

In this article, we introduce two new datasets that were also used to evaluate the proposed methodology. These datasets were created as part of the ALAMEDA Horizon 2020 Research and Innovation Action [59], which aims to bridge the early diagnosis and treatment gap of brain diseases [60].

Within the PD pilot of the ALAMEDA project, 15 PD patients were initially enrolled in the Special Outpatient Clinic of Parkinson’s Disease and Related Movement Disorders at the First Department of Neurology of the National and Kapodistrian University of Athens, at Eginitio Hospital. The study received approval from the Ethics Committee of the institution, and informed consent was obtained from each participant upon enrollment.

Patients were equipped with a set of wearable devices, including a smartwatch (Fitbit), a smart bracelet (GENEActiv), a pair of smart insoles (Novel Loadsol-ap), and a prototype smart belt, for continuous monitoring throughout one year, coupled with four 1-week intensive monitoring periods scheduled before each in-clinic re-evaluation on a 3-month basis [61]. In the end, 11 PD patients contributed to the final datasets.

In this study, we leverage only the ALAMEDA PD tremor dataset [57] and the ALAMEDA PD dyskinesia dataset [58]. These datasets contain 92 features extracted from raw tri-axial accelerometer data that were collected from wrist-worn GENEActiv bracelets with a sampling rate of 100 Hz during in-clinic visits. These features were generated after applying certain preprocessing steps, following the methodology described below. Moreover, tremor and dyskinesia scores from the MDS-UPDRS III and IV assessments, respectively, conducted during in-clinic visits were used as labels. Both datasets are publicly available through Zenodo. The broader dataset, which includes the GENEActiv raw accelerometer data and additional clinical annotations [62], from which we extracted our features and clinical labels, is also available on Zenodo.

Data processing and analysis pipeline

Separate data processing pipelines were developed for each of the three PD motor symptoms (tremor, bradykinesia, and dyskinesia), each consisting of several preprocessing steps and an independently trained ML model, as illustrated in Fig. 2. In the PD Digital Biomarker DREAM Challenge [56], in the subtasks based on the Levodopa Response Study dataset, conventional signal processing techniques for manual feature extraction outperformed DL architectures, such as CNNs, for automatic feature extraction. So, in the current study we focus on extending the best-performing approaches that correspond to hand-crafted features with signal processing techniques, and we do not evaluate DL methods.

Fig. 2.

Illustration of the proposed data processing and analysis pipeline

Data preprocessing

First, to preserve the correlation between adjacent data points, the missing values on the three axes (X, Y, and Z) were imputed, ensuring minimal data loss and maintaining data integrity. Specifically, the acceleration time series were interpolated within each task separately, using the time method in both forward and backward directions with limit equal to the half window length that was used for segmentation, as it will be explained later. The remaining missing data were dropped. In this way, we tolerated up to 50% consecutive initial missing data per window.

The next step corresponds to the bandpass filtering of the tri-axial interpolated acceleration time series to keep only the frequency bands relevant to each symptom. The sampling rate of the devices employed in the Levodopa Response Study dataset was 50 Hz, which, according to the Nyquist-Shannon sampling theorem [63], allows the detection of movements up to 25 Hz. So, both wrist-worn devices used are suitable for monitoring human movements that range from 0 to 20 Hz, in general [64]. However, only specific frequency components within the monitored range are physiologically relevant to each symptom.

More specifically, PD tremor is characterized by recurrent oscillations in the 3–12 Hz range, manifesting as repeated patterns in the time domain of the acceleration signals and as distinct power peaks at characteristic frequencies in the frequency domain. Rest, postural, and kinetic tremor during slow movement occur at 3–7 Hz, 4–9 Hz, and 7–12 Hz, respectively [15, 65]. Since a single classifier was trained to detect all types of tremor without distinguishing between rest, postural, or kinetic tremor, a unified bandpass filter with cut-off frequencies [2.5 Hz, 12.5 Hz] was applied. In this way, low-frequency sensor drift, gross movements, the gravitational force influence, as well as high-frequency noise, were filtered out to isolate the frequency components that are most representative of tremor activity.

In contrast, bradykinesia and dyskinesia correspond to slower, less periodic movements that are mainly observed in lower frequency bands. Bradykinesia is characterized by reduced magnitude and speed of movement, with the majority of acceleration energy concentrated below 3 Hz, a range shown to effectively capture bradykinesia with wearables [45]. Dyskinesia, on the other hand, manifests as involuntary, irregular, and excessive movements with higher acceleration amplitudes in the time domain but a less rhythmic structure. Dyskinesia generally manifests at lower frequencies than the rhythmic oscillations of tremor and often overlaps spectrally with voluntary movements [41]. So, for both symptoms, high-frequency tremor, as well as the gravitational force influence and low-frequency sensor drift were filtered out with cut-off frequencies [0.25 Hz, 3.5 Hz], to isolate the signal characteristics that are most indicative of bradykinesia and dyskinesia, in line with prior work [24].

In both cases, a fourth-order infinite impulse response (IIR) Butterworth filter was applied. IIR filters are well suited for this study because the primary interest lies in preserving signal amplitude rather than its phase. Additionally, Butterworth filters are maximally flat, as their frequency response demonstrates nearly constant gain in the passband [66].

For each frequency band of the maintained ones, we computed the magnitude and the first principal component (PC1) of the filtered signals, as proposed in [24], to attenuate the dependency on both sensor placement (right or left wrist) and its orientation (lateral or distal, upright or inverted). This was particularly helpful, as the placement of each sensor varied according to the most affected side of each enrolled patient and the symptoms to be detected have no fixed direction, in general. In this way, symptoms can be detected independently for both hands using the same processing pipeline and limb-agnostic models. The last two steps of the proposed signal transformation process are illustrated in Fig. 3.

Fig. 3.

Signal transformation process for tremor, bradykinesia, and dyskinesia detection. (a)-(b) Raw tri-axial acceleration signal after interpolation. (c) Filtered signal for tremor detection using a [2.5 Hz, 12.5 Hz] bandpass filter. (d) Filtered signal for bradykinesia and dyskinesia detection using a [0.25 Hz, 3.5 Hz] bandpass filter. (e) Transformed signal (magnitude and first principal component of the filtered signal) used for feature extraction in tremor detection. (f) Transformed signal (magnitude and first principal component of the filtered signal) used for feature extraction in bradykinesia and dyskinesia detection

Then, the magnitude and the first principal component were segmented into overlapping time windows of various lengths. The window lengths tested span across 2.56 s, 5.12 s, 10.24 s, and 20.48 s. Considering that the sampling rate of the devices employed in the Levodopa Response Study dataset was 50 Hz, these windows correspond to 128, 256, 512, and 1024 samples, respectively. For the ALAMEDA datasets, given that the devices had a sampling rate of 100 Hz, the same window lengths correspond to 256, 512, 1024, and 2048 samples, respectively. The window lengths were intentionally selected as powers of 2 to facilitate the fast Fourier transform (FFT) applied at a later stage. Moreover, all consecutive windows had 50% overlap with each other to ensure that insightful momentary patterns are not missed due to boundary effects.

The selected window lengths belong to intervals commonly encountered in the literature for detecting PD tremor, bradykinesia, and dyskinesia. The high-frequency rhythmic oscillations of tremor can be reliably captured even in short windows (2–5 s), which encompass multiple oscillation cycles, while the slower and less periodic movements related to bradykinesia and dyskinesia may require longer windows (10–20 s) to accumulate sufficient information for their characterization. Testing all four window lengths (2.56, 5.12, 10.24, and 20.48 s) across all three symptom detection tasks enables the exploration of physiologically meaningful temporal scales and the identification of optimal configurations for each condition. However, no longer time windows were explored, as the duration of specific labeled exercises in the datasets is 15 s, and segmentation was performed individually per exercise. Larger time windows would include idle periods or segments from adjacent exercises associated with different labels, reducing the quality of the assigned labels and the relevance of the extracted features.

Next, 92 features were extracted from each window in both time and frequency domains. To extract spectral features, we applied discrete Fourier transform (DFT), as detailed in Eq. 1 using the computationally efficient FFT algorithm.

1

where N is the window length, x(n) the n-th sample of the signal and X(k) the k-th frequency component of the signal.

The extracted features can be grouped into two main categories. The first category includes descriptive statistical and amplitude-based features, such as mean value, mean absolute value, standard deviation, variance, average absolute difference from the mean value, ratio of values above the mean value, median value, deviation from the median value, interquartile range, skewness, kurtosis, minimum and maximum values, peak-to-peak amplitude, and the ratios of positive and negative values, as listed in Table 1. The second category encompasses dynamic and complexity-related features that capture temporal patterns of the signal, as well as spectral features that characterize its frequency domain properties. These include zero crossing ratio, peak ratio, rest period ratio, slope sign change ratio, root mean square value, energy, sample entropy, fluctuation after applying detrended fluctuation analysis, power, dominant frequency, dominant frequency power ratio, power around the dominant frequency, spectral entropy, and spectral flatness, as listed in Table 2.

Table 1.

