Early Detection of Parkinson’s Disease Using a Single-Arm Wearable Sensor and Convolutional Neural Networks

Early Detection of Parkinson’s Disease Using a Single-Arm Wearable Sensor and Convolutional Neural Networks | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Early Detection of Parkinson’s Disease Using a Single-Arm Wearable Sensor and Convolutional Neural Networks Hyejin Choi, Changhong Youm, Hwayoung Park, Bohyun Kim, Juseon Hwang, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7285962/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Early-stage Parkinson’s disease (PD) is characterized by subtle motor symptoms that complicate diagnosis and often delay intervention. Timely and accurate identification is critical for effective management, emphasizing the need for objective, non-invasive diagnostic methods. Methods This study developed a non-invasive approach for early PD detection using wearable sensors and a convolutional neural network (CNN) during a 6-min walk test. The test was segmented into 1-min intervals, extracting three straight-walking and three turning gait phases per minute. Time-series data were collected from 78 patients with early-stage PD and 50 healthy controls across six body locations. Results The CNN achieved 95.6% accuracy when classifying PD status using gyroscope data from the left arm during the first-minute straight-walking phase. Furthermore, repeated-measures analysis of variance and post hoc tests indicated that a 1- to 2-min measurement window was sufficient for reliable detection, supporting the feasibility of time-efficient clinical screening. Conclusions These findings suggest that a wearable sensor, placed on a single arm and used to capture first-minute straight gait data, can provide highly accurate and non-invasive early PD detection. Future research should evaluate medication effects, extend validation to broader disease stages, and explore unsupervised learning approaches to identify latent motor phenotypes and enable personalized monitoring. Parkinson’s disease gait wearable sensors artificial intelligence deep learning neurodegeneration Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 BACKGROUND Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by motor symptoms, including bradykinesia, rigidity, resting tremor, and postural instability [ 1 ]. Gait disturbances are particularly disabling and often emerge early, manifesting as reduced step length, decreased walking speed, festination, and impaired turning [ 2 , 3 ]. Subtle asymmetries, such as unilateral shuffling and diminished arm swing, frequently appear in early-stage PD but may go undetected without overt gait complaints [ 4 , 5 ], especially since some features overlap with normal aging [ 6 ]. Current clinical assessments, including the Hoehn and Yahr (H&Y) Scale and the Unified Parkinson’s Disease Rating Scale, rely on clinician observations and are subject to inter-rater variability and subjectivity [ 7 , 8 ]. Although laboratory-based systems such as three-dimensional motion capture and instrumented walkways provide precise gait measurements, their high cost and operational complexity limit their routine clinical use [ 9 ]. In contrast, wearable sensors offer a portable, low-cost, and scalable solution for continuous, real-world gait monitoring [ 10 ]. As a result, there is growing interest in leveraging wearable sensor data to develop objective tools for PD detection and severity assessment [ 11 ]. The 6-min walk test (SMWT) is widely used to assess functional endurance under semi-natural conditions. It captures both the straight and turning gait phases, enabling a more ecologically valid evaluation of mobility [ 12 ]. While previous studies have primarily focused on total walking distance as the main outcome [ 13 , 14 ], phase-specific gait dynamics—particularly during turning—may better reveal early motor impairments. Turning requires more complex motor coordination and often exhibits greater variability, making it a sensitive indicator of dysfunction [ 15 ]. Despite recommendations to use segment-based analysis, few studies have examined temporal variations in gait quality during the SMWT. Segment-based analysis involves dividing the walking test into smaller, distinct time intervals to evaluate specific gait phases, such as straight walking and turning, separately. For example, Bohannon et al. and Valet et al. reported that performance during the first 2 min approximates full-test outcomes, although their analyses focused exclusively on walking distance [ 13 , 14 ]. Hadouiri et al. extended this approach in patients with multiple sclerosis by analyzing gait data minute by minute, revealing time-dependent changes in spatiotemporal metrics [ 16 ]. However, no comparable analysis has been conducted for PD using deep learning techniques. Moreover, existing studies have primarily focused on lower-limb data, often overlooking upper-limb and trunk movements. Deep learning methods, particularly convolutional neural networks (CNNs), have shown promise in detecting PD from wearable sensor data, outperforming traditional machine learning techniques in capturing non-linear and phase-dependent gait features [ 17 – 19 ]. Our previous work demonstrated that CNNs trained on full-duration SMWT data could effectively distinguish patients with early-stage PD from healthy controls [ 20 ]. However, CNN performance across specific gait phases and short time intervals has not been systematically investigated. Determining whether early-stage PD can be accurately classified within a shorter timeframe, such as the first or last minute, could streamline assessments and reduce participant burden [ 21 ]. Therefore, this study aimed to: (1) evaluate CNN-based classification of individuals with early-stage PD using multi-segment wearable sensor data collected during straight and turning gait phases of the SMWT; and (2) compare classification performance across different time intervals and sensor locations to determine the minimal reliable measurement duration and optimal sensor placement, with particular focus on the sensor previously shown to yield the highest classification accuracy. METHODS Participants Overall, 78 individuals with early-stage PD and 50 age-matched controls were enrolled. All participants were diagnosed with idiopathic PD according to the United Kingdom Parkinson's Disease Society Brain Bank criteria, as confirmed by a neurologist. Inclusion criteria were H&Y stages 1–2 and Mini-Mental State Examination scores ≥ 24. Individuals with comorbid musculoskeletal, cardiovascular, or cognitive conditions, or those with recent orthopedic surgery, were excluded. Only participants who completed the full SMWT were included in the final analysis. The recruitment flow diagram is shown in Fig. 1 , and demographic data are presented in Table 1 . The study protocol was approved by the Institutional Review Board of Dong-A University Medical Center (IRB number: DAUHIRB-22-089) and conducted in accordance with relevant guidelines and regulations. All participants provided written informed consent. The study was registered with the Clinical Research Information Service of Korea (KCT0009353). Table 1 Physical and clinical characteristics of participants Early PDs (n = 78) Controls (n = 50) p -value Sex (male/female) 37 / 41 21 / 29 0.588 a Age (years) 67.51 ± 7.08 65.80 ± 5.51 0.149 b Height (cm) 161.20 ± 7.72 161.14 ± 8.38 0.907 b Body mass (kg) 63.96 ± 10.54 63.68 ± 10.04 0.884 b BMI (kg/m 2 ) 24.51 ± 3.13 24.40 ± 2.43 0.847 b Disease duration (years) 5.50 ± 3.96 - - Treatment duration (years) 4.19 ± 3.25 - - L-Dopa equivalent dose (mg/day) 541.90 ± 286.54 - - MMSE (scores) 28.12 ± 1.87 27.22 ± 1.90 0.003 c MoCA (scores) 26.23 ± 2.88 24.86 ± 2.99 0.001 c UPDRS Total (scores) 47.68 ± 19.54 - - UPDRS III (scores) 25.35 ± 13.63 - - H&Y Scale (stages 1 and 2) 28 / 50 - - SMWT (m) 414.24 ± 85.16 494.23 ± 51.42 < 0.001 c The data are presented as the mean ± standard deviation. Significant difference: p < 0.05; PD: Parkinson’s disease; BMI: Body mass index; L-Dopa: Levodopa; MMSE: Mini-Mental State Examination; MoCA: Montreal Cognitive Assessment; UPDRS: Unified Parkinson’s Disease Rating Scale; H&Y: Hoehn and Yahr; SMWT: 6-min walk test. a p ­value of Fisher’s exact test. b p ­value of the independent t­test. c p ­value of the Mann–Whitney U test. Experimental Procedures All participants completed the SMWT along a 20-m hallway delineated by cones placed at each end (Fig. 2 a). The task was performed under standardized instructions: walk as fast as possible without running, with safety monitoring by study personnel. All tests were conducted approximately 2 h after medication intake during the "ON" medication state. Six Xsens DOT sensors (Movella Technologies, Enschede, Netherlands) were affixed using a stretchable belt at the following locations: the left and right upper arms (5 cm above the lateral humeral epicondyle), left and right thighs (10 cm above the lateral femoral epicondyle), thoracic spine (T10), and lumbar spine (aligned with the midpoint of the posterior superior iliac spine) (Fig. 2 a). Sensor dimensions were 36.3 × 30.35 × 10.8 mm, and weight was 11.2 g. All sensors sampled tri-axial acceleration and gyroscopic data at 60 Hz, using the East-North-Up coordinate system. Raw data were transmitted via Bluetooth 5.0 to an iPad (iOS 15.6.1; Apple, Cupertino, CA) using the MovellaDOT application. All subsequent analyses were performed using MATLAB R2023a (MathWorks, Natick, MA). Figure 2 b illustrates the overall framework adopted in this study for classifying people with early-stage PD and controls using wearable sensor data. The process encompasses four primary stages: collecting raw TS data via wearable sensors, applying preprocessing techniques to prepare the data, splitting the dataset into training and testing subsets, and finally, training a CNN model to assess classification performance. Data Pre-processing and Gait Phase Detection Raw TS data from all six sensors were segmented into six 1-min intervals to facilitate time-dependent analysis. Within each 1-min interval, straight and turning gait phases were detected using lumbar gyroscope data, following the approach of El-Gohary et al. [ 22 ] Turning gait was identified when angular velocity exceeded 15°/s, and the end of turning was marked when it fell below 5°/s. To exclude transient movements or artifacts, a minimum turning duration of 0.5 s (30 frames) was enforced based on gait cycle dynamics and prior literature. Straight segments were retained only if they lasted for ≥ 5 s (300 frames). Up to three valid segments were extracted per phase (straight or turning) per minute. The identified frame indices were synchronized across all sensors to preserve temporal alignment. Data pre-processing was implemented using MATLAB and Python (Python 3.10; Python Software Foundation). Normalization, Filtering, and Resampling To mitigate sensor placement variability and anthropometric differences, all signals were normalized by subtracting the median value per axis (X, Y, and Z). A second-order low-pass Butterworth filter with a 6-Hz cutoff was applied to reduce high-frequency noise. Subsequently, the segmented gait data were resampled using dynamic time-warping-based interpolation to standardize the input length: 896 and 448 frames for straight and turning segments, respectively. Segments that did not meet the required length were zero-padded. To satisfy the CNN input constraints, all final inputs were adjusted to 896 frames (to match a 224 × 224 pixel resolution, using a width factor of four). Image representations were generated using recurrence plots [ 23 ], which were applied to both the individual axes and the resultant values derived from the accelerometer and gyroscope data. For each of the 1st to 6th minutes, three straight and three turning gait phases were extracted per minute from six sensor locations. This process yielded eight TS image types per segment based on the acceleration and angular velocity data across all gait phases. Data Generation To train and evaluate the CNN models, the accumulated imaging dataset was analyzed incrementally by minutes. Specifically, datasets from 1 min were analyzed first, followed by cumulative datasets from 1 to 2 min, 1–3 min, and so on up to 6 min. For each dataset, training, validation, and test sets were randomly split by subject in ratios of 50%, 20%, and 30%, respectively. As sample counts varied across minutes (Additional file 1), a representative 1-min example illustrates the oversampling procedure. That dataset initially comprised 234 early-stage PD samples and 150 control samples per segment across six sensors (left arm, right arm, left thigh, right thigh, thoracic spine, and lumbar spine). To address this imbalance, data were randomly oversampled using the imbalanced-learn (version 0.10.1) Python package by generating new minority-class samples to achieve parity [ 24 ]. Consequently, 234 samples per class (468 total) were generated per sensor segment for the 1-min dataset. The same balancing strategy was applied to all other minute-level datasets. CNN Model Training Model training and classification were conducted for each accumulated minute using the processed TS image dataset. For variance analysis, a five-fold cross-validation was applied. The batch size used during training was 128. Adam optimizers were employed, with a learning rate of 1e-05. After 150 epochs, the proposed model converged, and training was halted using early stopping. To reduce the risk of overfitting, only the best model was retained. This implies that during the training phase, the model was saved if the validation accuracy of a given epoch surpassed the highest accuracy achieved thus far. To classify people with early-stage PD and controls based on the processed TS images, the following three CNN architectures were employed: ResNet [ 25 ], DenseNet [ 26 ], and SqueezeNet [ 27 ]. The input image size for all CNN models was set to 224 × 224 pixels. We applied the ResNet model comprising 18 layers. The DenseNet-121 