Statistical and amplitude-based features extracted from acceleration signals for PD motor symptom detection

Feature name	Description	Domain	Signal applied	Mathematical expression
Mean value	Mean value of the signal (1st moment)	Time Frequency	Magnitude PC1
Mean absolute value	Mean value of the absolute values of the signal	Time	PC1
Standard deviation	Standard deviation from the mean value of the signal	Time Frequency	Magnitude PC1
Variance	Variance from the mean value of the signal (2nd moment)	Time Frequency	Magnitude PC1
Average absolute difference from mean value	Average absolute difference of the values of the signal from their mean value	Time Frequency	Magnitude PC1
Above mean value ratio	Ratio of data samples above the mean value of the signal	Time Frequency	Magnitude PC1
Median value	Median value of the signal	Time Frequency	Magnitude PC1	after sorting
Median value deviation	Deviation from the median value of the signal	Time Frequency	Magnitude PC1
Interquartile range	Interquartile range of the signal (the difference between the 75th and the 25th percentile of the data)	Time Frequency	Magnitude PC1	after sorting
Skewness	Measure of the asymmetry of the distribution (3rd moment)	Time Frequency	Magnitude PC1
Kurtosis	Measure of the deviation of the distribution from the normal distribution (4th moment)	Time Frequency	Magnitude PC1
Minimum value	Minimum value of the signal	Time Frequency	Magnitude PC1	\min(x)
Maximum value	Maximum value of the signal	Time Frequency	Magnitude PC1	\max(x)
Peak-to-peak amplitude	Difference between the maximum and the minimum value of the signal	Time Frequency	Magnitude PC1
Positive value ratio	Ratio of the positive data samples of the signal	Time	PC1
Negative value ratio	Ratio of the negative data samples of the signal	Time	PC1

Table 2.

Dynamic, complexity, and spectral features extracted from acceleration signals for PD motor symptom detection

Feature name	Description	Domain	Signal applied	Mathematical expression
Zero crossing ratio	Ratio of zero crossings of the signal	Time	PC1
Peak ratio	Ratio of the peaks of the signal	Time Frequency	Magnitude PC1
Rest period ratio	Ratio of data samples of the signal below a certain threshold (0.2)	Time Frequency	Magnitude PC1
Slope sign change ratio	Ratio of the slope sign changes of the signal	Time Frequency	Magnitude PC1
Root mean square value	Root mean square value of the signal	Time Frequency	Magnitude PC1
Energy	Energy of the signal	Time Frequency	Magnitude PC1
Sample entropy	Sample entropy of the signal	Time	Magnitude PC1	, where A and B are the number of vector pairs with distance < r for length m+1 and m
Fluctuation	Measure of the self-affinity of the signal computed by the detrended fluctuation analysis	Time	Magnitude PC1	, where n the length of segments and the local least-squares linear trend
Total Power	Total power of the signal	Frequency	Magnitude PC1
Dominant frequency	Dominant frequency of the signal	Frequency	Magnitude PC1
Dominant frequency power ratio	Ratio of the power of the dominant frequency of the signal	Frequency	Magnitude PC1
Power around the dominant frequency	Power around the dominant frequency in a range of ± window length/40 data samples	Frequency	Magnitude PC1	, where
Spectral entropy	Entropy of the signal in the frequency domain	Frequency	Magnitude PC1	, where p(k) is the normalized power spectral density
Spectral flatness	Flatness of the signal in the frequency domain	Frequency	Magnitude PC1

These features capture distinct, yet complementary, aspects of PD motor symptoms. Tremor is primarily characterized by high variance-related features, such as standard deviation, root mean square (RMS), and peak-to-peak amplitude, frequent zero crossings and slope sign changes, a dominant frequency peak with high power within the 3–12 Hz range, as well as low sample entropy and high regularity in detrended fluctuation analysis (DFA) that reflect the rhythmic and periodic nature of tremor oscillations. In contrast, bradykinesia manifests as low-amplitude, low-variance, and low-energy motion with sparse zero crossings, long rest periods, reduced mean, deviation, and RMS values, reflecting the slowness and diminished movement amplitude typical of this symptom. Finally, dyskinesia is characterized by large-amplitude fluctuations with high variance, skewness, and kurtosis that indicate high asymmetry and intermittency of the acceleration signal, as well as broad, low-frequency power and high spectral entropy that reflect the unstructured and unpredictable nature of involuntary dyskinetic movements.

Together, these complementary time-, frequency-, and complexity-based features provide a physiologically grounded representation of PD tremor, bradykinesia, and dyskinesia. While individual features capture different aspects of these symptoms, their combination can reveal higher-order relationships and synergistic patterns that better characterize the complex dynamics of PD motor symptoms and enhance the discriminative power of the ML models. All features were extracted manually using signal processing techniques. DNNs for automatic feature extraction were not explored in the current approach, because they were outperformed by conventional signal processing methods in the PD Digital Biomarker DREAM Challenge [56].

Moreover, the datasets used in our experiments are highly imbalanced. In the Levodopa Response Study dataset, tremor is present in 28.4%, bradykinesia in 29.4%, and dyskinesia in only 10.7% of the observations, while in the ALAMEDA datasets, tremor and dyskinesia are present in 24.1% and 29.5% of the observations, respectively. The class distribution was largely unaffected by different window lengths after segmentation, with no consistent pattern of increase or decrease in minority class percentages observed as the window length changed, as illustrated in Fig. 4.

Fig. 4.

Class distributions across symptoms and window lengths, as well as for the original data before segmentation from the Levodopa Response study dataset

To enable the detection of the minority class, which corresponds to the presence of each one of the three symptoms, we employed several oversampling and undersampling techniques. Regarding oversampling, we evaluated both random oversampling (ROS), which randomly resamples observations from the minority class, and SMOTE, which creates new observations of the minority class based on the kNN algorithm. For SMOTE, we experimented with different numbers of nearest neighbors used to generate synthetic samples (). Regarding undersampling, we tested random undersampling (RUS), which randomly removes some samples from the majority class.

Additionally, to ensure that features with greater values do not affect the fitting more than features with smaller range of values, features extracted underwent standardization based on the following mathematical expression:

2

where is the original feature value at index i, μ is the mean, σ is the standard deviation and is the standardized feature value at index i. This also enabled the subsequent feature importance analysis that we performed for all three symptoms to identify the features that contribute the most to the prediction outcomes.

The final preprocessing step corresponds to dimensionality reduction. We set a variance threshold to ensure that the features are informative enough to be maintained in the training process. Additionally, we experimented with two more methods. The first one corresponds to the selection of the k best features based on the analysis of variance (ANOVA) filter which computes the F-value () for each feature (f) according to the following mathematical formula:

3

where n is the total number of instances in the dataset, is the number of instances in class j, is the mean of feature f across both classes, is the mean of feature f for class j and is the value of the feature f for the i-th instance in class j.

The second method of dimensionality reduction corresponds to the projection of the original data onto the subspace created by the first k more informative principal components. In this case, the new reduced dataset () is produced as follows:

4

where is the centered original dataset and is the matrix of the k eigenvectors that correspond to the k greater eigenvalues of the covariance matrix of the centered dataset which is defined by the following mathematical expression:

5

In both cases, we experimented with different numbers of selected features or principal components (k), starting from 10 and increasing in increments of 5, up to the full set of features remaining after removing completely uninformative features using a variance threshold of 0.

Machine learning models

After completing the preprocessing, we used the selected transformed features to train various ML models. We experimented with a wide range of models to identify those that perform best for the specific problems we address. The ML methods tested in the current study include LR, kNN, MLP, SVM, DT, AdaBoost, and RF.

An LR model estimates the probability of a PD motor symptom occurring using the logistic function which is written as:

6

where is the weight vector, is the feature vector and b is the bias term [67]. In our experiments, we used the liblinear solver to train the LR model and L2 regularization, exploring various values of the inverse regularization parameter (). We also tested the balanced option for the class weight hyperparameter to account for class imbalance. The training process was set to terminate upon convergence or after reaching the maximum number of iterations, which was set conservatively to 5000.

A kNN model is a non-parametric model which predicts the occurrence of a PD motor symptom by assigning the class of the k nearest instances in the feature space, as Eq. 7 indicates:

7

where is the class of the i-th neighbor in the set of the k nearest neighbors [68]. In our experiments, we tested various values of k larger than typically used due to the dataset size and the class imbalance (). We evaluated both uniform and distance-based weighting methods and experimented with Manhattan, Euclidean, and Minkowski distance metrics. The underlying search algorithm was automatically selected by the classifier among Ball Tree, KD Tree, and brute-force search.

An MLP model is based on a feedforward neural network with neurons organized into multiple layers and non-linear activation functions [69]. In our experiments, we used a rectified linear unit (ReLU) activation function, defined as follows:

8

We also used the adaptive moment estimation (ADAM) optimizer to train the MLP classifier with batch size set to auto. We experimented with different numbers of layers and layer sizes, including the following configurations: [(10,), (50,), (100,), (200,), (100, 50)]. The training process was set to stop upon convergence, which is defined as no improvement in the loss function by at least 0.0001 for 10 consecutive iterations, or when a maximum of 5000 iterations is reached.

An SVM model determines the presence or absence of a PD motor symptom based on a hyperplane that maximizes the margin between the two classes [70]. In our experiments, a radial basis function (RBF) kernel was used, expressed as follows:

9

to transform the data into a higher-dimensional space, enabling linear separation. We explored different values of the regularization parameter () and tested the balanced option for the class weight hyperparameter to address class imbalance. The training process was set to terminate upon convergence, when satisfying the Karush-Kuhn-Tucker conditions within a tolerance of 0.001.

A DT model predicts the occurrence of a PD motor symptom by assigning each instance to a leaf node that corresponds to a specific class, as determined by feature-based splits at the intermediate nodes. During training, in our experiments, the feature splits at the intermediate nodes were chosen to minimize the Gini impurity (G), defined as:

10

where is the proportion of instances in class j [71]. To address class imbalance, we also tried the balanced option for the class weight hyperparameter. The training process was terminated when the tree reached the specified maximum depth, for which we explored the following range of values or when no further splits were possible due to the requirement of having at least a specific number of samples in each leaf, for which we tested the following values .