model, composed of four dense blocks with 6, 12, 24, and 16 channels, was also adopted. Similarly, SqueezeNet v1.0 was employed. These architectures were selected because of their proven performance and efficiency in image-based classification. The performance of the binary classification models was evaluated using standard metrics, including accuracy, precision, recall, and F1-score. Accuracy was used as the primary evaluation metric for classification problems that are well-balanced and not skewed or affected by class imbalance. In addition, confusion matrix components—true positive, true negative, false positive, and false negative—were computed to derive the precision and recall values. Precision indicates the proportion of correct positive predictions, whereas recall quantifies the number of actual positive cases accurately identified by the classifier. This metric is also known as sensitivity. The F1-score integrates both precision and recall and serves as a comprehensive metric, commonly defined as the harmonic mean of these two components. The harmonic mean is considered more appropriate than the arithmetic mean for averaging ratios such as precision recall [ 28 , 29 ]. Statistical Analysis The Shapiro–Wilk test assessed multivariate normality. Fisher’s exact test, the Mann–Whitney U test, or the independent t-test evaluated differences in physical and clinical characteristics between people with early-stage PD and controls. Statistical analyses were performed using IBM SPSS Statistics for Windows, version 21.0 (IBM Corp., Armonk, NY), MATLAB, and Python. Statistical significance was set at p < 0.05. Repeated-measures analysis of variance (ANOVA) assessed the effect of measurement duration (1–6 min) on classification accuracy across top-performing sensor-variable combinations. Bonferroni-adjusted post hoc t-tests compared specific time intervals. RESULTS CNN Classification Performance Across Gait Phases The classification accuracy from 1 to 6 min was evaluated for each sensor using time-series (TS) images processed with three CNN architectures: residual neural network (ResNet), dense convolutional network (DenseNet), and SqueezeNet. The highest accuracy from any model for each accelerometer and gyroscope variable was analyzed across all sensors for both straight (Fig. 3) and turning (Fig. 4) gait phases, where Gyr and Acc denote gyroscope and accelerometer data, respectively. X, Y, and Z represent directional axes. XYZ refers to the resultant magnitude of all three axes. Accuracy varied significantly between gait phases. As shown in Fig. 3, heat-map accuracies were consistently higher during straight gait across all sensor-variable combinations and time points. The left arm gyroscope X-axis data (Larm_Gyr_X) during straight gait yielded the highest accuracy—95.6% (precision, recall, F1-score = 0.956)—at 1 min, surpassing the full 6-min accuracy of 94.6% (precision, recall, F1-score = 0.946), suggesting that short-duration straight gait effectively captures early-stage PD features. In addition to the left arm, the sensors on the right arm (Rarm) and both thighs (Lthi and Rthi) showed strong performance during straight gait, with several variables achieving >90% accuracy at 6 min. Notable examples included Rarm_Gyr_Z (accuracy: 0.907; precision: 0.908; recall: 0.907; F1-score: 0.907), Lthi_Acc_X (accuracy: 0.905; precision: 0.906; recall: 0.905; F1-score: 0.905), Lthi_Acc_Y (accuracy: 0.919; precision: 0.919; recall: 0.919; F1-score: 0.919), Lthi_Gyr_X (accuracy: 0.903; precision: 0.903; recall: 0.903; F1-score: 0.903), Lthi_Gyr_Y (accuracy: 0.901; precision: 0.901; recall: 0.901; F1-score: 0.901), Lthi_Gyr_Z (accuracy: 0.901; precision: 0.902; recall: 0.901; F1-score: 0.901), Lthi_Acc_XYZ (accuracy: 0.909; precision: 0.910; recall: 0.909; F1-score: 0.909), and Rthi_Gyr_Y (accuracy: 0.902; precision: 0.903; recall: 0.902; F1-score: 0.902). By contrast, classification accuracy during turning gait was lower and more variable (Fig. 4). The top-performing turning variable was Rthi_Acc_Y, reaching 80.9% accuracy (precision: 0.809; recall: 0.809; F1-score: 0.809). Other leading variables included Rthi_Gyr_Z (accuracy: 0.808; precision: 0.808; recall: 0.808; F1-score: 0.808), Lthi_Gyr_Z (accuracy: 0.807; precision: 0.807; recall: 0.807; F1-score: 0.806), and Lthi_Acc_X (accuracy: 0.800; precision: 0.800; recall: 0.800; F1-score: 0.800). Despite these results, a performance gap relative to straight gait was evident across all sensors. To further illustrate model performance, Figures 5 and 6 show confusion matrices for the highest-performing sensor–variable combinations. Variables with >90% accuracy were selected for the straight gait phase and those exceeding 80% for the turning phase. The confusion matrices are based on the best results from ResNet, DenseNet, and SqueezeNet. We present representative images and corresponding frequency-domain patterns for individuals with people with early-stage PD and controls, focusing on the variables with the highest classification accuracy for each gait phase (i.e., straight and turning) at the most informative sensor location (Fig. 7 and Fig. 8). The representative images and graphs illustrate characteristic gait patterns from each minute of SMWT, capturing both straight walking and turning movements across time intervals from minute 1 to minute 6. Determining the Optimal Measurement Duration Given the superior performance of the straight gait, the analysis focused on identifying the optimal measurement duration and sensor location using straight-phase data. This approach prioritizes continuous gait patterns characteristic of early-stage PD. A. Analysis of Optimal Sensor and Variable Combinations To assess how measurement duration influenced classification, a repeated-measures ANOVA was performed using classification accuracy (averaged across the three CNN models) as the dependent variable. The top 10 sensor–variable combinations were identified based on mean accuracy across six-time points. Eight of the top 10 variables were from the left arm (Larm), highlighting its consistent performance; the left thigh (Lthi) and lumbar spine (Lumba) ranked 9th and 10th, respectively. The top-ranked variables were: Larm_Gyr_X: 93.7% (F = 5.24, p = 0.013); Larm_Gyr_Y: 93.0% (F = 1.56, p = 0.257); Larm_Acc_XYZ: 92.3% (F = 33.36, p < 0.001); Larm_Gyr_Z: 92.2% (F = 28.88, p < 0.001); Larm_Acc_X: 92.1% (F = 16.29, p < 0.001); Larm_Gyr_XYZ: 91.1% (F = 17.12, p < 0.001); Larm_Acc_Y: 90.6% (F = 48.57, p < 0.001); Larm_Acc_Z: 90.4% (F = 0.33, p = 0.886); Lthi_Acc_XYZ: 87.2% (F = 31.04, p < 0.001); and Lumba_Acc_Y: 86.5% (F = 2.61, p = 0.092). These variables demonstrated robust classification performance and statistical significance over time. B. Determining the Optimal Measurement Duration Based on Top Variables To identify the minimum duration required for stable classification, repeated-measures ANOVA and Bonferroni-corrected paired t-tests were conducted across 1–6 min intervals for each top variable. If no significant difference emerged between 1 and 6 min, additional comparisons were made between 1 and 2 min. For Larm_Gyr_X, Gyr_Y, Acc_XYZ, Gyr_Z, Acc_X, Gyr_XYZ, and Acc_Z, no significant differences were observed between 1 and 6 min (adjusted p ≥ 0.579), or between 1 and 2 min (adjusted p ≥ 0.075), indicating that 1-min measurements suffice. Larm_Acc_Y required 2 min (1 vs. 2 min: adjusted p = 0.034; 2 vs. 6 min: adjusted p = 1.000), and Lthi_Acc_XYZ likewise required ≥ 2 min (1 vs. 6 min: adjusted p = 0.048; 2 vs. 6 min: adjusted p = 0.490). In contrast, Lumba_Acc_Y was stable even at 1 min (adjusted p = 1.000). Results are presented in Additional file 2. DISCUSSION This study extends our previous work [ 20 ] by analyzing gait phase- and time-dependent classification of early-stage PD using wearable sensors and CNNs. While earlier studies focused on total gait metrics or walking distance [ 13 , 14 ], our findings highlight the diagnostic value of segmenting gait—particularly straight walking—and demonstrate the feasibility of reducing measurement duration without compromising accuracy. Classification accuracy was consistently higher during straight gait across all sensors, likely due to its repetitive, stable nature [ 17 , 30 ], which enabled CNNs to detect subtle motor deviations more effectively. The Larm_Gyr_X achieved the highest accuracy (95.6% at 1 min), highlighting the diagnostic importance of upper-limb motion [ 31 ]. In contrast, turning gait introduced greater variability and individualized compensatory strategies [ 15 , 32 , 33 ], which may have limited the CNN’s ability to learn consistent patterns [ 34 ]. The top-performing turning feature, right-thigh mediolateral acceleration (Rthi_Acc_Y), achieved only 80.9% accuracy at 6 min. Li et al. reported higher classification accuracy during turning using CNNs and lower-limb and trunk data, their protocol excluded upper-limb sensors and employed short, self-paced walking [ 35 ]. In contrast, our approach used sustained, fast-paced walking with full-body sensors, potentially capturing more consistent motor deviations in early-stage PD [ 4 , 36 ]. Compared with our previous full-signal model without gait segmentation, which achieved 83.5% accuracy using lumbar gyroscope data [ 20 ], the present phase-specific framework significantly improved both classification performance and interpretability. Specifically, Larm_Gyr_X achieved 95.6% accuracy during 1-min straight gait, demonstrating the value of analyzing distinct gait phases to uncover task-specific features [ 37 ]. Sensor-wise, upper-limb sensors—particularly the left and right arms—outperformed trunk sensors during straight gait, frequently exceeding 90% accuracy. Thigh sensors were more predictive during turning, likely reflecting their role in pivoting and balance [ 38 ]. Notably the lumbar spine sensor, which was previously optimal for non-segmented data, showed only moderate performance here, indicating that sensor relevance is highly phase-specific [ 39 ]. The strong performance of the left arm sensor aligns with early-stage PD asymmetry in arm swing [ 40 ] and supports context-specific sensor selection [ 41 ]. The left-arm sensors consistently outperformed right-arm sensors during straight gait, possibly reflecting PD-related asymmetry [ 3 ]. While PD often presents unilaterally [ 42 ], this asymmetry does not always correspond to limb dominance [ 43 ], suggesting individual variability [ 44 ]. During turning, the slightly better performance of the right-thigh sensor may indicate lateralized pivoting strategies or turning preferences [ 45 ]. These asymmetries should be considered both clinically and in model design [ 46 ]. To further characterize gait, we analyzed frequency-domain patterns for each minute of the SMWT, focusing on both straight and turning gait phases in individuals with early-stage PD and healthy controls (Fig. 7 and Fig. 8 ). For straight gait, the highest classification accuracy was observed in the Gyr_X component of the left arm, while for turning gait it was the Acc_Y component of the right thigh. Differences in frequency patterns between people with early-stage PD and controls likely influenced the generated images and contributed to classification accuracy [ 20 , 47 ]. Future work should include additional non-linear analyses to extract discriminative features related to straight and turning gait patterns in the context of bilateral coordination, asymmetry, and variability [ 48 , 49 ]. Furthermore, factors such as sex differences and symptom severity should be examined to better understand and quantify their impact. These features should then be analyzed together to clarify their contributions to classification performance. Although the SMWT is a clinical standard, our findings suggest that 1 min of straight gait data may be sufficient for accurate early-stage PD classification using left arm sensors [ 50 ]. Most top-performing features, including Larm_Gyr_X, showed no significant accuracy difference between 1 and 6 min, with peak performance at 1 min. These results support the feasibility of brief, task-specific gait assessments that reduce physical burden and simplify clinical workflows [ 13 , 14 ]. Previous studies also endorsed abbreviated testing: Bohannon et al. [ 13 ] reported strong correlations between 2-min and full SMWT distances (r = 0.968), while Valet et al. [ 14 ] found that the 2-min distance accounted for 98% of total SMWT performance in patients with stroke. However, distance-based metrics may overlook subtle motor deficits. In contrast, our CNN-based time-series analysis captured finer spatiotemporal gait abnormalities relevant to early-stage PD [ 51 ]. The stability of the classification accuracy across time intervals indicates that properly segmented, short gait recordings can effectively detect motor impairments [ 14 ]. This aligns with recent multiple sclerosis studies showing time-dependent gait changes even within short intervals [ 16 ]. Unlike prior studies focusing on stride length or gait speed, our phase-specific deep learning approach enabled granular analysis of motor features. Most top variables stabilized within 1–2 min, supporting the feasibility of time-efficient screening protocols [ 52 , 53 ]. Despite these contributions, this study has several limitations. First, all data were collected in the medication “ON” state, which may have dampened symptom severity. Future research should include “ON” and “OFF” states to evaluate generalizability [ 54 ]. Second, although we observed significant group differences in MMSE and MoCA scores, larger and more demographically diverse samples are needed to improve generalizability. Validation in external cohorts with different regional and population characteristics is also warranted to ensure robustness. Third, this study focused only on early-stage PD (Hoehn and Yahr stages 1–2). Including later stages could enhance understanding of disease progression and improve classification performance. Accounting for turning direction and limb dominance may further improve model accuracy, as these factors influence motor strategies and sensor-derived gait characteristics during the SMWT. Finally, while our CNN-based supervised learning approach demonstrated high classification accuracy, its limited interpretability remains a challenge. Future studies should integrate explainable AI (XAI) techniques and explore unsupervised or semi-supervised methods such as clustering or representation learning to uncover latent gait phenotypes and improve model interpretability [ 39 ]. Applying this analytical framework to older adults with a history of falls may also enable early classification of prodromal or early-stage PD by detecting subtle gait abnormalities predictive of fall risk, while also offering a low-cost, scalable solution for quantitative gait assessment in various settings, including home-based monitoring [ 49 ]. Longitudinal analysis of minute-by-minute non-linear feature trends could additionally provide insights into disease trajectory and progression [ 55 ]. In conclusion, this study demonstrated that patients with early-stage PD can be accurately classified using deep learning models applied to phase-segmented gait data collected from wearable sensors. By separating straight and turning gait phases and analyzing time-dependent classification performance, we identified straight gait––particularly within the first minute––as the most diagnostically informative segment. The left arm gyroscope data (Larm_Gyr_X) consistently achieved the highest classification accuracy, highlighting the importance of upper-limb motion in early-stage PD detection. These findings support the development of time-efficient, phase-specific, and interpretable digital gait assessments for early PD screening. Future research should extend this framework to include diverse disease stages and medication states and explore unsupervised learning approaches to uncover latent motor phenotypes. Declarations Ethics approval and consent to participate : All study procedures involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. The study protocol and supplementary information files were approved by the Institutional Review Board of Dong-A University Hospital (approval number: DAUHIRB-22-089) (see ethics approval letter in Additional file 3). All patients provided written informed consent before data collection. The study was registered in the Clinical Research Information Service in the Republic of Korea (KCT0009353). Consent for publication : Not applicable. Availability of data and materials : The datasets supporting this study’s findings are available from the corresponding author upon reasonable request. Code availability : We do not have an open-source code available. The codes for training and testing the deep learning models were written in Python 3.10 using PyTorch 1.9.1 and torchvision 0.10.1. Data management and feature-processing scripts were written in Python 3.10 using pandas 1.3.3 and NumPy 1.21.2. The analysis code is available upon request from the corresponding author. Funding : This research was supported by grants from the National Research Foundation of Korea (NRF), funded by the Korean government (MSIT) (No. 2022R1A2C100933711; Changhong Youm); the Basic Science Research Program through the NRF, funded by the Ministry of Education (No. 2022R1A6A3A0108756411; Hwayoung Park); and the Ministry of Education of the Republic of Korea and the NRF (No. 2024S1A5B5A16021673; Hwayoung Park). This study received no specific grants from funding agencies in the public, commercial, or non-profit sectors. The funding sources had no role in the study design, data collection, analysis, or interpretation or in writing the manuscript. Competing interests : All authors declare no financial or non-financial competing interests. Author contributions : HC, CY, HP, BK, JH, and SC conceived and designed the study. HC, HP, BK, and SC recruited the participants. HC, CY, HP, BK, JH, and SC performed the data acquisition. HC, CY, HP, BK, JH, and SC analyzed and interpreted the data. HC, CY, HP, BK, JH, and SC drafted the article. All authors read and approved the final version of the manuscript submitted. Acknowledgments : The authors thank all participants of this study. This work was supported by a Dong-A University Research Fund. The authors also thank Editage (www.editage.co.kr) for English language editing. References Yin W, Zhu W, Gao H, Niu X, Shen C, Fan X, et al. Gait analysis in the early stage of Parkinson’s disease with a machine learning approach. Front Neurol. 2024;15:1472956. Mermelstein S, Barbosa P, Kaski D. Neurological gait assessment. Pract Neurol. 2024;24:11–21. Mirelman A, Bonato P, Camicioli R, Ellis TD, Giladi N, Hamilton JL, et al. Gait impairments in Parkinson’s disease. Lancet Neurol. 2019;18:697–708. Huo H, Zhang C, Liu W, Zhao C, Ma L, Wang J, et al. Early detection of Parkinson’s disease using a multi area graph convolutional network. Sci Rep. 2025;15:5561. Lin C-H, Wang F-C, Kuo T-Y, Huang P-W, Chen S-F, Fu L-C. Early detection of Parkinson’s disease by neural network models. IEEE Access. 2022;10:19033–44. Aboutorabi A, Arazpour M, Bahramizadeh M, Hutchins SW, Fadayevatan R. The effect of aging on gait parameters in able-bodied older subjects: a literature review. Aging Clin Exp Res. 2016;28:393–405. Mirelman A, Ben Or Frank M, Melamed M, Granovsky L, Nieuwboer A, Rochester L, et al. Detecting sensitive mobility features for Parkinson’s disease stages via machine learning. Mov Disord. 2021;36:2144–55. Rizzo G, Copetti M, Arcuti S, Martino D, Fontana A, Logroscino G. Accuracy of clinical diagnosis of Parkinson disease: a systematic review and meta-analysis. Neurology. 2016;86:566–76. Buckley C, Alcock L, McArdle R, Rehman RZU, Del Din S, Mazzà C, et al. The role of movement analysis in diagnosing and monitoring neurodegenerative conditions: insights from gait and postural control. Brain Sci. 2019;9:34. Daneault JF, Vergara-Diaz G, Parisi F, Admati C, Alfonso C, Bertoli M, et al. Accelerometer data collected with a minimum set of wearable sensors from subjects with Parkinson’s disease. Sci Data. 2021;8:48. Brognara L, Palumbo P, Grimm B, Palmerini L. Assessing gait in Parkinson’s disease using wearable motion sensors: a systematic review. Diseases. 2019;7:18. Juneau P, Baddour N, Burger H, Bavec A, Lemaire ED. Amputee fall risk classification using machine learning and smartphone sensor data from 2-minute and 6-minute walk tests. Sens (Basel). 2022;22:1749. Bohannon RW, Bubela D, Magasi S, McCreath H, Wang YC, Reuben D, et al. Comparison of walking performance over the first 2 minutes and the full 6 minutes of the six-minute walk test. BMC Res Notes. 2014;7:269. Valet M, Pierchon L, Lejeune T. The 2-min walk test could replace the 6-min walk test in ambulant persons with subacute or chronic stroke: a two-stage retrospective study. Int J Rehabil Res. 2023;46:41–5. Ghislieri M, Agostini V, Rizzi L, Fronda C, Knaflitz M, Lanotte M. Foot–floor contact sequences: a metric for gait assessment in Parkinson’s disease after deep brain stimulation. Sens (Basel). 2024;24:6593. Hadouiri N, Monnet E, Gouelle A, Sagawa Y Jr, Decavel P. Locomotor strategy to perform 6-minute walk test in people with multiple sclerosis: a prospective observational study. Sens (Basel). 2023;23:3407. El Maachi I, Bilodeau GA, Bouachir WD. Deep 1D-Convnet for accurate Parkinson disease detection and severity prediction from gait. Expert Syst Appl. 2020;143:113075. Yang X, Ye Q, Cai G, Wang Y, Cai G. PD-ResNet for classification of Parkinson’s disease from gait. IEEE J Transl Eng Health Med. 2022;10:2200111. Pedrero-Sánchez JF, Belda-Lois JM, Serra-Añó P, Inglés M, López-Pascual J. Classification of healthy, Alzheimer and Parkinson populations with a multi-branch neural network. Biomed Signal Process Control. 2022;75:103617. Choi H, Youm C, Park H, Kim B, Hwang J, Cheon SM, et al. Convolutional neural network based detection of early stage Parkinson’s disease using the six minute walk test. Sci Rep. 2024;14:22648. Gloeckl R, Teschler S, Jarosch I, Christle JW, Hitzl W, Kenn K. Comparison of two- and six-minute walk tests in detecting oxygen desaturation in patients with severe chronic obstructive pulmonary disease—a randomized crossover trial. Chron Respir Dis. 2016;13:256–63. El-Gohary M, Pearson S, McNames J, Mancini M, Horak F, Mellone S, et al. Continuous monitoring of turning in patients with movement disability. Sens (Basel). 2013;14:356–69. Eckmann JP, Kamphorst SO, Ruelle D. Recurrence plots of dynamical systems. Europhys Lett. 1987;4:973–7. Lemaître G, Nogueira F, Aridas CK. Imbalanced-learn: a python toolbox to tackle the curse of imbalanced datasets in machine learning. J Mach Learn Res. 2017;18:1–5. Tank VH, Ghosh R, Gupta V, Sheth N, Gordon S, He W, et al. Drug eluting stents versus bare metal stents for the treatment of extracranial vertebral artery disease: a meta-analysis. J Neurointerv Surg. 2016;8:770–4. Huang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely connected convolutional networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE; 2017. pp. 2261-9. Iandola FN, Han S, Moskewicz MW, Ashraf K, Dally WJ, Keutzer K. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and < 0.5 MB model size. arXiv. 2016. 10.48550/arXiv.1602.07360 . Bernardo LS, Damaševičius R, Ling SH, de Albuquerque VHC, Tavares JMR. Modified SqueezeNet architecture for Parkinson’s disease detection based on keypress data. Biomedicines. 2022;10:2746. Uchitomi H, Ming X, Zhao C, Ogata T, Miyake Y. Classification of mild Parkinson’s disease: data augmentation of time-series gait data obtained via inertial measurement units. Sci Rep. 2023;13:12638. Rehman RZU, Buckley C, Mico-Amigo ME, Kirk C, Dunne-Willows M, Mazza C, et al. Accelerometry-based digital gait characteristics for classification of Parkinson’s disease: what counts? IEEE Open J Eng Med Biol. 2020;1:65–73. Mancini M, Afshari M, Almeida Q, Amundsen-Huffmaster S, Balfany K, Camicioli R, et al. Digital gait biomarkers in Parkinson’s disease: susceptibility/risk, progression, response to exercise, and prognosis. npj Parkinsons Dis. 2025;11:51. Turcato AM, Godi M, Giardini M, Arcolin I, Nardone A, Giordano A, et al. Abnormal gait pattern emerges during curved trajectories in high-functioning Parkinsonian patients walking in line at normal speed. PLoS ONE. 2018;13:e0197264. Song J, Sigward S, Fisher B, Salem GJ. Altered dynamic postural control during step turning in persons with early-stage Parkinson’s disease. Parkinsons Dis. 2012;2012:386962. Meng L, Pang J, Yang Y, Chen L, Xu R, Ming D. Inertial-based gait metrics during turning improve the detection of early-stage Parkinson’s disease patients. IEEE Trans Neural Syst Rehabil Eng. 2023;31:1472–82. Li X, Huang X, Pang J, Meng L, Ming D. A convolutional neural network based classification method for mild to moderate Parkinson’s disease at turns. In: Asian Pacific Conference on Medical and Biological Engineering; 2024. pp. 371–8. Atrsaei A, Corrà MF, Dadashi F, Vila-Chã N, Maia L, Mariani B, et al. Gait speed in clinical and daily living assessments in Parkinson’s disease patients: performance versus capacity. npj Parkinsons Dis. 2021;7:24. Pedrero-Sánchez JF, Belda-Lois JM, Serra-Añó P, Mollà-Casanova S, López-Pascual J. Classification of Parkinson’s disease stages with a two-stage deep neural network. Front Aging Neurosci. 2023;15:1152917. Mancini M, El-Gohary M, Pearson S, McNames J, Schlueter H, Nutt JG, et al. Continuous monitoring of turning in Parkinson’s disease: rehabilitation potential. NeuroRehabilitation. 2015;37:3–10. https://doi.org/10.3233/NRE-151236 . Warmerdam E, Hausdorff JM, Atrsaei A, Zhou Y, Mirelman A, Aminian K, et al. Long-term unsupervised mobility assessment in movement disorders. Lancet Neurol. 2020;19:462–70. Navarro-López V, Fernández-Vázquez D, Molina-Rueda F, Cuesta-Gómez A, García-Prados P, Del-Valle-Gratacós M, et al. Arm-swing kinematics in Parkinson’s disease: a systematic review and meta-analysis. Gait Posture. 2022;98:85–95. Ferraris C, Amprimo G, Masi G, Vismara L, Cremascoli R, Sinagra S, et al. Evaluation of arm swing features and asymmetry during gait in Parkinson’s disease using the azure kinect sensor. Sens (Basel). 2022;22:6282. Barrett MJ, Wylie SA, Harrison MB, Wooten GF. Handedness and motor symptom asymmetry in Parkinson’s disease. J Neurol Neurosurg Psychiatry. 2011;82:1122–4. Yust-Katz S, Tesler D, Treves TA, Melamed E, Djaldetti R. Handedness as a predictor of side of onset of Parkinson’s disease. Parkinsonism Relat Disord. 2008;14:633–5. Van Rooden SM, Visser M, Verbaan D, Marinus J. Handedness associated to side of onset of Parkinson’s disease? Parkinsonism Relat Disord. 2009;15:546–7. Park H, Youm C, Lee M, Noh B, Cheon SM. Turning characteristics of the more-affected side in Parkinson’s disease patients with freezing of gait. Sens (Basel). 2020;20:3098. Park H, Shin S, Youm C, Cheon SM, Lee M, Noh B. Classification of Parkinson’s disease with freezing of gait based on 360° turning analysis using 36 kinematic features. J Neuroeng Rehabil. 2021;18:177. Vanmechelen I, Bekteshi S, Haberfehlner H, Feys H, Desloovere K, Aerts JM, et al. Reliability and discriminative validity of wearable sensors for the quantification of upper limb movement disorders in individuals with dyskinetic cerebral palsy. Sens (Basel). 2023;23:1574. Pietrosanti L, Saggio G, Patera M, Suppa A, Giannini F, Verrelli CM. Modeling of shoulder–elbow movement with exponential parameter identification during walking gaits for healthy subjects and patients with Parkinson’s disease. Appl Sci. 2025;15:668. Del Din S, Godfrey A, Rochester L. Validation of an accelerometer to quantify a comprehensive battery of gait characteristics in healthy older adults and Parkinson’s disease: toward clinical and at home use. IEEE J Biomed Health Inf. 2016;20:838–47. Kim DS, Schuetz N, Johnson A, Tolas A, Mantena S, O’Sullivan JW, et al. Unlocking insights: clinical associations from the largest 6-minute walk test collection via the my Heart Counts cardiovascular Health Study, a fully digital smartphone platform. Prog Cardiovasc Dis. 2025;89:45–52. Galán-Mercant A, Ortiz A, Herrera-Viedma E, Tomas MT, Fernandes B, Moral-Munoz JA. Assessing physical activity and functional fitness level using convolutional neural networks. Knowl Based Syst. 2019;185:104939. Pappas MC, Baudendistel ST, Schmitt AC, Au KLK, Hass CJ. Acclimatization of force production during walking in persons with Parkinson’s disease. J Biomech. 