An AdaBoost model adopts an ensemble boosting approach, according to which multiple decision trees are trained sequentially, with each tree focusing on misclassified instances from the previous individual models to progressively improve performance. In our study, we experimented with a range of shallow DTs, each with a maximum depth in the range of . In the end, the occurrence of a PD motor symptom is predicted based on the weighted sum of the individual classifiers as follows:

11

where is the class predicted by the t-th decision tree, is the weight of this prediction and is dependent on the error of the t-th decision tree (), calculated as , and T is the number of decision trees [72].

Finally, an RF model implements an ensemble bagging approach, predicting the occurrence of a PD motor symptom based on the majority vote of the predicted outcomes of multiple decision trees, as follows:

12

where is the class predicted by the t-th decision tree and T is the number of decision trees [73]. In our study, we experimented with different numbers of individual DTs (). Each tree was trained on a bootstrap sample of the training data to enhance model variance and had a maximum depth in the range and a minimum number of leaf samples in the range . To address class imbalance, we also evaluated the use of the balanced subsample option for the class weight hyperparameter, which adjusts weights based only on the bootstrap sample drawn for each tree.

To ensure the optimal performance of the proposed pipeline, we systematically tuned the hyperparameters of all ML models and preprocessing steps described above. This was performed with a grid search strategy combined with stratified group 5-fold cross-validation optimizing for the F1-score to account for class imbalance. The search space for each model and preprocessing method was defined based on domain knowledge, literature recommendations, dataset characteristics, and computational constraints. The specific hyperparameter ranges evaluated for each model and preprocessing step are detailed above. Eventually, the best hyperparameter values identified were used to refit the pipeline on the entire training set.

Performance evaluation methodology

Evaluation strategy

The predictive capability of the proposed models was evaluated based on stratified group 5-fold cross-validation. In general, cross-validation is more robust than a single train-test split, as it averages performance achieved across multiple splits. In addition, 5-fold cross-validation is less computationally expensive than LOSO cross-validation and tends to produce more stable and reliable performance estimates with lower variance across folds in small, imbalanced datasets such as the ones used in this study.

Moreover, we employed a group-wise cross-validation strategy, in which each group consisted of all the observations of one or more patients in the datasets. This was essential, as the independent and identically distributed (i.i.d.) assumption was not satisfied in our case due to the presence of groups of dependent samples in the datasets. By placing each subject in a different testing fold and ensuring that every subject was never in both testing and training sets, we eventually prevented data leakage. In addition, with this design, since missing values were interpolated independently for each patient, no information from the test partitions could influence the interpolation of the training data.

Furthermore, given the high class imbalance of the datasets, with some patients even never manifesting specific symptoms during the monitoring period, we applied the stratification method. As a result, the class distribution was preserved in each fold, preventing potential misleading results due to skewed splits. Additionally, when resampling techniques were used to address class imbalance, they were applied exclusively to the training partitions, ensuring that no test data were resampled or used to influence the generation of new samples in the training set.

Finally, nested cross-validation was used in our experiments, where the outer loop estimated the generalization error of the models, while the inner loop was used to tune the hyperparameter values, as illustrated in Fig. 5.

Fig. 5.

Illustration of nested 5-fold cross-validation, where the outer loop evaluates model performance and the inner loop tunes hyperparameters within each fold

Evaluation metrics

In addition to the evaluation strategy, careful consideration was given to the selection of suitable evaluation metrics. In the current study, we employed various evaluation metrics to quantify the performance and generalization ability of the proposed methods. These metrics include: accuracy, precision, recall, F1-score, their macro and weighted averages, as well as the area under the receiver operating characteristic curve, the area under the precision-recall curve, Mathew’s correlation coefficient, and Cohen’s kappa score.

Firstly, accuracy (ACC), one of the most widely used performance indicators in classification problems, is defined as the ratio of the overall correctly classified instances to the total instances:

13

where TP stands for true positives, FP for false positives, TN for true negatives, and FN for false negatives.

Precision (PREC) corresponds to the proportion of correctly predicted positive instances out of all the instances predicted as positive:

14

Recall (REC) corresponds to the proportion of correctly predicted positive instances out of all the actual positive instances:

15

The F1-score (F1) is the harmonic mean of precision and recall:

16

The macro averages of precision, recall, and F1-score correspond to the mean of these metrics across classes, giving equal importance to both majority and minority classes, while the weighted averages compute their weighted mean according to the support of each class, as shown in the following formulas:

17

18

where corresponds to precision, recall or F1-score for class i, is the number of instances within the class i and n is the total number of instances.

The values of all the above metrics range from 0 to 100%. Values above 50% indicate that a model performs better than random chance in detecting the presence or absence of a PD motor symptom, with higher values corresponding to better performance.

However, in highly imbalanced datasets, like the ones used in the current study, accuracy may be overly high due to the successful detection of the majority class even if the minority class is poorly detected. Similarly, precision, recall, and F1-score depend on whether the majority class is considered the positive one or not, thus, they may also be misleading. While the weighted averages of these metrics offer a better estimation of the overall performance of a model, they remain biased toward the majority class, whereas macro averages treat all classes equally and, therefore, provide a more balanced, though not entirely unbiased, view of model performance in imbalanced datasets.

In our case, detecting the minority class is crucial, as it corresponds to the presence of a PD motor symptom. So, the need for additional evaluation metrics is evident, as it was highlighted in [49]. This issue was partially addressed in the PD Digital Biomarker DREAM Challenge [56] by using the AUPRC to evaluate the feature sets submitted. In the current study, we employed both AUROCC and AUPRC which typically constitute better options for imbalanced datasets.

The ROC curve plots the true positive rate (TPR) against the false positive rate (FPR), while the precision-recall curve plots precision against recall across various classification thresholds. The areas computed under these curves, AUROCC and AUPRC, respectively, provide a measure of the trade-off between the respective metrics, and their values range from 0 to 1. Values above 0.5 indicate that a model performs better than random chance in detecting the presence or absence of a PD motor symptom, with higher values corresponding to better performance.

During the evaluation, we also used the Mathew’s correlation coefficient (MCC) which provides a balanced measure of performance even in imbalanced datasets and measures the agreement between true and predicted labels, considering all classes according to the following formula:

19

Finally, Cohen’s kappa (CK) score was also used in our experiments, as it has proven to be more reliable in medical applications with imbalanced datasets [74]. It measures the agreement between true and predicted labels, accounting for the probability of this agreement occurring by chance, as expressed by the following equation:

20

where is the observed agreement and is the expected agreement by chance.

The values of the last two metrics range from −1 to 1, with positive values indicating the presence of agreement and, thus, the ability of the model to detect the presence or absence of a PD motor symptom better than random chance. Again, the higher these values are, the more accurate a model is.

For all evaluation metrics employed, their mean values along with the respective standard deviations across the 5 folds of the cross-validation are reported in the Results section, to assess model stability. Moreover, calibration curves were generated, depicting the relationship between the predicted probabilities and the observed fractions of positive samples across probability bins to assess the reliability of the model probabilistic predictions.

Statistical analysis and testing

To assess whether performance differences across various configurations, such as segmentation window lengths, resampling methods, dimensionality reduction techniques, and ML models, are statistically significant, we employed the Wilcoxon signed-rank test [75], a robust non-parametric paired statistical test, widely used for ML model comparisons in binary classification problems [76].

In our experiments, this test was applied to the cross-validation fold results obtained for each of the evaluation metrics described in Sect. 2.2.2. Given two paired performance values from two different configurations, the Wilcoxon test evaluates whether their median difference is zero, which under the null hypothesis () indicates equivalent performance. To verify whether the best performing configurations achieved significantly higher performance compared to the others, a one-sided alternative hypothesis () was used.

21

where is the performance difference between two configurations on the i-th fold.

Zero differences are removed and absolute differences are ranked in ascending order. The Wilcoxon test statistic is then defined as the sum of ranks () corresponding to positive differences:

22

while the p-value is defined as the probability, under the null hypothesis, of obtaining a test statistic at least as large as the observed statistic:

23

for 5 folds and possible sign assignments. A p-value < 0.05 was considered statistically significant. With 5 folds, this occurs only when the best performing configurations outperform the others in every fold, so that all differences are positive, resulting in with .

Model interpretability and feature importance analysis

To interpret the best-performing models and identify the most influential predictors for each symptom detection task, Shapley additive explanations (SHAP) [77] were employed. The SHAP framework quantifies the contribution of each feature to the model prediction by assigning it a unique importance value, known as a SHAP value. It is inspired by cooperative game theory and the concept of Shapley values. In this context, predicting the class of each observation is viewed as a game, with the features acting as players, and the Shapley values fairly distributing the model prediction among the features. The Shapley value for feature i and model g is defined as:

24

where S is a subset of the feature set excluding feature i and M the total number of features in the dataset. This expression represents the average marginal contribution of feature i across all possible feature combinations.

Global feature importance was visualized using SHAP summary (beeswarm) plots, which rank predictors according to their mean absolute SHAP values:

25

where the feature i, the observation j and N the total number of observations in the dataset. In these plots, each point represents the SHAP value of a feature for a given observation, indicating how that feature influences the model prediction relative to the baseline. Positive SHAP values push predictions toward symptom presence, while negative values push them toward symptom absence. The color scale reflects the actual feature value, demonstrating whether higher or lower feature values are associated with increased predicted symptom probability. The horizontal spread of SHAP values for each feature reflects the variability of its impact across observations, where a wider spread indicates a stronger overall influence on the model predictions.