2023;148:111477. Kosak M, Smith T. Comparison of the 2-, 6-, and 12-minute walk tests in patients with stroke. J Rehabil Res Dev. 2005;42:103–7. Wu X, Ma L, Wei P, Shan Y, Chan P, Wang K, et al. Wearable sensor devices can automatically identify the ON-OFF status of patients with Parkinson’s disease through an interpretable machine learning model. Front Neurol. 2024;15:1387477. Plotnik M, Wagner JM, Adusumilli G, Gottlieb A, Naismith RT. Gait asymmetry, and bilateral coordination of gait during a six-minute walk test in persons with multiple sclerosis. Sci Rep. 2020;10:12382. Additional File Additional file 3 is not available with this version. Additional Declarations No competing interests reported. Supplementary Files Additionalfile1.SupplementaryTable1.docx Additionalfile2.SupplementaryTable2.docx Additionalfile4.Researchdatarawdata.xlsx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7285962","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":498295338,"identity":"b0b91756-aa86-4ef4-9781-5d37891efce3","order_by":0,"name":"Hyejin Choi","email":"","orcid":"","institution":"The Graduate School of Dong-A University","correspondingAuthor":false,"prefix":"","firstName":"Hyejin","middleName":"","lastName":"Choi","suffix":""},{"id":498295339,"identity":"0f916ae2-eb0c-4f1f-add7-b75b3f0f7300","order_by":1,"name":"Changhong Youm","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAt0lEQVRIiWNgGAWjYDACCTBpw9gA4SYQrSWNdC2HSdDCP7vH8DFvznnZ/hkJjB9+MKTlE7bkzhljY95tt41n3EhgluxhyLFsIKTFQCLHTBqoJbHhRgKDNANDhQFBW6BaziXOB9rymxQtBxI33EhgA9qSQ1iLxI20YsO525KNN5552GbZY5BGWAv/jOSND95us5Oddzz58I0fFcmEtTAwcMAUgaKGGA0MDOwPiFI2CkbBKBgFIxgAAMSZOXu0+WMMAAAAAElFTkSuQmCC","orcid":"","institution":"The Graduate School of Dong-A University","correspondingAuthor":true,"prefix":"","firstName":"Changhong","middleName":"","lastName":"Youm","suffix":""},{"id":498295341,"identity":"3248253c-4b94-4397-ab86-b9884e4bdac6","order_by":2,"name":"Hwayoung Park","email":"","orcid":"","institution":"Dong-A University","correspondingAuthor":false,"prefix":"","firstName":"Hwayoung","middleName":"","lastName":"Park","suffix":""},{"id":498295343,"identity":"a2dcec3e-c3bf-4b10-8a9d-82c9b6eb2b58","order_by":3,"name":"Bohyun Kim","email":"","orcid":"","institution":"Dong-A University","correspondingAuthor":false,"prefix":"","firstName":"Bohyun","middleName":"","lastName":"Kim","suffix":""},{"id":498295346,"identity":"5f3e13ca-3c9c-46ca-b094-53e8079b3ca4","order_by":4,"name":"Juseon Hwang","email":"","orcid":"","institution":"The Graduate School of Dong-A University","correspondingAuthor":false,"prefix":"","firstName":"Juseon","middleName":"","lastName":"Hwang","suffix":""},{"id":498295348,"identity":"4c723001-351e-4deb-8a72-aec6bd8c3e1b","order_by":5,"name":"Sang-Myung Cheon","email":"","orcid":"","institution":"Dong-A University","correspondingAuthor":false,"prefix":"","firstName":"Sang-Myung","middleName":"","lastName":"Cheon","suffix":""}],"badges":[],"createdAt":"2025-08-04 01:38:19","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7285962/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7285962/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":88890556,"identity":"133297ca-a413-4210-81f3-764bd949b4f5","added_by":"auto","created_at":"2025-08-12 12:46:19","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":218077,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy flow diagram depicting participant recruitment and analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviation\u003c/strong\u003e: PD: Parkinson’s disease.\u003c/p\u003e","description":"","filename":"fig1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/1b324569015df1c349ab1117.jpg"},{"id":88890559,"identity":"585d5587-a8e4-4eec-8241-798b04ecf742","added_by":"auto","created_at":"2025-08-12 12:46:19","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":389174,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of the methodology and analytical framework. (a) The 6-min walk test. The main task of walking back and forth on a 20-m course for 6 min. Placement of wearable sensors. The participant wore wearable sensors attached to the upper arms (5 cm above the lateral humeral epicondyle) and thighs (10 cm above the lateral femoral epicondyle) of both sides, thoracic spine (T10), and lumbar spine (center of the left and right posterior superior iliac spines). (b) Framework for classifying people with early-stage PD and controls based on wearable sensors.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations\u003c/strong\u003e: PD: Parkinson’s disease; Cons: controls; CNN: convolutional neural network; ResNet: residual neural network; DenseNet: dense convolutional network; RM ANOVA: repeated-measures analysis of variance.\u003c/p\u003e","description":"","filename":"fig2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/f48302b7fe3765c1d8728301.jpg"},{"id":88890567,"identity":"ed940fd0-e9e3-4907-abc8-4563cb19ea21","added_by":"auto","created_at":"2025-08-12 12:46:19","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":485637,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eClassification accuracy of all variables across sensor locations during straight gait in people with early-stage PD and controls\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe figure presents the highest classification accuracy among the three CNN models (ResNet, DenseNet, and SqueezeNet) for each accelerometer and gyroscope variable across all sensor locations during the straight gait phase.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations\u003c/strong\u003e: PD: Parkinson’s disease; CNN: convolutional neural network; ResNet: residual neural network; DenseNet: dense convolutional network; Acc: accelerometer; Gyr: gyroscope; XYZ: resultant values of accelerometer and gyroscope data; Larm: left arm; Rarm: right arm; Lthi: left thigh; Rthi: right thigh; Thora: thoracic spine; Lumbar: lumbar spine.\u003c/p\u003e","description":"","filename":"fig3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/4d89ceddfc22948058046c52.jpg"},{"id":88890569,"identity":"33953413-2eb7-4e3f-b632-e8c83dc5578a","added_by":"auto","created_at":"2025-08-12 12:46:19","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":472125,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eClassification accuracy of all variables across sensor locations during turning gait in people with early-stage PD and controls\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe figure presents the highest classification accuracy among the three CNN models (ResNet, DenseNet, and SqueezeNet) for each accelerometer and gyroscope variable across all sensor locations during the turning gait phase.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations\u003c/strong\u003e: PD: Parkinson’s disease; CNN: convolutional neural network; ResNet: residual neural network; DenseNet: dense convolutional network; Acc: accelerometer; Gyr: gyroscope; XYZ: resultant values of accelerometer and gyroscope data; Larm: left arm; Rarm: right arm; Lthi: left thigh; Rthi: right thigh; Thora: thoracic spine; Lumbar: lumbar spine.\u003c/p\u003e","description":"","filename":"fig4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/48919d3ee9cf1411d3ce680a.jpg"},{"id":88894425,"identity":"fb07f8f8-3d9a-4c5a-9e99-8c05672b2c72","added_by":"auto","created_at":"2025-08-12 13:02:19","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":313132,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConfusion matrices for top-performing variables (accuracy \u0026gt; 90%) during straight gait at 6 min\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe confusion matrices represent the classification results for the top-performing variables based on the highest classification accuracy among the three CNN models (ResNet, DenseNet, and SqueezeNet).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations\u003c/strong\u003e: Larm: left arm; Rarm: right arm; Lthi: left thigh; Rthi: right thigh; Acc: accelerometer; Gyr: gyroscope; XYZ: resultant values of accelerometer and gyroscope data; ResNet: residual neural network; DenseNet: dense convolutional network; PD: Parkinson’s disease; Cons: controls.\u003c/p\u003e","description":"","filename":"fig5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/a83e018a89002d9c633081c1.jpg"},{"id":88896919,"identity":"a932da17-a2bf-4f8c-9172-193aab158b0d","added_by":"auto","created_at":"2025-08-12 13:10:19","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":81228,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eConfusion matrices for top-performing variables (accuracy \u0026gt; 80%) during turning gait at 6 min\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe confusion matrices represent the classification results for the top-performing variables based on the highest classification accuracy among the three CNN models (ResNet, DenseNet, and SqueezeNet).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations\u003c/strong\u003e: Lthi: left thigh; Rthi: right thigh; Acc: accelerometer; Gyr: gyroscope; ResNet: residual neural network; DenseNet: dense convolutional network; PD: Parkinson’s disease; Cons: controls.\u003c/p\u003e","description":"","filename":"fig6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/7cfe5309a1023a40fbb1df27.jpg"},{"id":88896921,"identity":"50045dde-0ed8-48a1-af4e-bca1f56b2439","added_by":"auto","created_at":"2025-08-12 13:10:19","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":168878,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExample of representative images and corresponding frequency patterns during the straight gait phase in people with early-stage PD and controls\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis figure presents representative image-based inputs and their associated frequency-domain plots derived from the gyroscope x-axis data of the left arm, which was identified as the highest-performing variable among all body segments. Data are segmented by minute (from 1 to 6 min) during the SMWT, focusing on the straight gait phase. Each row corresponds to a subgroup defined by sex (male or female) and clinical status (Controls, H\u0026amp;Y stage 1, and H\u0026amp;Y stage 2), illustrating temporal changes in signal morphology and frequency characteristics over time.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations\u003c/strong\u003e: PD: Parkinson’s disease; SMWT: 6-min walk test; H\u0026amp;Y: Hoehn and Yahr.\u003c/p\u003e","description":"","filename":"fig7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/3abe6daf5a48d73251ffb49b.jpg"},{"id":88892864,"identity":"b6acca8c-87ab-4b80-9ea8-2673a99f529c","added_by":"auto","created_at":"2025-08-12 12:54:19","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":159222,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExample of representative images and corresponding frequency patterns during the turning gait phase in people with early-stage PD and controls\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis figure presents representative image-based inputs and their associated frequency-domain plots derived from the acceleration y-axis data of the right thigh, which was identified as the highest-performing variable among all body segments. Data are segmented by minute (from 1 to 6 min) during the SMWT, focusing on the turning gait phase. Each row corresponds to a subgroup defined by sex (male or female) and clinical status (Controls, H\u0026amp;Y stage 1, and H\u0026amp;Y stage 2), illustrating temporal changes in signal morphology and frequency characteristics over time.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAbbreviations\u003c/strong\u003e: PD: Parkinson’s disease; SMWT: 6-min walk test; H\u0026amp;Y: Hoehn and Yahr.\u003c/p\u003e","description":"","filename":"fig8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/86e4da3291117f2fc7d7f8a7.jpg"},{"id":91480665,"identity":"604ff340-a695-421e-a217-01e272d7a45d","added_by":"auto","created_at":"2025-09-17 03:01:55","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3592901,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/41937d20-3f1a-423a-931b-1868e986265b.pdf"},{"id":88890557,"identity":"64023665-cd53-46ab-9243-9c7d400e5233","added_by":"auto","created_at":"2025-08-12 12:46:19","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":19951,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile1.SupplementaryTable1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/795872ff30a6e893df8da5fb.docx"},{"id":88894419,"identity":"74cd3401-0908-44e1-a9cf-f7eee3a843ba","added_by":"auto","created_at":"2025-08-12 13:02:19","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":44179,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile2.SupplementaryTable2.docx","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/cf888c5b00d314d7b9101f16.docx"},{"id":88890562,"identity":"f9c75c41-3537-4302-8e0f-9e4b51ecab6e","added_by":"auto","created_at":"2025-08-12 12:46:19","extension":"xlsx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":19405,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile4.Researchdatarawdata.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7285962/v1/6a3656dea338acb276cd6b50.xlsx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Early Detection of Parkinson’s Disease Using a Single-Arm Wearable Sensor and Convolutional Neural Networks","fulltext":[{"header":"BACKGROUND","content":"\u003cp\u003eParkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by motor symptoms, including bradykinesia, rigidity, resting tremor, and postural instability [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Gait disturbances are particularly disabling and often emerge early, manifesting as reduced step length, decreased walking speed, festination, and impaired turning [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Subtle asymmetries, such as unilateral shuffling and diminished arm swing, frequently appear in early-stage PD but may go undetected without overt gait complaints [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], especially since some features overlap with normal aging [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eCurrent clinical assessments, including the Hoehn and Yahr (H\u0026amp;Y) Scale and the Unified Parkinson’s Disease Rating Scale, rely on clinician observations and are subject to inter-rater variability and subjectivity [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. Although laboratory-based systems such as three-dimensional motion capture and instrumented walkways provide precise gait measurements, their high cost and operational complexity limit their routine clinical use [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. In contrast, wearable sensors offer a portable, low-cost, and scalable solution for continuous, real-world gait monitoring [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. As a result, there is growing interest in leveraging wearable sensor data to develop objective tools for PD detection and severity assessment [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe 6-min walk test (SMWT) is widely used to assess functional endurance under semi-natural conditions. It captures both the straight and turning gait phases, enabling a more ecologically valid evaluation of mobility [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. While previous studies have primarily focused on total walking distance as the main outcome [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], phase-specific gait dynamics—particularly during turning—may better reveal early motor impairments. Turning requires more complex motor coordination and often exhibits greater variability, making it a sensitive indicator of dysfunction [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eDespite recommendations to use segment-based analysis, few studies have examined temporal variations in gait quality during the SMWT. Segment-based analysis involves dividing the walking test into smaller, distinct time intervals to evaluate specific gait phases, such as straight walking and turning, separately. For example, Bohannon et al. and Valet et al. reported that performance during the first 2 min approximates full-test outcomes, although their analyses focused exclusively on walking distance [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Hadouiri et al. extended this approach in patients with multiple sclerosis by analyzing gait data minute by minute, revealing time-dependent changes in spatiotemporal metrics [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. However, no comparable analysis has been conducted for PD using deep learning techniques. Moreover, existing studies have primarily focused on lower-limb data, often overlooking upper-limb and trunk movements.\u003c/p\u003e\u003cp\u003eDeep learning methods, particularly convolutional neural networks (CNNs), have shown promise in detecting PD from wearable sensor data, outperforming traditional machine learning techniques in capturing non-linear and phase-dependent gait features [\u003cspan additionalcitationids=\"CR18\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e–\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Our previous work demonstrated that CNNs trained on full-duration SMWT data could effectively distinguish patients with early-stage PD from healthy controls [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. However, CNN performance across specific gait phases and short time intervals has not been systematically investigated. Determining whether early-stage PD can be accurately classified within a shorter timeframe, such as the first or last minute, could streamline assessments and reduce participant burden [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTherefore, this study aimed to: (1) evaluate CNN-based classification of individuals with early-stage PD using multi-segment wearable sensor data collected during straight and turning gait phases of the SMWT; and (2) compare classification performance across different time intervals and sensor locations to determine the minimal reliable measurement duration and optimal sensor placement, with particular focus on the sensor previously shown to yield the highest classification accuracy.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cp\u003e\u003cb\u003eParticipants\u003c/b\u003e\u003c/p\u003e\u003cp\u003eOverall, 78 individuals with early-stage PD and 50 age-matched controls were enrolled. All participants were diagnosed with idiopathic PD according to the United Kingdom Parkinson's Disease Society Brain Bank criteria, as confirmed by a neurologist. Inclusion criteria were H\u0026amp;Y stages 1–2 and Mini-Mental State Examination scores ≥ 24. Individuals with comorbid musculoskeletal, cardiovascular, or cognitive conditions, or those with recent orthopedic surgery, were excluded. Only participants who completed the full SMWT were included in the final analysis. The recruitment flow diagram is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, and demographic data are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The study protocol was approved by the Institutional Review Board of Dong-A University Medical Center (IRB number: DAUHIRB-22-089) and conducted in accordance with relevant guidelines and regulations. All participants provided written informed consent. The study was registered with the Clinical Research Information Service of Korea (KCT0009353).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cdiv class=\"gridtable\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePhysical and clinical characteristics of participants\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eEarly PDs (n = 78)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eControls (n = 50)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cem\u003ep\u003c/em\u003e-value\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSex (male/female)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e37 / 41\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e21 / 29\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.588\u003csup\u003ea\u003c/sup\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAge (years)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e67.51 ± 7.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e65.80 ± 5.51\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.149\u003csup\u003eb\u003c/sup\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHeight (cm)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e161.20 ± 7.72\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e161.14 ± 8.38\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.907\u003csup\u003eb\u003c/sup\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBody mass (kg)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e63.96 ± 10.54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e63.68 ± 10.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.884\u003csup\u003eb\u003c/sup\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBMI (kg/m\u003csup\u003e2\u003c/sup\u003e)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e24.51 ± 3.13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e24.40 ± 2.43\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.847\u003csup\u003eb\u003c/sup\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eDisease duration (years)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e5.50 ± 3.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTreatment duration (years)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e4.19 ± 3.25\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eL-Dopa equivalent dose (mg/day)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e541.90 ± 286.54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMMSE (scores)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e28.12 ± 1.87\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e27.22 ± 1.90\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.003\u003csup\u003ec\u003c/sup\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMoCA (scores)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e26.23 ± 2.88\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e24.86 ± 2.99\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.001\u003csup\u003ec\u003c/sup\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eUPDRS Total (scores)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e47.68 ± 19.54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eUPDRS III (scores)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e25.35 ± 13.63\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eH\u0026amp;Y Scale (stages 1 and 2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e28 / 50\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSMWT (m)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e414.24 ± 85.16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e494.23 ± 51.42\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt; 0.001\u003csup\u003ec\u003c/sup\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e\u003cp\u003eThe data are presented as the mean ± standard deviation. Significant difference: \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05; PD: Parkinson’s disease; BMI: Body mass index; L-Dopa: Levodopa; MMSE: Mini-Mental State Examination; MoCA: Montreal Cognitive Assessment; UPDRS: Unified Parkinson’s Disease Rating Scale; H\u0026amp;Y: Hoehn and Yahr; SMWT: 6-min walk test.\u003c/p\u003e\u003cp\u003e\u003csup\u003ea\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e­value of Fisher’s exact test.\u003c/p\u003e\u003cp\u003e\u003csup\u003eb\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e­value of the independent t­test.\u003c/p\u003e\u003cp\u003e\u003csup\u003ec\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e­value of the Mann–Whitney U test.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003c/div\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eExperimental Procedures\u003c/b\u003e\u003c/p\u003e\u003cp\u003eAll participants completed the SMWT along a 20-m hallway delineated by cones placed at each end (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). The task was performed under standardized instructions: walk as fast as possible without running, with safety monitoring by study personnel. All tests were conducted approximately 2 h after medication intake during the \"ON\" medication state.\u003c/p\u003e\u003cp\u003eSix Xsens DOT sensors (Movella Technologies, Enschede, Netherlands) were affixed using a stretchable belt at the following locations: the left and right upper arms (5 cm above the lateral humeral epicondyle), left and right thighs (10 cm above the lateral femoral epicondyle), thoracic spine (T10), and lumbar spine (aligned with the midpoint of the posterior superior iliac spine) (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). Sensor dimensions were 36.3 × 30.35 × 10.8 mm, and weight was 11.2 g. All sensors sampled tri-axial acceleration and gyroscopic data at 60 Hz, using the East-North-Up coordinate system. Raw data were transmitted via Bluetooth 5.0 to an iPad (iOS 15.6.1; Apple, Cupertino, CA) using the MovellaDOT application. All subsequent analyses were performed using MATLAB R2023a (MathWorks, Natick, MA). Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb illustrates the overall framework adopted in this study for classifying people with early-stage PD and controls using wearable sensor data. The process encompasses four primary stages: collecting raw TS data via wearable sensors, applying preprocessing techniques to prepare the data, splitting the dataset into training and testing subsets, and finally, training a CNN model to assess classification performance.\u003c/p\u003e\u003cp\u003e\u003cb\u003eData Pre-processing and Gait Phase Detection\u003c/b\u003e\u003c/p\u003e\u003cp\u003eRaw TS data from all six sensors were segmented into six 1-min intervals to facilitate time-dependent analysis. Within each 1-min interval, straight and turning gait phases were detected using lumbar gyroscope data, following the approach of El-Gohary et al. [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e] Turning gait was identified when angular velocity exceeded 15°/s, and the end of turning was marked when it fell below 5°/s. To exclude transient movements or artifacts, a minimum turning duration of 0.5 s (30 frames) was enforced based on gait cycle dynamics and prior literature. Straight segments were retained only if they lasted for ≥ 5 s (300 frames). Up to three valid segments were extracted per phase (straight or turning) per minute. The identified frame indices were synchronized across all sensors to preserve temporal alignment. Data pre-processing was implemented using MATLAB and Python (Python 3.10; Python Software Foundation).\u003c/p\u003e\u003cp\u003e\u003cb\u003eNormalization, Filtering, and Resampling\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo mitigate sensor placement variability and anthropometric differences, all signals were normalized by subtracting the median value per axis (X, Y, and Z). A second-order low-pass Butterworth filter with a 6-Hz cutoff was applied to reduce high-frequency noise.\u003c/p\u003e\u003cp\u003eSubsequently, the segmented gait data were resampled using dynamic time-warping-based interpolation to standardize the input length: 896 and 448 frames for straight and turning segments, respectively. Segments that did not meet the required length were zero-padded. To satisfy the CNN input constraints, all final inputs were adjusted to 896 frames (to match a 224 × 224 pixel resolution, using a width factor of four). Image representations were generated using recurrence plots [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], which were applied to both the individual axes and the resultant values derived from the accelerometer and gyroscope data. For each of the 1st to 6th minutes, three straight and three turning gait phases were extracted per minute from six sensor locations. This process yielded eight TS image types per segment based on the acceleration and angular velocity data across all gait phases.