Results

In this section, we present the results of the experiments conducted, following the methodology outlined earlier. We investigate the impact of various factors on the performance of the proposed approaches for detecting three Parkinson’s disease motor symptoms (tremor, bradykinesia, and dyskinesia), using the Levodopa Response Study dataset. These factors include the window length used for segmentation, resampling techniques, dimensionality reduction methods, and ML models. We also perform a feature importance analysis to identify the features that contributed most to the prediction outcomes. Then, we benchmark performance against the winning solutions of the PD Digital Biomarker DREAM Challenge [56] for tremor, bradykinesia, and dyskinesia detection, using the same dataset and AUPRC as a common evaluation metric. Finally, we present the evaluation results of the proposed approaches on the ALAMEDA datasets.

Impact of window length on performance

As described in Sect. 2, we experimented with four different window lengths for segmentation: 2.56 s, 5.12 s, 10.24 s, and 20.48 s that correspond to 128, 256, 512, and 1024 samples for the Levodopa Response Study dataset. Table 3 presents the performance of simple LR models in detecting tremor, bradykinesia, and dyskinesia across the explored window lengths, without any resampling or dimensionality reduction technique applied.

Table 3.

Impact of window length on the performance of LR classifiers for tremor, bradykinesia, and dyskinesia detection, using the Levodopa Response study dataset, without applying any resampling or dimensionality reduction techniques

Window length	ACC	PREC	REC	F1	AUROCC	AUPRC	MCC	CK
Tremor
128	65.8%	70.2%	65.8%	67.2%	0.637	0.564	0.251	0.243
256	67.9%	71.6%	67.9%	69.1%	0.656	0.583	0.289	0.280
512	68.2%	73.2%	68.2%	69.5%	0.665	0.588	0.305	0.291
1024	69.8%	72.6%	69.8%	70.6%	0.694	0.664	0.371	0.363
Bradykinesia
128	64.9%	70.4%	64.9%	66.3%	0.648	0.595	0.273	0.261
256	66.9%	71.6%	66.9%	68.1%	0.663	0.607	0.304	0.292
512	67.1%	71.0%	67.1%	68.1%	0.658	0.603	0.296	0.287
1024	68.1%	72.4%	68.1%	69.2%	0.678	0.624	0.330	0.318
Dyskinesia
128	61.1%	85.1%	61.1%	68.8%	0.588	0.373	0.112	0.078
256	62.2%	85.5%	62.2%	69.8%	0.610	0.389	0.132	0.094
512	63.2%	86.7%	63.2%	70.7%	0.627	0.402	0.149	0.102
1024	70.1%	86.1%	70.1%	75.5%	0.668	0.442	0.227	0.185

In general, the following pattern was observed: longer windows led to better performance across all three PD symptom detection tasks. This improvement was statistically significant for both tremor and dyskinesia detection according to the Wilcoxon signed-rank test, where 1024-sample windows outperformed shorter ones with T = 15 and p = 0.0312 for AUPRC and accuracy. Although this trend was also observed for bradykinesia, it did not reach statistical significance.

Importantly, class distribution remained largely unaffected by window length, as shown in Fig. 4, suggesting that the performance gains are not due to improved class balance. Only a slight increase in the proportion of tremor-positive instances was observed when using the longest window (20.48 s) for segmentation, which may have modestly contributed to the corresponding improvement in performance.

From a physiological perspective, this trend is consistent with the temporal characteristics of PD motor symptoms. In particular, bradykinesia and dyskinesia, which manifest as slow, reduced movements and irregular, excessive motions, respectively, may not be fully captured by shorter windows. Longer windows provide a richer representation of these symptoms by accumulating more temporal information from the acceleration signals, potentially enhancing their detectability. In contrast, while high-frequency tremor oscillations can be theoretically effectively captured even in shorter windows, longer windows capture more oscillation cycles, potentially improving the robustness, stability, and confidence of the ML models in practice.

Additionally, this consistent improvement in performance may also reflect labeling consistency. Longer windows capture a larger portion of the exercises performed, potentially leading to more accurate labeling of the observations in the dataset, which in turn may lead to better model generalization. The issue of mislabeled observations due to segmentation was also highlighted in [36], where the authors proposed an expectation-maximization clustering preprocessing step to address this problem. However, this approach was not adopted in the current study to prevent the risk of propagating labeling errors.

Based on these findings, the longest window (20.48 s) was used for segmentation in the subsequent analyses. Although further increasing the window length might seem advantageous for capturing more contextual information, this would also reduce the reliability of the labels and the relevance of the extracted features, as longer windows would include idle or irrelevant periods beyond the completion of each exercise, as explained in Sect. 2.

Impact of resampling on performance

Three different resampling techniques (RUS, ROS, and SMOTE) were evaluated in our experiments. Table 4 presents their impact on the performance of simple LR classifiers in detecting tremor, bradykinesia, and dyskinesia, using 1024-sample windows and no dimensionality reduction. SMOTE consistently demonstrated a slight improvement in classification performance, outperforming ROS in most cases, while RUS resulted in slightly decreased performance across all three cases of PD motor symptom detection. However, these differences were not statistically significant (p > 0.05).

Table 4.

Impact of different resampling techniques on the performance of LR classifiers for tremor, bradykinesia, and dyskinesia detection, using the Levodopa Response study dataset, 1024-sample windows and no dimensionality reduction

Re-sampling method	ACC	PREC	REC	F1	AUROCC	AUPRC	MCC	CK
Tremor
None	69.8%	72.6%	69.8%	70.6%	0.694	0.664	0.371	0.363
RUS	69.4%	72.1%	69.4%	70.1%	0.688	0.657	0.359	0.352
ROS	69.7%	72.3%	69.7%	70.5%	0.691	0.660	0.366	0.359
SMOTE	70.0%	72.5%	70.0%	70.7%	0.693	0.663	0.370	0.363
Bradykinesia
None	68.1%	72.4%	68.1%	69.2%	0.678	0.624	0.330	0.318
RUS	67.5%	72.3%	67.5%	68.7%	0.677	0.624	0.326	0.312
ROS	67.7%	72.2%	67.7%	68.9%	0.675	0.622	0.325	0.312
SMOTE	68.4%	72.6%	68.4%	69.5%	0.680	0.626	0.335	0.324
Dyskinesia
None	70.1%	86.1%	70.1%	75.5%	0.668	0.442	0.227	0.185
RUS	69.0%	86.0%	69.0%	74.7%	0.665	0.442	0.221	0.177
ROS	70.3%	86.1%	70.3%	75.7%	0.670	0.443	0.230	0.187
SMOTE	70.8%	86.0%	70.8%	76.0%	0.667	0.438	0.228	0.188

More specifically, regarding tremor detection, only SMOTE offered a marginal improvement in accuracy, weighted average recall and F1-score, and Cohen’s kappa score. Regarding bradykinesia detection, SMOTE demonstrated a more clear improvement across all evaluation metrics, though still without statistical significance. Finally, regarding dyskinesia detection, both oversampling techniques (ROS and SMOTE) marginally improved classification results, with SMOTE demonstrating higher values in accuracy, weighted average recall and F1-score, and Cohen’s kappa score.

These findings can be explained by the underlying principles of each resampling method. First, RUS addresses class imbalance by discarding a large portion of majority-class samples, which may result in significant information loss, especially in relatively small datasets, like the Levodopa Response Study dataset, and instead of improving performance, deteriorate it. On the other hand, ROS mitigates class imbalance by replicating minority-class samples, which may lead to model overfitting due to overly redundant information. Therefore, it is reasonable that these two resampling methods did not yield improved performance in most cases. SMOTE, in contrast, generates synthetic minority-class samples, retaining overall data variability and avoiding information loss. This methodological advantage could explain the consistent slight improvement in performance achieved with this method across all three symptoms.

However, the overall limited performance gain from all resampling methods may be linked to both inter- and intra-class distribution in the feature space. When classes are already well separable, class imbalance plays a minor role, and resampling adds little benefit, since the model can already distinguish the classes effectively. Conversely, when classes are poorly separable, which is the most realistic scenario in complex physiological data, the high intra-class variability of PD symptoms may lead to noisy or unrepresentative resampling results. For example, SMOTE may generate synthetic samples that do not correspond to realistic physiological patterns, thereby limiting the impact of resampling on model performance.

Impact of dimensionality reduction on performance

Two different dimensionality reduction techniques, selecting the k best features based on the ANOVA filter and transforming the features based on the first k principal components (PCA), were explored. Table 5 presents their impact on the performance of simple LR classifiers in detecting tremor, bradykinesia, and dyskinesia, applying 1024-sample windows for segmentation and SMOTE. PCA led to marginally improved classification performance in certain evaluation metrics for tremor and bradykinesia detection, though without statistical significance (p > 0.05), while the ANOVA filter resulted in slightly reduced performance across all three PD motor symptom detection tasks.

Table 5.

Impact of different dimensionality reduction techniques on the performance of LR classifiers for tremor, bradykinesia, and dyskinesia detection, using the Levodopa Response study dataset, 1024-sample windows and SMOTE

Dimension reduction method	ACC	PREC	REC	F1	AUROCC	AUPRC	MCC	CK
Tremor
None	70.0%	72.5%	70.0%	70.7%	0.693	0.663	0.370	0.363
ANOVA	69.7%	72.4%	69.7%	70.3%	0.691	0.661	0.365	0.358
PCA	70.0%	72.6%	70.0%	70.7%	0.694	0.664	0.372	0.365
Bradykinesia
None	68.4%	72.6%	68.4%	69.5%	0.680	0.626	0.335	0.324
ANOVA	66.9%	71.7%	66.9%	68.1%	0.669	0.618	0.313	0.301
PCA	68.5%	72.6%	68.5%	69.6%	0.680	0.626	0.335	0.323
Dyskinesia
None	70.8%	86.0%	70.8%	76.0%	0.667	0.438	0.228	0.188
ANOVA	70.1%	85.9%	70.1%	75.5%	0.662	0.435	0.221	0.180
PCA	70.2%	85.9%	70.2%	75.6%	0.663	0.435	0.221	0.181

The reduced effectiveness of ANOVA-based selection, which ranks features by their univariate association with class labels, suggests that even low-variance or weakly correlated features contribute useful complementary information, either individually or in combination with other features. In contrast, PCA achieved comparable or slightly higher performance with fewer dimensions by decorrelating features and capturing the dominant variance directions, mitigating the risk of overfitting. However, this approach does not ensure the preservation of symptom-relevant physiological variance, since the most discriminative features do not necessarily correspond to the principal components of highest variance.

Overall, the limited impact of both dimensionality reduction techniques indicates that the original feature set already encodes physiologically relevant and largely non-redundant information, capturing distinct complementary temporal, spectral, and complexity properties of PD motor symptoms. After standardization, the LR classifiers effectively handled the moderate feature dimensionality, highlighting the quality and discriminative power of the handcrafted feature engineering performed. Therefore, further dimensionality reduction is unlikely to yield substantial gains without compromising interpretability.

Impact of different ML methods on performance

As described in Sect. 2, seven different ML models were evaluated in our experiments: LR, kNN, MLP, SVM, DT, AdaBoost, and RF. Table 6 presents the performance of each model in detecting tremor, bradykinesia, and dyskinesia. In all cases, 1024-sample windows were used, as they demonstrated the best classification performance. Additionally, given the marginal improvements observed with SMOTE and PCA in some cases, each model was tested with and without these preprocessing methods.

Table 6.

Performance of different ML algorithms for tremor, bradykinesia, and dyskinesia detection, using the Levodopa Response study dataset and 1024-sample windows

Model	ACC	PREC	REC	F1	AUROCC	AUPRC	MCC	CK
Tremor
LR	70.0%	72.6%	70.0%	70.7%	0.694	0.664	0.372	0.365
kNN	63.0%	70.1%	63.0%	64.1%	0.656	0.646	0.296	0.273
MLP	70.0%	71.9%	70.0%	70.6%	0.687	0.654	0.359	0.355
SVM	70.5%	72.2%	70.5%	71.0%	0.689	0.656	0.366	0.362
DT	65.8%	67.4%	65.8%	66.3%	0.634	0.599	0.260	0.257
AdaBoost	67.5%	70.6%	67.5%	68.3%	0.670	0.640	0.324	0.316
RF	71.9%	72.5%	71.9%	72.0%	0.689	0.655	0.375	0.373
Bradykinesia
LR	68.5%	72.6%	68.5%	69.6%	0.680	0.626	0.335	0.323
kNN	64.7%	72.7%	64.7%	66.1%	0.675	0.636	0.320	0.294
MLP	68.9%	71.3%	68.9%	69.7%	0.663	0.604	0.311	0.306
SVM	67.6%	71.8%	67.6%	68.7%	0.672	0.619	0.319	0.306
DT	67.5%	70.5%	67.5%	68.5%	0.655	0.597	0.292	0.286
AdaBoost	68.8%	71.5%	68.8%	69.7%	0.667	0.608	0.315	0.310
RF	71.7%	73.0%	71.7%	72.1%	0.685	0.626	0.356	0.353
Dyskinesia
LR	70.8%	86.0%	70.8%	76.0%	0.667	0.438	0.228	0.188
kNN	68.0%	85.9%	68.0%	73.9%	0.663	0.443	0.216	0.169
MLP	73.1%	85.3%	73.1%	77.6%	0.653	0.413	0.216	0.187
SVM	76.3%	86.2%	76.3%	80.1%	0.671	0.427	0.252	0.226
DT	72.1%	84.5%	72.1%	76.7%	0.630	0.386	0.181	0.155
AdaBoost	69.7%	84.8%	69.7%	75.0%	0.628	0.395	0.179	0.150
RF	83.4%	86.1%	83.4%	84.5%	0.647	0.402	0.270	0.264

Regarding tremor detection, RF and LR outperformed the other ML models. The RF classifier achieved accuracy, weighted average precision, weighted average recall, weighted average F1-score, 0.689 ± 0.030 AUROCC, 0.655 ± 0.057 AUPRC, 0.375 ± 0.064 MCC, and 0.373 ± 0.066 CK. On the other hand, the LR classifier achieved accuracy, weighted average precision, weighted average recall, weighted average F1-score, 0.694 ± 0.033 AUROCC, 0.664 ± 0.047 AUPRC, 0.372 ± 0.067 MCC, and 0.373 ± 0.063 CK. In both cases, the macro average scores were slightly lower than the weighted averages, with macro average recall and macro average F1-score for RF, and macro average recall and macro average F1-score for LR. However, the minority class (tremor) was successfully detected with recall of and , respectively. Both models significantly outperformed kNN, DT, and AdaBoost in terms of accuracy and AUPRC according to the Wilcoxon signed-rank test (T = 15, p = 0.0312), but differences with MLP and SVM were not statistically significant.

Regarding bradykinesia detection, the RF classifier achieved the highest scores in most evaluation metrics among the ML methods tested. More specifically, it yielded accuracy, weighted average precision, weighted average recall, weighted average F1-score, 0.685 ± 0.069 AUROCC, 0.626 ± 0.062 AUPRC, 0.356 ± 0.130 MCC, and 0.353 ± 0.129 CK. Macro average scores were slightly lower, with macro average recall and macro average F1-score, but the minority class (bradykinesia) was successfully detected with recall. Notably, RF significantly outperformed kNN, SVM, and DT in terms of accuracy according to the Wilcoxon signed-rank test (T = 15, p = 0.0312). However, the kNN classifier achieved the highest AUPRC, reaching 0.636 ± 0.036, and detected bradykinesia with recall, but with a lower F1-score of .

Regarding dyskinesia detection, the RF classifier demonstrated the highest scores across most evaluation metrics, achieving accuracy, weighted average precision, weighted average recall, weighted average F1-score, 0.647 ± 0.035 AUROCC, 0.402 ± 0.088 AUPRC, 0.270 ± 0.083 MCC, and 0.264 ± 0.086 CK, significantly outperforming all other models in terms of accuracy (T = 15, p = 0.0312). Macro average scores were lower, with macro average recall and macro average F1-score, and the minority class (dyskinesia) was marginally detected with recall. However, the SVM classifier outperformed the RF model in terms of weighted average precision, AUROCC, and AUPRC, with scores of , 0.671 ± 0.053, and 0.427 ± 0.090, respectively, detecting dyskinesia more successfully with recall, though these differences were not significant (p > 0.05). Notably, the SVM model significantly outperformed all other classifiers except RF in terms of accuracy (T = 15, p = 0.0312). Lastly, again, the kNN classifier achieved the highest AUPRC, reaching 0.443 ± 0.072 and detected dyskinesia with a higher recall of but with a substantially lower F1-score of .

Overall, RF consistently outperformed the other models across tremor, bradykinesia, and dyskinesia detection tasks. This can be attributed to its ensemble nature, which enhances generalization even in relatively small, imbalanced datasets, like the Levodopa Response Study dataset. By aggregating multiple weaker classifiers, RF is less prone to overfitting and noise, and is able to capture non-linear, complex relationships among heterogeneous time-, frequency-, and complexity-based features. Additionally, LR achieved competitive results for tremor detection, suggesting that several tremor-related amplitude and frequency features may manifest in a relatively linear fashion with tremor intensity. In contrast, the strong performance of the SVM model with the RBF kernel for dyskinesia detection may reflect the inherently more complex, non-linear movement patterns of this symptom. Lastly, while kNN yielded higher recall and AUPRC for bradykinesia and dyskinesia detection, its overall lower performance in other metrics likely indicates overfitting, typical behavior of kNN classifiers in high-dimensional, imbalanced datasets, as they rely solely on local distances, and thus, may overemphasize small, noisy minority-class clusters and misclassify borderline test data.

Notably, for all three symptoms, as shown in Fig. 6, the RF models effectively rank predicted probabilities, with higher probabilities consistently corresponding to subsets with more positive samples, indicating a strong discrimination between positive and negative observations. The tremor and bradykinesia detection models are well calibrated, with predicted probabilities closely reflecting observed fractions of positives and only slight overconfidence in some cases, highlighting the potential that they hold for clinical decision support. On the other hand, the model that detects dyskinesia, the rarest symptom in the dataset, exhibits greater overconfidence for medium to high predicted probabilities. This behavior is typical in highly imbalanced datasets, where positive samples are scarce. Nonetheless, this indicates that the model confidently identifies instances resembling positives, reflecting its ability to capture subtle discriminative patterns even with limited positive training samples.

Fig. 6.

Calibration plots for the best performing models, illustrating the agreement between predicted probabilities and observed fraction of positive samples for (a) tremor, (b) bradykinesia, and (c) dyskinesia detection

Feature importance analysis and physiological interpretation

To gain a deeper understanding of the decision mechanisms of the best performing models and assess their biological plausibility, feature importance was analyzed using SHAP values. Since the RF classifiers outperformed the other models, the TreeExplainer implementation of SHAP was employed, which provides computationally efficient SHAP value estimates for tree-based models. Figure 7 presents the ten most influential features for the detection of tremor, bradykinesia, and dyskinesia. Notably, the observed high-ranked features are consistent with known physiological characteristics of each symptom, highlighting the distinct temporal, spectral, and complexity properties that influence model predictions in each case, as detailed below.

Fig. 7.