\u003c/p\u003e\u003cp\u003e\u003cb\u003eData Generation\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo train and evaluate the CNN models, the accumulated imaging dataset was analyzed incrementally by minutes. Specifically, datasets from 1 min were analyzed first, followed by cumulative datasets from 1 to 2 min, 1–3 min, and so on up to 6 min. For each dataset, training, validation, and test sets were randomly split by subject in ratios of 50%, 20%, and 30%, respectively. As sample counts varied across minutes (Additional file 1), a representative 1-min example illustrates the oversampling procedure. That dataset initially comprised 234 early-stage PD samples and 150 control samples per segment across six sensors (left arm, right arm, left thigh, right thigh, thoracic spine, and lumbar spine). To address this imbalance, data were randomly oversampled using the imbalanced-learn (version 0.10.1) Python package by generating new minority-class samples to achieve parity [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Consequently, 234 samples per class (468 total) were generated per sensor segment for the 1-min dataset. The same balancing strategy was applied to all other minute-level datasets.\u003c/p\u003e\u003cp\u003e\u003cb\u003eCNN Model Training\u003c/b\u003e\u003c/p\u003e\u003cp\u003eModel training and classification were conducted for each accumulated minute using the processed TS image dataset. For variance analysis, a five-fold cross-validation was applied. The batch size used during training was 128. Adam optimizers were employed, with a learning rate of 1e-05. After 150 epochs, the proposed model converged, and training was halted using early stopping. To reduce the risk of overfitting, only the best model was retained. This implies that during the training phase, the model was saved if the validation accuracy of a given epoch surpassed the highest accuracy achieved thus far.\u003c/p\u003e\u003cp\u003eTo classify people with early-stage PD and controls based on the processed TS images, the following three CNN architectures were employed: ResNet [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], DenseNet [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], and SqueezeNet [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. The input image size for all CNN models was set to 224 × 224 pixels. We applied the ResNet model comprising 18 layers. The DenseNet-121 model, composed of four dense blocks with 6, 12, 24, and 16 channels, was also adopted. Similarly, SqueezeNet v1.0 was employed. These architectures were selected because of their proven performance and efficiency in image-based classification.\u003c/p\u003e\u003cp\u003eThe performance of the binary classification models was evaluated using standard metrics, including accuracy, precision, recall, and F1-score. Accuracy was used as the primary evaluation metric for classification problems that are well-balanced and not skewed or affected by class imbalance. In addition, confusion matrix components—true positive, true negative, false positive, and false negative—were computed to derive the precision and recall values. Precision indicates the proportion of correct positive predictions, whereas recall quantifies the number of actual positive cases accurately identified by the classifier. This metric is also known as sensitivity. The F1-score integrates both precision and recall and serves as a comprehensive metric, commonly defined as the harmonic mean of these two components. The harmonic mean is considered more appropriate than the arithmetic mean for averaging ratios such as precision recall [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e].\u003c/p\u003e\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\u003cp\u003eThe Shapiro–Wilk test assessed multivariate normality. Fisher’s exact test, the Mann–Whitney U test, or the independent t-test evaluated differences in physical and clinical characteristics between people with early-stage PD and controls. Statistical analyses were performed using IBM SPSS Statistics for Windows, version 21.0 (IBM Corp., Armonk, NY), MATLAB, and Python. Statistical significance was set at \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05.\u003c/p\u003e\u003cp\u003eRepeated-measures analysis of variance (ANOVA) assessed the effect of measurement duration (1–6 min) on classification accuracy across top-performing sensor-variable combinations. Bonferroni-adjusted post hoc t-tests compared specific time intervals.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eCNN Classification Performance Across Gait Phases\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe classification accuracy from 1 to 6 min was evaluated for each sensor using time-series (TS) images processed with three CNN architectures: residual neural network (ResNet), dense convolutional network (DenseNet), and SqueezeNet. The highest accuracy from any model for each accelerometer and gyroscope variable was analyzed across all sensors for both straight (Fig. 3) and turning (Fig. 4) gait phases, where Gyr and Acc denote gyroscope and accelerometer data, respectively. X, Y, and Z represent directional axes. XYZ refers to the resultant magnitude of all three axes.\u003c/p\u003e\n\u003cp\u003eAccuracy varied significantly between gait phases. As shown in Fig. 3, heat-map accuracies were consistently higher during straight gait across all sensor-variable combinations and time points. The left arm gyroscope X-axis data (Larm_Gyr_X) during straight gait yielded the highest accuracy\u0026mdash;95.6% (precision, recall, F1-score = 0.956)\u0026mdash;at 1 min, surpassing the full 6-min accuracy of 94.6% (precision, recall, F1-score = 0.946), suggesting that short-duration straight gait effectively captures early-stage PD features.\u003c/p\u003e\n\u003cp\u003eIn addition to the left arm, the sensors on the right arm (Rarm) and both thighs (Lthi and Rthi) showed strong performance during straight gait, with several variables achieving \u0026gt;90% accuracy at 6 min. Notable examples included Rarm_Gyr_Z (accuracy: 0.907; precision: 0.908; recall: 0.907; F1-score: 0.907), Lthi_Acc_X (accuracy: 0.905; precision: 0.906; recall: 0.905; F1-score: 0.905), Lthi_Acc_Y (accuracy: 0.919; precision: 0.919; recall: 0.919; F1-score: 0.919), Lthi_Gyr_X (accuracy: 0.903; precision: 0.903; recall: 0.903; F1-score: 0.903), Lthi_Gyr_Y (accuracy: 0.901; precision: 0.901; recall: 0.901; F1-score: 0.901), Lthi_Gyr_Z (accuracy: 0.901; precision: 0.902; recall: 0.901; F1-score: 0.901), Lthi_Acc_XYZ (accuracy: 0.909; precision: 0.910; recall: 0.909; F1-score: 0.909), and Rthi_Gyr_Y (accuracy: 0.902; precision: 0.903; recall: 0.902; F1-score: 0.902). By contrast, classification accuracy during turning gait was lower and more variable (Fig. 4). The top-performing turning variable was Rthi_Acc_Y, reaching 80.9% accuracy (precision: 0.809; recall: 0.809; F1-score: 0.809). Other leading variables included Rthi_Gyr_Z (accuracy: 0.808; precision: 0.808; recall: 0.808; F1-score: 0.808), Lthi_Gyr_Z (accuracy: 0.807; precision: 0.807; recall: 0.807; F1-score: 0.806), and Lthi_Acc_X (accuracy: 0.800; precision: 0.800; recall: 0.800; F1-score: 0.800). Despite these results, a performance gap relative to straight gait was evident across all sensors.\u003c/p\u003e\n\u003cp\u003eTo further illustrate model performance, Figures 5 and 6 show confusion matrices for the highest-performing sensor\u0026ndash;variable combinations. Variables with \u0026gt;90% accuracy were selected for the straight gait phase and those exceeding 80% for the turning phase. The confusion matrices are based on the best results from ResNet, DenseNet, and SqueezeNet.\u003c/p\u003e\n\u003cp\u003eWe present representative images and corresponding frequency-domain patterns for individuals with people with early-stage PD and controls, focusing on the variables with the highest classification accuracy for each gait phase (i.e., straight and turning) at the most informative sensor location (Fig. 7 and Fig. 8). The representative images and graphs illustrate characteristic gait patterns from each minute of SMWT, capturing both straight walking and turning movements across time intervals from minute 1 to minute 6.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eDetermining the Optimal Measurement Duration\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGiven the superior performance of the straight gait, the analysis focused on identifying the optimal measurement duration and sensor location using straight-phase data. This approach prioritizes continuous gait patterns characteristic of early-stage PD.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eA. Analysis of Optimal Sensor and Variable Combinations\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eTo assess how measurement duration influenced classification, a repeated-measures ANOVA was performed using classification accuracy (averaged across the three CNN models) as the dependent variable. The top 10 sensor\u0026ndash;variable combinations were identified based on mean accuracy across six-time points. Eight of the top 10 variables were from the left arm (Larm), highlighting its consistent performance; the left thigh (Lthi) and lumbar spine (Lumba) ranked 9th and 10th, respectively. The top-ranked variables were: Larm_Gyr_X: 93.7% (F = 5.24, \u003cem\u003ep\u003c/em\u003e = 0.013); Larm_Gyr_Y: 93.0% (F = 1.56, \u003cem\u003ep\u003c/em\u003e = 0.257); Larm_Acc_XYZ: 92.3% (F = 33.36, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001); Larm_Gyr_Z: 92.2% (F = 28.88, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001); Larm_Acc_X: 92.1% (F = 16.29, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001); Larm_Gyr_XYZ: 91.1% (F = 17.12, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001); Larm_Acc_Y: 90.6% (F = 48.57, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001); Larm_Acc_Z: 90.4% (F = 0.33, \u003cem\u003ep\u003c/em\u003e = 0.886); Lthi_Acc_XYZ: 87.2% (F = 31.04, \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001); and Lumba_Acc_Y: 86.5% (F = 2.61, \u003cem\u003ep\u003c/em\u003e = 0.092). These variables demonstrated robust classification performance and statistical significance over time.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eB. Determining the Optimal Measurement Duration Based on Top Variables\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eTo identify the minimum duration required for stable classification, repeated-measures ANOVA and Bonferroni-corrected paired t-tests were conducted across 1\u0026ndash;6 min intervals for each top variable. If no significant difference emerged between 1 and 6 min, additional comparisons were made between 1 and 2 min.\u003c/p\u003e\n\u003cp\u003eFor Larm_Gyr_X, Gyr_Y, Acc_XYZ, Gyr_Z, Acc_X, Gyr_XYZ, and Acc_Z, no significant differences were observed between 1 and 6 min (adjusted \u003cem\u003ep\u003c/em\u003e \u0026ge; 0.579), or between 1 and 2 min (adjusted \u003cem\u003ep\u003c/em\u003e \u0026ge; 0.075), indicating that 1-min measurements suffice. Larm_Acc_Y required 2 min (1 vs. 2 min: adjusted\u003cem\u003e\u0026nbsp;p\u003c/em\u003e = 0.034; 2 vs. 6 min: adjusted \u003cem\u003ep\u003c/em\u003e = 1.000), and Lthi_Acc_XYZ likewise required\u0026nbsp;\u0026ge;\u0026nbsp;2 min (1 vs. 6 min: adjusted\u003cem\u003e\u0026nbsp;p\u003c/em\u003e = 0.048; 2 vs. 6 min: adjusted \u003cem\u003ep\u003c/em\u003e = 0.490). In contrast, Lumba_Acc_Y was stable even at 1 min (adjusted \u003cem\u003ep\u003c/em\u003e = 1.000). Results are presented in Additional file 2.\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eThis study extends our previous work [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e] by analyzing gait phase- and time-dependent classification of early-stage PD using wearable sensors and CNNs. While earlier studies focused on total gait metrics or walking distance [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e], our findings highlight the diagnostic value of segmenting gait\u0026mdash;particularly straight walking\u0026mdash;and demonstrate the feasibility of reducing measurement duration without compromising accuracy.