Top 10 important features, ranked by their mean absolute SHAP values for (a) tremor, (b) bradykinesia, and (c) dyskinesia detection

First, regarding tremor detection, the dominant frequency of the first principal component signal (PC1_fft_dom_freq) proved to be the most influential feature, with higher dominant frequencies strongly pushing predictions toward tremor presence (a 1-standard deviation (SD) increase raised the predicted probability by 2.3%), while lower dominant frequencies pushed predictions toward tremor absence. This spectral pattern is physiologically meaningful, as PD tremor manifests at a high frequency in the range of 3–12 Hz. Similarly, features reflecting temporal oscillations, consistent with periodic PD tremor manifestations, such as the ratios of peaks, sign slope changes, and zero crossings (Magnitude_peaks_rt, Magnitude_ssc_rt, PC1_zero_cross_rt, PC1_ssc_rt, and PC1_peaks_rt) followed in importance, demonstrating the same behavior, with increases of 1 SD leading to 1.7–2.9% increase in tremor predicted probability.

In contrast, other influential features demonstrated the opposite behavior, with lower values pushing predictions toward tremor presence and higher values pushing predictions toward tremor absence. These features include the spectral entropy (PC1_fft_entropy) and the ratio of values above the mean in the frequency domain (Magnitude_fft_above_mean_rt). Both of them are consistent with the high concentration of spectral power around the dominant frequency of PD tremor. Interestingly, higher sample entropy (Magnitude_sampen) and lower detrended fluctuation analysis coefficients of the magnitude signal (Magnitude_dfa) pushed predictions toward tremor presence, a pattern not directly expected from the physiology of tremor but likely reflecting combinatorial interactions with other features, potentially capturing subtle oscillation amplitude variance, highlighting the existence of multivariate dependencies beyond individual feature effects.

Regarding bradykinesia detection, the two most influential features were the standard deviation (Magnitude_std_dev) and the range of the magnitude signal (Magnitude_maxmin_diff) in the time domain, with lower values strongly pushing predictions toward bradykinesia presence and higher values pushing predictions toward bradykinesia absence (increases of 1 SD reduced the predicted probability by 1.1–1.6%). Other influential temporal features, including skewness (Magnitude_skewness), maximum value (Magnitude_max), variance (Magnitude_var), and mean deviation of the magnitude signal (Magnitude_avg_diff_mean), demonstrated the same behavior, with increases of 1 SD decreasing the bradykinesia predicted probability by 0.2–1.6%. All these temporal patterns are physiologically meaningful for bradykinesia, as they indicate low amplitude, slowness, and low variability in acceleration signals.

On the other hand, the most influential spectral features for bradykinesia detection demonstrated the opposite behavior, with higher values pushing predictions toward bradykinesia presence (increases of 1 SD led to 0.6–1.7% increase in the predicted probability) and lower values pushing predictions toward bradykinesia absence. These features corresponded to the ratios of slop sign changes, peaks, and rest periods in the frequency domain (PC1_fft_ssc_rt, PC1_fft_peaks_rt, PC1_fft_rest_rt, and Magnitude_fft_rest_rt). These patterns reflect the broader, flatter spectral content in low frequencies, typical of bradykinesia, where reduced, non-periodic motion disperses spectral power across multiple small peaks rather than a single dominant frequency.

Finally, regarding dyskinesia detection, the detrended fluctuation analysis coefficient of the magnitude signal (Magnitude_dfa) was the most influential feature, with lower and moderate values pushing predictions toward dyskinesia presence and higher values pushing predictions toward dyskinesia absence (a 1 SD increase reduced the predicted probability by 1.2%). Sample entropy of the first principal component (PC1_sampen) followed, contributing positively to the prediction of dyskinesias (a 1 SD increase raised the predicted probability by 1.3%). Both patterns indicate that more irregular and less self-similar movements are more likely to be classified as dyskinesias, which is consistent with the physiology of this symptom. Conversely, a lower rest ratio in the magnitude signal (Magnitude_rest_rt) increased the predicted probability of dyskinesia, with a 1 SD decrease leading to a 1.2% rise in dyskinesia likelihood. This is also consistent with the excessive involuntary dyskinetic movements, that inherently reduce resting periods. Similarly, higher ratios of peaks and slope sign changes (Magnitude_peaks_rt, Magnitude_ssc_rt), as well as greater deviation from the median (Magnitude_med_dev) also pushed predictions toward dyskinesia presence. These patterns capture the increased variability and irregular amplitude fluctuations that are typical of erratic motions observed during dyskinetic episodes.

The most influential spectral features for dyskinesia detection reinforced the time domain findings. Higher ratios of spectral power above the mean (PC1_fft_above_mean_rt), greater spectral entropy (PC1_fft_entropy), higher ratios of spectral peaks (PC1_fft_peaks_rt), as well as larger spectral interquartile range (PC1_fft_iqr) values all contributed positively to dyskinesia predictions, with increases of 1 SD raising the dyskinesia predicted probability by 0.8–1.2%. These patterns indicate a broader, more complex spectral distribution at low frequencies, rather than power concentrated around a single dominant frequency, which is consistent with the erratic, non-periodic nature of dyskinetic movements. Overall, these findings suggest that dyskinesia is characterized by irregular, high-variability motion in the time domain and dispersed energy in the frequency domain, aligning with its known physiological presentation as uncontrolled, involuntary motor activity.

Comparison with PD Digital Biomarker DREAM challenge results

As described in Sect. 1, this study substantially expanded the scope of the PD Digital Biomarker DREAM challenge by detecting PD motor symptoms in both upper extremities, independently of which side is the most or least affected and which limb is active or at rest, across all 18 motor tasks, rather than a limited subset. This approach more accurately simulates free-living scenarios in which neither the patient’s activity nor the hand used is known.

Despite the significant increase in complexity, our approach demonstrated performance comparable to that of the winning solutions of the PD Digital Biomarker DREAM Challenge, achieving up to 0.664, 0.636, and 0.443 AUPRC for tremor, bradykinesia, and dyskinesia detection, respectively. Notably, across all three cases, this increase in problem difficulty resulted in only a slight to moderate reduction in the AUPRC obtained, underscoring the scalability of the proposed method.

More specifically, for dyskinesia detection, complexity increased by 450%, while performance decreased only by 7.13%, indicating that each 1% drop in AUPRC corresponded to a substantial 63.13% increase in complexity. Similarly, for tremor detection, complexity increased by 175% with only a 11.47% reduction in AUPRC, translating to a 15.26% increase in complexity per 1% drop in AUPRC. Regarding bradykinesia detection, complexity increased by 128.57% with a moderate 33.05% reduction of AUPRC, still resulting in a net positive trade-off of 3.89% complexity gain per 1% performance loss.

However, in the PD Digital Biomarker DREAM Challenge, bradykinesia detection is less indicative of true model performance due to the greater influence of metadata. While metadata were used in all three cases in the original challenge, their impact was particularly strong in bradykinesia detection, where even a simple baseline model trained solely on metadata achieved an overly high AUPRC of 0.813 [56]. In contrast, our analysis was based exclusively on features extracted from tri-axial acceleration time series, deliberately omitting metadata to highlight the predictive power of accelerometer data alone.

Furthermore, unlike evaluation in the PD Digital Biomarker DREAM challenge, which relied on a single predefined train-test split, our method was evaluated using a more robust stratified group 5-fold cross-validation strategy, ensuring greater reliability. While the results are not directly comparable due to our expanded scope, use of only accelerometer data, and different evaluation strategy, this benchmark demonstrates the scalability and effectiveness of the proposed approach in addressing increasingly difficult problems with fewer data, making a step toward detecting PD motor symptoms in uncontrolled real-world settings, ultimately enabling tailor-made treatment suggestions.

Evaluation on the ALAMEDA datasets

The proposed pre-trained RF classifiers that achieved the highest performance in the experiments above were also evaluated on the ALAMEDA PD tremor and dyskinesia datasets without any further tuning. In this external evaluation, tremor was detected with 0.879 ± 0.082 AUPRC, recall, and F1-score, with macro average recall and F1-score of and , respectively, and weighted average recall and F1-score of and , respectively. Dyskinesia was detected with 0.648 ± 0.113 AUPRC, recall, F1-score, with macro average recall and F1-score of and , respectively, and weighted average recall and F1-score and , respectively.

These results demonstrate that the proposed approach detected symptoms more accurately on MDS-UPDRS assessments in controlled clinical settings than in simulated free-living environments, which is expected due to the higher variability and noise in the latter. Moreover, the ML models were retrained based on the ALAMEDA PD tremor and dyskinesia datasets. However, in both cases, retraining did not significantly improve classification results, indicating that the proposed ML models generalize well on previously unseen data without the need for task-specific adjustments. Importantly, SHAP-based feature importance analysis revealed that the majority of the most influential features remained consistent across datasets for the proposed classifiers, demonstrating the same directionality in their contribution toward symptom prediction, despite slight differences in ranking and SHAP value range.

Specifically, for tremor detection, dominant frequency, ratios of peaks, slope sign changes, and zero crossings, along with the detrended fluctuation coefficient remained among the top predictors, confirming that the proposed tremor detection model rely on physiologically stable markers of rhythmic, oscillations, typical of PD tremor. Similarly, for dyskinesia detection, the detrended fluctuation coefficient and sample entropy, together with rest and peak ratios in the time domain, as well as spectral features, like deviation from the median and interquartile range, were again among the most influential features. These patterns reaffirm that dyskinesia is consistently characterized by irregular, high-variability movements with broad spectral power in low frequencies. Findings for both symptoms demonstrate that consistent temporal and spectral features capture symptom-specific physiological signatures that remain robust across different environments (both controlled and simulated free-living) and datasets, highlighting that the extracted features could act as highly informative digital biomarkers and the proposed approach holds significant potential for adoption in real-world PD monitoring.