\u003c/p\u003e\u003cp\u003eClassification accuracy was consistently higher during straight gait across all sensors, likely due to its repetitive, stable nature [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], which enabled CNNs to detect subtle motor deviations more effectively. The Larm_Gyr_X achieved the highest accuracy (95.6% at 1 min), highlighting the diagnostic importance of upper-limb motion [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. In contrast, turning gait introduced greater variability and individualized compensatory strategies [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e], which may have limited the CNN\u0026rsquo;s ability to learn consistent patterns [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. The top-performing turning feature, right-thigh mediolateral acceleration (Rthi_Acc_Y), achieved only 80.9% accuracy at 6 min. Li et al. reported higher classification accuracy during turning using CNNs and lower-limb and trunk data, their protocol excluded upper-limb sensors and employed short, self-paced walking [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. In contrast, our approach used sustained, fast-paced walking with full-body sensors, potentially capturing more consistent motor deviations in early-stage PD [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e]. Compared with our previous full-signal model without gait segmentation, which achieved 83.5% accuracy using lumbar gyroscope data [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], the present phase-specific framework significantly improved both classification performance and interpretability. Specifically, Larm_Gyr_X achieved 95.6% accuracy during 1-min straight gait, demonstrating the value of analyzing distinct gait phases to uncover task-specific features [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eSensor-wise, upper-limb sensors\u0026mdash;particularly the left and right arms\u0026mdash;outperformed trunk sensors during straight gait, frequently exceeding 90% accuracy. Thigh sensors were more predictive during turning, likely reflecting their role in pivoting and balance [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. Notably the lumbar spine sensor, which was previously optimal for non-segmented data, showed only moderate performance here, indicating that sensor relevance is highly phase-specific [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. The strong performance of the left arm sensor aligns with early-stage PD asymmetry in arm swing [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e] and supports context-specific sensor selection [\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe left-arm sensors consistently outperformed right-arm sensors during straight gait, possibly reflecting PD-related asymmetry [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. While PD often presents unilaterally [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e], this asymmetry does not always correspond to limb dominance [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e], suggesting individual variability [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. During turning, the slightly better performance of the right-thigh sensor may indicate lateralized pivoting strategies or turning preferences [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. These asymmetries should be considered both clinically and in model design [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eTo further characterize gait, we analyzed frequency-domain patterns for each minute of the SMWT, focusing on both straight and turning gait phases in individuals with early-stage PD and healthy controls (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e). For straight gait, the highest classification accuracy was observed in the Gyr_X component of the left arm, while for turning gait it was the Acc_Y component of the right thigh.\u003c/p\u003e\u003cp\u003eDifferences in frequency patterns between people with early-stage PD and controls likely influenced the generated images and contributed to classification accuracy [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. Future work should include additional non-linear analyses to extract discriminative features related to straight and turning gait patterns in the context of bilateral coordination, asymmetry, and variability [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. Furthermore, factors such as sex differences and symptom severity should be examined to better understand and quantify their impact. These features should then be analyzed together to clarify their contributions to classification performance.\u003c/p\u003e\u003cp\u003eAlthough the SMWT is a clinical standard, our findings suggest that 1 min of straight gait data may be sufficient for accurate early-stage PD classification using left arm sensors [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e]. Most top-performing features, including Larm_Gyr_X, showed no significant accuracy difference between 1 and 6 min, with peak performance at 1 min. These results support the feasibility of brief, task-specific gait assessments that reduce physical burden and simplify clinical workflows [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Previous studies also endorsed abbreviated testing: Bohannon et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e] reported strong correlations between 2-min and full SMWT distances (r\u0026thinsp;=\u0026thinsp;0.968), while Valet et al. [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] found that the 2-min distance accounted for 98% of total SMWT performance in patients with stroke. However, distance-based metrics may overlook subtle motor deficits. In contrast, our CNN-based time-series analysis captured finer spatiotemporal gait abnormalities relevant to early-stage PD [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eThe stability of the classification accuracy across time intervals indicates that properly segmented, short gait recordings can effectively detect motor impairments [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. This aligns with recent multiple sclerosis studies showing time-dependent gait changes even within short intervals [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Unlike prior studies focusing on stride length or gait speed, our phase-specific deep learning approach enabled granular analysis of motor features. Most top variables stabilized within 1\u0026ndash;2 min, supporting the feasibility of time-efficient screening protocols [\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e, \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eDespite these contributions, this study has several limitations. First, all data were collected in the medication \u0026ldquo;ON\u0026rdquo; state, which may have dampened symptom severity. Future research should include \u0026ldquo;ON\u0026rdquo; and \u0026ldquo;OFF\u0026rdquo; states to evaluate generalizability [\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e]. Second, although we observed significant group differences in MMSE and MoCA scores, larger and more demographically diverse samples are needed to improve generalizability. Validation in external cohorts with different regional and population characteristics is also warranted to ensure robustness. Third, this study focused only on early-stage PD (Hoehn and Yahr stages 1\u0026ndash;2). Including later stages could enhance understanding of disease progression and improve classification performance. Accounting for turning direction and limb dominance may further improve model accuracy, as these factors influence motor strategies and sensor-derived gait characteristics during the SMWT. Finally, while our CNN-based supervised learning approach demonstrated high classification accuracy, its limited interpretability remains a challenge. Future studies should integrate explainable AI (XAI) techniques and explore unsupervised or semi-supervised methods such as clustering or representation learning to uncover latent gait phenotypes and improve model interpretability [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. Applying this analytical framework to older adults with a history of falls may also enable early classification of prodromal or early-stage PD by detecting subtle gait abnormalities predictive of fall risk, while also offering a low-cost, scalable solution for quantitative gait assessment in various settings, including home-based monitoring [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. Longitudinal analysis of minute-by-minute non-linear feature trends could additionally provide insights into disease trajectory and progression [\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eIn conclusion, this study demonstrated that patients with early-stage PD can be accurately classified using deep learning models applied to phase-segmented gait data collected from wearable sensors. By separating straight and turning gait phases and analyzing time-dependent classification performance, we identified straight gait\u0026ndash;\u0026ndash;particularly within the first minute\u0026ndash;\u0026ndash;as the most diagnostically informative segment. The left arm gyroscope data (Larm_Gyr_X) consistently achieved the highest classification accuracy, highlighting the importance of upper-limb motion in early-stage PD detection. These findings support the development of time-efficient, phase-specific, and interpretable digital gait assessments for early PD screening. Future research should extend this framework to include diverse disease stages and medication states and explore unsupervised learning approaches to uncover latent motor phenotypes.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e: All study procedures involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. The study protocol and supplementary information files were approved by the Institutional Review Board of Dong-A University Hospital (approval number: DAUHIRB-22-089) (see ethics approval letter in Additional file 3). All patients provided written informed consent before data collection. The study was registered in the Clinical Research Information Service in the Republic of Korea (KCT0009353).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for\u0026nbsp;publication\u003c/strong\u003e: Not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e: The datasets supporting this study\u0026rsquo;s findings are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability\u003c/strong\u003e: We do not have an open-source code available. The codes for training and testing the deep learning models were written in Python 3.10 using PyTorch 1.9.1 and torchvision 0.10.1. Data management and feature-processing scripts were written in Python 3.10 using pandas 1.3.3 and NumPy 1.21.2. The analysis code is available upon request from the corresponding author.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e: This research was supported by grants from the National Research Foundation of Korea (NRF), funded by the Korean government (MSIT) (No. 2022R1A2C100933711; Changhong Youm); the Basic Science Research Program through the NRF, funded by the Ministry of Education (No. 2022R1A6A3A0108756411; Hwayoung Park); and the Ministry of Education of the Republic of Korea and the NRF (No. 2024S1A5B5A16021673; Hwayoung Park). This study received no specific grants from funding agencies in the public, commercial, or non-profit sectors. The funding sources had no role in the study design, data collection, analysis, or interpretation or in writing the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e: All authors declare no financial or non-financial competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e: HC, CY, HP, BK, JH, and SC conceived and designed the study. HC, HP, BK, and SC recruited the participants. HC, CY, HP, BK, JH, and SC performed the data acquisition. HC, CY, HP, BK, JH, and SC analyzed and interpreted the data. HC, CY, HP, BK, JH, and SC drafted the article. All authors read and approved the final version of the manuscript submitted.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e: The authors thank all participants of this study. This work was supported by a Dong-A University Research Fund. The authors also thank Editage (www.editage.co.kr) for English language editing.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eYin W, Zhu W, Gao H, Niu X, Shen C, Fan X, et al. Gait analysis in the early stage of Parkinson\u0026rsquo;s disease with a machine learning approach. Front Neurol. 2024;15:1472956.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMermelstein S, Barbosa P, Kaski D. Neurological gait assessment. Pract Neurol. 2024;24:11\u0026ndash;21.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMirelman A, Bonato P, Camicioli R, Ellis TD, Giladi N, Hamilton JL, et al. Gait impairments in Parkinson\u0026rsquo;s disease. Lancet Neurol. 2019;18:697\u0026ndash;708.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHuo H, Zhang C, Liu W, Zhao C, Ma L, Wang J, et al. Early detection of Parkinson\u0026rsquo;s disease using a multi area graph convolutional network. Sci Rep. 2025;15:5561.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLin C-H, Wang F-C, Kuo T-Y, Huang P-W, Chen S-F, Fu L-C. Early detection of Parkinson\u0026rsquo;s disease by neural network models. IEEE Access. 2022;10:19033\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAboutorabi A, Arazpour M, Bahramizadeh M, Hutchins SW, Fadayevatan R. The effect of aging on gait parameters in able-bodied older subjects: a literature review. Aging Clin Exp Res. 2016;28:393\u0026ndash;405.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMirelman A, Ben Or Frank M, Melamed M, Granovsky L, Nieuwboer A, Rochester L, et al. Detecting sensitive mobility features for Parkinson\u0026rsquo;s disease stages via machine learning. Mov Disord. 2021;36:2144\u0026ndash;55.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRizzo G, Copetti M, Arcuti S, Martino D, Fontana A, Logroscino G. Accuracy of clinical diagnosis of Parkinson disease: a systematic review and meta-analysis. Neurology. 2016;86:566\u0026ndash;76.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBuckley C, Alcock L, McArdle R, Rehman RZU, Del Din S, Mazz\u0026agrave; C, et al. The role of movement analysis in diagnosing and monitoring neurodegenerative conditions: insights from gait and postural control. Brain Sci. 2019;9:34.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDaneault JF, Vergara-Diaz G, Parisi F, Admati C, Alfonso C, Bertoli M, et al. Accelerometer data collected with a minimum set of wearable sensors from subjects with Parkinson\u0026rsquo;s disease. Sci Data. 2021;8:48.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBrognara L, Palumbo P, Grimm B, Palmerini L. Assessing gait in Parkinson\u0026rsquo;s disease using wearable motion sensors: a systematic review. Diseases. 2019;7:18.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJuneau P, Baddour N, Burger H, Bavec A, Lemaire ED. Amputee fall risk classification using machine learning and smartphone sensor data from 2-minute and 6-minute walk tests. Sens (Basel). 2022;22:1749.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBohannon RW, Bubela D, Magasi S, McCreath H, Wang YC, Reuben D, et al. Comparison of walking performance over the first 2 minutes and the full 6 minutes of the six-minute walk test. BMC Res Notes. 2014;7:269.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eValet M, Pierchon L, Lejeune T. The 2-min walk test could replace the 6-min walk test in ambulant persons with subacute or chronic stroke: a two-stage retrospective study. Int J Rehabil Res. 2023;46:41\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGhislieri M, Agostini V, Rizzi L, Fronda C, Knaflitz M, Lanotte M. Foot\u0026ndash;floor contact sequences: a metric for gait assessment in Parkinson\u0026rsquo;s disease after deep brain stimulation. Sens (Basel). 2024;24:6593.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHadouiri N, Monnet E, Gouelle A, Sagawa Y Jr, Decavel P. Locomotor strategy to perform 6-minute walk test in people with multiple sclerosis: a prospective observational study. Sens (Basel). 2023;23:3407.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEl Maachi I, Bilodeau GA, Bouachir WD. Deep 1D-Convnet for accurate Parkinson disease detection and severity prediction from gait. Expert Syst Appl. 2020;143:113075.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang X, Ye Q, Cai G, Wang Y, Cai G. PD-ResNet for classification of Parkinson\u0026rsquo;s disease from gait. IEEE J Transl Eng Health Med. 2022;10:2200111.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePedrero-S\u0026aacute;nchez JF, Belda-Lois JM, Serra-A\u0026ntilde;\u0026oacute; P, Ingl\u0026eacute;s M, L\u0026oacute;pez-Pascual J. Classification of healthy, Alzheimer and Parkinson populations with a multi-branch neural network. Biomed Signal Process Control. 2022;75:103617.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChoi H, Youm C, Park H, Kim B, Hwang J, Cheon SM, et al. Convolutional neural network based detection of early stage Parkinson\u0026rsquo;s disease using the six minute walk test. Sci Rep. 2024;14:22648.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGloeckl R, Teschler S, Jarosch I, Christle JW, Hitzl W, Kenn K. Comparison of two- and six-minute walk tests in detecting oxygen desaturation in patients with severe chronic obstructive pulmonary disease\u0026mdash;a randomized crossover trial. Chron Respir Dis. 2016;13:256\u0026ndash;63.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEl-Gohary M, Pearson S, McNames J, Mancini M, Horak F, Mellone S, et al. Continuous monitoring of turning in patients with movement disability. Sens (Basel). 2013;14:356\u0026ndash;69.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEckmann JP, Kamphorst SO, Ruelle D. Recurrence plots of dynamical systems. Europhys Lett. 1987;4:973\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLema\u0026icirc;tre G, Nogueira F, Aridas CK. Imbalanced-learn: a python toolbox to tackle the curse of imbalanced datasets in machine learning. J Mach Learn Res. 2017;18:1\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTank VH, Ghosh R, Gupta V, Sheth N, Gordon S, He W, et al. Drug eluting stents versus bare metal stents for the treatment of extracranial vertebral artery disease: a meta-analysis. J Neurointerv Surg. 2016;8:770\u0026ndash;4.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHuang G, Liu Z, Van Der Maaten L, Weinberger KQ. Densely connected convolutional networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE; 2017. pp. 2261-9.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIandola FN, Han S, Moskewicz MW, Ashraf K, Dally WJ, Keutzer K. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and \u0026lt;\u0026thinsp;0.5 MB model size. arXiv. 2016. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.48550/arXiv.1602.07360\u003c/span\u003e\u003cspan address=\"10.48550/arXiv.1602.07360\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBernardo LS, Damaševičius R, Ling SH, de Albuquerque VHC, Tavares JMR. Modified SqueezeNet architecture for Parkinson\u0026rsquo;s disease detection based on keypress data. Biomedicines. 2022;10:2746.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eUchitomi H, Ming X, Zhao C, Ogata T, Miyake Y. Classification of mild Parkinson\u0026rsquo;s disease: data augmentation of time-series gait data obtained via inertial measurement units. Sci Rep. 2023;13:12638.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRehman RZU, Buckley C, Mico-Amigo ME, Kirk C, Dunne-Willows M, Mazza C, et al. Accelerometry-based digital gait characteristics for classification of Parkinson\u0026rsquo;s disease: what counts? IEEE Open J Eng Med Biol. 2020;1:65\u0026ndash;73.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMancini M, Afshari M, Almeida Q, Amundsen-Huffmaster S, Balfany K, Camicioli R, et al. Digital gait biomarkers in Parkinson\u0026rsquo;s disease: susceptibility/risk, progression, response to exercise, and prognosis. npj Parkinsons Dis. 2025;11:51.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTurcato AM, Godi M, Giardini M, Arcolin I, Nardone A, Giordano A, et al. Abnormal gait pattern emerges during curved trajectories in high-functioning Parkinsonian patients walking in line at normal speed. PLoS ONE. 2018;13:e0197264.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSong J, Sigward S, Fisher B, Salem GJ. Altered dynamic postural control during step turning in persons with early-stage Parkinson\u0026rsquo;s disease. Parkinsons Dis. 2012;2012:386962.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMeng L, Pang J, Yang Y, Chen L, Xu R, Ming D. Inertial-based gait metrics during turning improve the detection of early-stage Parkinson\u0026rsquo;s disease patients. IEEE Trans Neural Syst Rehabil Eng. 2023;31:1472\u0026ndash;82.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi X, Huang X, Pang J, Meng L, Ming D. A convolutional neural network based classification method for mild to moderate Parkinson\u0026rsquo;s disease at turns. In: Asian Pacific Conference on Medical and Biological Engineering; 2024. pp. 371\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAtrsaei A, Corr\u0026agrave; MF, Dadashi F, Vila-Ch\u0026atilde; N, Maia L, Mariani B, et al. Gait speed in clinical and daily living assessments in Parkinson\u0026rsquo;s disease patients: performance versus capacity. npj Parkinsons Dis. 2021;7:24.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePedrero-S\u0026aacute;nchez JF, Belda-Lois JM, Serra-A\u0026ntilde;\u0026oacute; P, Moll\u0026agrave;-Casanova S, L\u0026oacute;pez-Pascual J. Classification of Parkinson\u0026rsquo;s disease stages with a two-stage deep neural network. Front Aging Neurosci. 2023;15:1152917.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMancini M, El-Gohary M, Pearson S, McNames J, Schlueter H, Nutt JG, et al. Continuous monitoring of turning in Parkinson\u0026rsquo;s disease: rehabilitation potential. NeuroRehabilitation. 2015;37:3\u0026ndash;10. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3233/NRE-151236\u003c/span\u003e\u003cspan address=\"10.3233/NRE-151236\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWarmerdam E, Hausdorff JM, Atrsaei A, Zhou Y, Mirelman A, Aminian K, et al. Long-term unsupervised mobility assessment in movement disorders. Lancet Neurol. 2020;19:462\u0026ndash;70.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNavarro-L\u0026oacute;pez V, Fern\u0026aacute;ndez-V\u0026aacute;zquez D, Molina-Rueda F, Cuesta-G\u0026oacute;mez A, Garc\u0026iacute;a-Prados P, Del-Valle-Gratac\u0026oacute;s M, et al. Arm-swing kinematics in Parkinson\u0026rsquo;s disease: a systematic review and meta-analysis. Gait Posture. 2022;98:85\u0026ndash;95.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eFerraris C, Amprimo G, Masi G, Vismara L, Cremascoli R, Sinagra S, et al. Evaluation of arm swing features and asymmetry during gait in Parkinson\u0026rsquo;s disease using the azure kinect sensor. Sens (Basel). 2022;22:6282.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBarrett MJ, Wylie SA, Harrison MB, Wooten GF. Handedness and motor symptom asymmetry in Parkinson\u0026rsquo;s disease. J Neurol Neurosurg Psychiatry. 2011;82:1122\u0026ndash;4.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYust-Katz S, Tesler D, Treves TA, Melamed E, Djaldetti R. Handedness as a predictor of side of onset of Parkinson\u0026rsquo;s disease. Parkinsonism Relat Disord. 2008;14:633\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVan Rooden SM, Visser M, Verbaan D, Marinus J. Handedness associated to side of onset of Parkinson\u0026rsquo;s disease? Parkinsonism Relat Disord. 2009;15:546\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePark H, Youm C, Lee M, Noh B, Cheon SM. Turning characteristics of the more-affected side in Parkinson\u0026rsquo;s disease patients with freezing of gait. Sens (Basel). 2020;20:3098.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePark H, Shin S, Youm C, Cheon SM, Lee M, Noh B. Classification of Parkinson\u0026rsquo;s disease with freezing of gait based on 360\u0026deg; turning analysis using 36 kinematic features. J Neuroeng Rehabil. 2021;18:177.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVanmechelen I, Bekteshi S, Haberfehlner H, Feys H, Desloovere K, Aerts JM, et al. Reliability and discriminative validity of wearable sensors for the quantification of upper limb movement disorders in individuals with dyskinetic cerebral palsy. Sens (Basel). 2023;23:1574.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePietrosanti L, Saggio G, Patera M, Suppa A, Giannini F, Verrelli CM. Modeling of shoulder\u0026ndash;elbow movement with exponential parameter identification during walking gaits for healthy subjects and patients with Parkinson\u0026rsquo;s disease. Appl Sci. 2025;15:668.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDel Din S, Godfrey A, Rochester L. Validation of an accelerometer to quantify a comprehensive battery of gait characteristics in healthy older adults and Parkinson\u0026rsquo;s disease: toward clinical and at home use. IEEE J Biomed Health Inf. 2016;20:838\u0026ndash;47.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim DS, Schuetz N, Johnson A, Tolas A, Mantena S, O\u0026rsquo;Sullivan JW, et al. Unlocking insights: clinical associations from the largest 6-minute walk test collection via the my Heart Counts cardiovascular Health Study, a fully digital smartphone platform. Prog Cardiovasc Dis. 2025;89:45\u0026ndash;52.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGal\u0026aacute;n-Mercant A, Ortiz A, Herrera-Viedma E, Tomas MT, Fernandes B, Moral-Munoz JA. Assessing physical activity and functional fitness level using convolutional neural networks. Knowl Based Syst. 2019;185:104939.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePappas MC, Baudendistel ST, Schmitt AC, Au KLK, Hass CJ. Acclimatization of force production during walking in persons with Parkinson\u0026rsquo;s disease. J Biomech. 2023;148:111477.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKosak M, Smith T. Comparison of the 2-, 6-, and 12-minute walk tests in patients with stroke. J Rehabil Res Dev. 2005;42:103\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWu X, Ma L, Wei P, Shan Y, Chan P, Wang K, et al. Wearable sensor devices can automatically identify the ON-OFF status of patients with Parkinson\u0026rsquo;s disease through an interpretable machine learning model. Front Neurol. 2024;15:1387477.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePlotnik M, Wagner JM, Adusumilli G, Gottlieb A, Naismith RT. Gait asymmetry, and bilateral coordination of gait during a six-minute walk test in persons with multiple sclerosis. Sci Rep. 2020;10:12382.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Additional File","content":"\u003cp\u003eAdditional file 3 is not available with this version.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Parkinson’s disease, gait, wearable sensors, artificial intelligence, deep learning, neurodegeneration","lastPublishedDoi":"10.21203/rs.3.rs-7285962/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7285962/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eEarly-stage Parkinson\u0026rsquo;s disease (PD) is characterized by subtle motor symptoms that complicate diagnosis and often delay intervention. Timely and accurate identification is critical for effective management, emphasizing the need for objective, non-invasive diagnostic methods.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eThis study developed a non-invasive approach for early PD detection using wearable sensors and a convolutional neural network (CNN) during a 6-min walk test. The test was segmented into 1-min intervals, extracting three straight-walking and three turning gait phases per minute. Time-series data were collected from 78 patients with early-stage PD and 50 healthy controls across six body locations.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eThe CNN achieved 95.6% accuracy when classifying PD status using gyroscope data from the left arm during the first-minute straight-walking phase. Furthermore, repeated-measures analysis of variance and post hoc tests indicated that a 1- to 2-min measurement window was sufficient for reliable detection, supporting the feasibility of time-efficient clinical screening.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eThese findings suggest that a wearable sensor, placed on a single arm and used to capture first-minute straight gait data, can provide highly accurate and non-invasive early PD detection. Future research should evaluate medication effects, extend validation to broader disease stages, and explore unsupervised learning approaches to identify latent motor phenotypes and enable personalized monitoring.\u003c/p\u003e","manuscriptTitle":"Early Detection of Parkinson’s Disease Using a Single-Arm Wearable Sensor and Convolutional Neural Networks","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-12 12:46:14","doi":"10.21203/rs.3.rs-7285962/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"38fa7293-76e0-424b-8ec8-49b4201e4631","owner":[],"postedDate":"August 12th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-17T19:09:30+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-12 12:46:14","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7285962","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7285962","identity":"rs-7285962","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}