Discussion

This study addresses the problem of automated detection of three cardinal PD motor symptoms: tremor, bradykinesia, and levodopa-induced dyskinesia. These symptoms have a significant negative impact on PD patients’ quality of life and exhibit high variability in both frequency and intensity, not only among different patients but also within the same patient throughout the day, following the cycle of the medication effect. Therefore, adopting objective, unobtrusive monitoring of these PD motor symptoms in clinical practice is of utmost importance to support personalized, evidence-based treatment suggestions.

To meet this need, herein, we developed a modular data processing and analysis pipeline using data collected from commercial wearable devices, such as smartwatches and smart bracelets with embedded accelerometers. These devices offer a patient-friendly alternative to more burdensome interventions, and thus, can lead to increased long-term adherence and acceptance by patients [24]. The proposed approach was evaluated on multiple datasets, including the Levodopa Response Study dataset and two newly introduced ALAMEDA datasets. These datasets contain accelerometer data from PD patients while performing both standard clinical assessment exercises and ADLs.

This pipeline enables the fully automated detection of PD tremor, bradykinesia, and dyskinesia in real-world settings, as the developed models are task- and hand-agnostic. Unlike previous approaches [24, 31, 52], they do not require any information on the task the patients are performing and the kind of motion that this involves, making the proposed approach versatile but also highly challenging. However, such flexibility is essential for transitioning from restricted clinical environments to free-living settings, eventually, enabling continuous, unobtrusive monitoring of PD symptoms in real-world applications for personalized disease management.

The proposed data processing pipeline consists of several preprocessing steps, including data interpolation to handle missing values, bandpass filtering to maintain only relevant frequency bands for each symptom, computation of magnitude and first principal component signals to attenuate dependency on sensor placement and orientation, segmentation into overlapped windows of various sizes, feature extraction to capture distinct symptom characteristics in both time and frequency domains, after applying FFT, resampling of observations to address class imbalance, feature standardization, and dimensionality reduction in the feature space. These preprocessing steps are accompanied by one of the following ML models: LR, kNN, MLP, SVM, DT, AdaBoost, or RF to finally predict the occurrence of PD motor symptoms.

The main findings indicate that larger segmentation windows approaching the full duration of each exercise led to improved classification results across all symptom detection tasks, suggesting that bradykinetic slow movements and dyskinetic irregular movements require longer windows to be effectively characterized, while tremor detection could also benefit from capturing additional oscillation cycles. Moreover, SMOTE slightly enhanced performance in some cases, but the overall limited impact of resampling suggests that model performance was constrained more by intrinsic intra-class variability than by class imbalance alone. Similarly, PCA demonstrated competitive performance, but the overall limited impact of dimensionality reduction highlighted the high quality, physiological relevance, and complementarity rather than redundancy of the initial hand-crafted feature set.

Among the ML models tested, the RF classifiers consistently demonstrated better classification performance in detecting tremor, bradykinesia, and dyskinesia, achieving 71.9%, 71.7%, and 83.4% accuracy, and 0.664, 0.636, and 0.443 AUPRC respectively. While LR and SVM outperformed RF in certain evaluation metrics for tremor and dyskinesia, RF demonstrated greater overall robustness and interpretability. Calibration analysis also revealed well-calibrated predictions for tremor and bradykinesia, and slightly overconfident yet well-ranked predicted probabilities for dyskinesia, which is consistent with the higher class imbalance observed in this case. Importantly, SHAP-based feature importance analysis exhibited physiologically meaningful patterns that differed across symptoms, strengthening the physiological plausibility of the proposed approach.

Notably, compared to the original PD Digital Biomarker DREAM Challenge [56], this study addresses a significantly more complex problem, detecting symptoms on both hands, incorporating a wider range of tasks, simulating free-living conditions, and relying solely on accelerometer data without additional metadata. Despite this substantially increased complexity (128–450%), the proposed models achieved comparable performance to the winning challenge solutions, with only a modest reduction in AUPRC (7–33%), highlighting the scalability of the proposed pipeline and its readiness for real-world application.

Furthermore, when evaluated on the new ALAMEDA datasets, the pre-trained RF models achieved AUPRC scores of 0.879 for tremor and 0.648 for dyskinesia. The higher performance, compared to the Levodopa Response Study, observed in these more controlled clinical environments, together with the fact that no further tuning was required, indicates the strong generalization capability of the proposed pipeline. Notably, the most influential features for each symptom remained largely consistent across datasets, exhibiting similar directions of contribution toward symptom prediction, highlighting the stability and physiological relevance of these digital biomarkers across different contexts. Another consistent pattern observed across both the Levodopa Response Study and the ALAMEDA datasets was the higher detection performance for tremor compared to dyskinesia, despite the overall improvement in the latter dataset. Given that the same processing and modeling framework was applied across datasets, this consistent performance gap may reflect intrinsic differences between the two motor symptoms, including their relative sensitivity to contextual or clinical factors.

Finally, the findings confirmed that conventional evaluation metrics, such as accuracy, precision, recall, and F1-score may be misleading in imbalanced datasets, like those used in this study. Therefore, metrics, like AUROCC, AUPRC, MCC, and CK should be prioritized, as they provide a more accurate assessment of classification performance in imbalanced datasets. However, even these metrics should be interpreted collectively. For example, while kNN yielded higher AUPRC values for bradykinesia and dyskinesia detection, its lower overall performance in other metrics revealed potential overfitting. Crucially, the adopted evaluation strategy properly handles groups of dependent samples from the same patients, preventing data leakage and providing an important methodological improvement over other prior studies.

Collectively, these findings demonstrate that task- and hand-agnostic models trained on accelerometer data from commercial wearable devices can effectively detect cardinal PD motor symptoms, like tremor, bradykinesia, and dyskinesia. The combination of interpretable, physiologically relevant features with a modular and generalizable processing pipeline lays the foundation for reliable digital biomarkers capable of translating objective symptom monitoring from controlled clinical settings to real-world applications.

Limitations

Although this study offers valuable contributions, achieving performance comparable to that of the winning solutions in the PD Digital Biomarker DREAM Challenge [56] despite addressing a more complex and challenging problem, the results remain moderate when evaluated on the Levodopa Response Study dataset. This can be largely attributed to several factors inherent in the dataset and the complexity of the problem addressed. These include the small cohort size, the short 4-day monitoring period, and high class imbalance, with some patients never exhibiting certain symptoms. Moreover, symptom occurrence varied both between patients and within the same patient, as the study captured transitions across the medication effect cycle (from ON to OFF states during the first day and from OFF to ON during the last), further complicating the predictions.

Similarly, our focus on detecting PD symptoms across all 18 motor tasks, rather than a small subset, in both upper limbs, regardless of whether the limb is most or least affected and active or at rest, relying solely on upper-extremity accelerometer data further increased the complexity of the problem. In particular, dyskinesia detection demonstrated the lowest performance in terms of AUPRC compared to other symptoms across both the Levodopa Response Study and the ALAMEDA datasets. Notably, this consistent pattern was observed despite the use of the same raw accelerometer data, processing and modeling framework across symptoms, suggesting that the reduced performance is unlikely to be driven by differences in data collection or analysis but may instead reflect limitations in the predictive power of accelerometer data alone for dyskinesia detection.

The inherent overlap between voluntary movements and involuntary dyskinesias makes their differentiation extremely challenging, as also noted in previous studies [52, 54]. Moreover, dyskinesia is a levodopa-induced adverse motor effect that is closely linked to medication intake and fluctuations in dopaminergic levels, typically occurring during periods of high levodopa concentration or transitions between motor states [19]. Consequently, the findings may suggest a natural limit to the predictive power of acceleration signals alone for dyskinesia detection, while performance could potentially further improve through the integration of contextual or clinical information, such as medication intake, motor state, the specific exercise performed, or the limb involved.

These challenges highlight the difficulty of accurately detecting PD motor symptoms with a minimum set of sensors under realistic and variable conditions, as opposed to more controlled settings in which patients are instructed to perform only a few standardized tasks over very short periods that are insufficient to reflect changes across the medication effect cycle. Notably, in line with that, improved results were obtained in the ALAMEDA datasets, where data were collected during more structured clinical assessments of approximately 30 minutes. This underscores the trade-off between achieving higher model performance under more restricted conditions and addressing the greater variability of more realistic free-living scenarios.

More broadly, there are limited publicly available datasets to evaluate the proposed approach and develop robust models for PD motor symptom detection. Most of them, including those used in this study, are relatively small, which may not only constrain performance but also limits the scope of the study and the problems that can be addressed. For example, the Levodopa Response Study dataset contains accelerometer data from only 28 PD patients over a short 4-day monitoring period. This short duration does not support long-term monitoring and is not suitable for estimating the progression of the disease. On the other hand, the ALAMEDA PD tremor and dyskinesia datasets contain accelerometer data from multiple clinical visits throughout a year, supporting longitudinal monitoring, but these data are collected from a small cohort of only 11 PD patients.

The small dataset size also limits the types of ML models that can be effectively trained. The fact that conventional signal processing techniques for manual feature extraction outperformed DL architectures, such as CNNs, for automated feature extraction in the PD Digital Biomarker DREAM Challenge [56] can partially be attributed to the small size of the Levodopa Response Study dataset. Given this limitation, DL methods were intentionally not evaluated in the current study, which focuses on extending the best performing approaches that rely on hand-crafted features based on signal processing techniques. However, manual feature engineering relies heavily on domain expertise, introducing a risk of bias and potentially missing additional hidden patterns in acceleration signals that may improve classification performance.

Finally, although the datasets used contain annotated observations of patients performing ADLs, these ADLs correspond to a pre-defined battery of motor tasks in a restricted laboratory environment. As a result, only a limited simulation of free-living conditions was achieved with the selected datasets. This trade-off was essential to ensure accurate annotations provided by clinicians to support the supervised learning methods implemented in the current study. However, obtaining reliable annotations in truly uncontrolled, free-living environments remains a challenge.

Future work

In the future, the proposed approach could be seamlessly integrated into real-world clinical workflows, requiring only that PD patients wear a commercial smartwatch or smart bracelet with an embedded accelerometer on either wrist. The collected data could then be securely transferred to a cloud-based platform or a web application offering tailored interfaces for clinicians and patients or their caregivers, where the proposed pre-trained pipeline would automatically infer the presence or absence of PD symptoms. To amplify data privacy, an alternative offline workflow could be adopted, allowing data uploads via a wired connection during routine clinical visits. This integration pathway is already under development within the ALAMEDA Innovation Hub [78], by providing summaries of symptom occurrences over the past few hours or days to support personalized data-driven clinical decision-making.

Moreover, the user interfaces could be enhanced by incorporating interactive visualizations of feature contributions to model decisions, based on the SHAP-based feature importance analysis presented in this study. These visualizations would provide both global and local explanations, helping clinicians understand the overall logic of the model, as well as the reasoning behind individual predictions. Feature interaction plots could further illustrate how coexisting symptom signatures jointly influence predictions, offering deeper clinical insights. Importantly, all the extracted features have clear physiological meanings, enabling the transparent implementation of this explainability layer, unlike DL approaches that are often treated like “black boxes”. Enhancing interpretability will not only improve clinical trust and patient acceptance but also enable feedback loop systems where neurologists can adapt treatment plans based on objective, interpretable evidence, underscoring the translational impact of this work.

Alternatively, future implementations could deploy the proposed models directly on smart wearable devices or in paired smartphones (e.g., via Bluetooth connectivity), adhering to edge AI principles. The models explored in our experiments are lightweight enough to run efficiently on smartphones, and most can be adapted for resource-constrained devices, such as smartwatches and smart bracelets, with minimal optimization through embedded ML techniques, like model quantization and pruning. These optimizations will ensure energy efficiency while maintaining high performance. Deploying our models on edge devices will enable local inference, reducing latencies and providing near real-time estimations, while enhancing data privacy by avoiding continuous cloud transmission of sensitive data. This hybrid architecture could enhance patient autonomy and acceptance, while maintaining meaningful clinical feedback loops, paving the way for personalized PD care.

In the future, we plan to employ larger datasets that contain accelerometer and other multimodal data from a larger cohort of PD patients over longer monitoring periods. This will enable us to evaluate our proposed approach in a more robust way, potentially improve classification performance, study the progression of the disease, and observe how these symptoms evolve over time. Larger datasets may also allow the use of LOSO cross-validation for evaluation, which is more closely aligned with clinical application scenarios. While LOSO is typically unstable in small, imbalanced datasets due to high variance in performance estimates, its application becomes more reliable as dataset size increases, and the respective increase in computational cost could be addressed through parallelization.

Additionally, employing larger datasets will support experimentation with state-of-the-art architectures, such as transformers and other DNNs, that require large amounts of data for effective training. We will also explore hybrid methods that combine the strengths of the hand-crafted features used in the current study with DL approaches, to potentially optimize performance. More advanced data augmentation and resampling techniques based on generative neural networks could also be employed to address the high class imbalance observed in the datasets. Data balancing and generative adversarial networks (GANs) for synthetic data generation have been shown to stabilize gait classification in rare diseases [79]. So, similar approaches could be adopted to potentially improve the detection of less frequent PD symptoms using wearable sensors.

Finally, multi-hour data collected under actual free-living conditions and not just simulated ones will be also used to validate the proposed approach and experiment with additional methods. Such data may be fully or partially unlabeled with variable quality of annotations, as patients may report their symptoms inconsistently or subjectively over time. To leverage these large, unlabeled or partially annotated datasets, state-of-the-art self-supervised and semi-supervised learning methods will be employed. In addition, the incorporation of medication logs, motor state (ON/OFF), and other clinical or contextual information will be explored to potentially increase the accuracy of the predictions, particularly for dyskinesia, a symptom closely related to levodopa intake and motor state fluctuations. Recorded medication intakes and motor states could also be compared against the predicted symptoms to validate their clinical relevance.

Conclusions

This study proposed and evaluated across multiple datasets a modular data processing and analysis pipeline for the automated detection of cardinal PD motor symptoms using accelerometer data from commercial wearable devices. The proposed context-agnostic models ignore the exercise performed and the hand used, as well as any other clinical or contextual information, including medication intakes, and detect tremor, badykinesia, and dyskinesia using only wrist accelerometer data. Despite the significantly increased complexity of the addressed problem compared with previous approaches and the minimal sensing setup used to ensure unobtrusiveness and patient acceptance, the proposed approach achieved competitive performance, demonstrating high scalability and strong generalization across datasets. In addition, the feature importance analysis revealed physiologically meaningful patterns, while the interpretability of the models can further support clinical adoption and increase trust in the proposed approach. These findings highlight the potential of commercial wearables and machine learning for passive, objective, and unobtrusive monitoring of PD motor symptoms in real-world settings. Ultimately, this is expected to assist clinicians in decision-making processes, supporting more tailor-made treatment suggestions and personalized, patient-centric PD management.

Abbreviations

ACC

Accuracy

AdaBoost

Adaptive boosting

ADAM

Adaptive moment estimation

ADL

Activities of daily living

AI

Artificial intelligence

ANOVA

Analysis of variance

AUPRC

Area under the precision-recall curve

AUROCC

Area under the receiver operating characteristic curve

CDRS

Clinical Dyskinesia Rating Scale

CK

Cohen’s kappa

CNN

Convolutional neural network

CWT

Continuous wavelet transform

DALY

Disability-adjusted life year

DFA

Detrended fluctuation analysis

DFT

Discrete Fourier transform

DL

Deep learning

DNN

Deep neural network

DREAM

Dialogue for Reverse Engineering Assessment and Methods

DT

Decision tree

DWT

Discrete wavelet transform

FFT

Fast Fourier transform

FN

False negatives

FP

False positives

FPR

False positive rate

GAN

Generative adversarial network

GDP

Gross domestic product

i.i.d.

Independent and identically distributed

IIR

Infinite impulse response

IMU

Inertial measurement unit

IoT

Internet of things

kNN

k-nearest neighbors

LOSO

Leave-one-subject-out

LR:

Logistic regression

LSTM

Long short-term memory

MCC

Mathew’s correlation coefficient

MDS

Movement Disorder Society

ML

Machine learning

MLP

Multilayer perceptron

NN

Neural network

PC1

First principal component

PCA

Principal component analysis

PD

Parkinson’s disease

PREC

Precision

RBF

Radial basis function

RF

Random forest

RDRS

Rush Dyskinesia Rating Scale

REC

Recall

ReLU

Rectified linear unit

RMS

Root mean square

RMSE

Root mean squared error

RNN

Recurrent neural network

ROCKET

Random convolutional kernel transform

ROS

Random oversampling

RUS

Random undersampling

SCOPA

Scales for Outcomes in Parkinson’s Disease

SD

Standard deviation

SDI

Socio-demographic index

SHAP

Shapley additive explanations

SMOTE

Synthetic minority oversampling technique

SVM

Support vector machine

TN

True negatives

TP

True positives

TPR

True positive rate

UDysRS

Unified Dyskinesia Rating Scale

UPDRS

Unified Parkinson’s Disease Rating Scale

VAE

Variational autoencoder

XGBoost

Extreme gradient boosting

Author contributions

Conceptualization: K.-M.G., I.R., and P.T.; Methodology: K.-M.G.; Software: K.-M.G.; Validation, K.-M.G.; Formal analysis: K.-M.G.; Investigation: K.-M.G.; Data curation: K.-M.G.; Resources: K.-M.G., I.R., and P.T.; Writing—original draft preparation: K.-M.G.; Writing—review and editing: I.R.; Supervision: I.R.; Project administration: I.R. and P.T.; Funding acquisition: I.R. and P.T.; All authors have read, reviewed and agreed to the published version of the manuscript.

Funding

This research was partially funded by the European Union through the Horizon Europe Programme, as part of the ENFIELD - European Lighthouse to Manifest Trustworthy and Green AI project (GA No. 101120657) and the Horizon 2020 Programme, as part of the ALAMEDA - Bridging the Early Diagnosis and Treatment Gap of Brain Diseases via Smart, Connected, Proactive and Evidence-based Technological Interventions project (GA No. 101017558). The funding bodies played no role in the design of the study, the collection, analysis, and interpretation of data and in writing the manuscript. Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union.

Data availability

Qualified researchers can access the “Levodopa Response Study” dataset upon agreement to specific conditions for use and completion of a 6-step application process through Synapse (https://www.synapse.org/#!Synapse:syn20681023/wiki/). The ALAMEDA datasets are publicly available through Zenodo (PD tremor dataset: https://zenodo.org/records/10782573 and PD dyskinesia dataset: https://zenodo.org/records/10724649).

Declarations

Ethics approval and consent to participate

The study received approval from the Ethics Committee of the Special Outpatient Clinic of Parkinson’s Disease and Related Movement Disorders at the First Department of Neurology of the National and Kapodistrian University of Athens, in Eginitio Hospital. All participants contributing to the ALAMEDA datasets provided their written informed consent upon enrollment. The study was conducted in accordance with the ethical principles of the Declaration of Helsinki on research involving human participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.