Convolutional neural network-based detection of early-stage Parkinson’s disease using the six-minute walk test

Convolutional neural network-based detection of early-stage Parkinson’s disease using the six-minute walk test | 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 Article Convolutional neural network-based detection of early-stage Parkinson’s disease using the six-minute walk test Hyejin Choi, Changhong Youm, Hwayoung Park, Bohyun Kim, Juseon Hwang, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4482534/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 30 Sep, 2024 Read the published version in Scientific Reports → Version 1 posted 11 You are reading this latest preprint version Abstract The heterogeneity of Parkinson’s disease (PD) generates significant challenges for accurate diagnosis, especially in early-stage disease, when symptoms may be very subtle. This study aimed to determine the accuracy of a convolutional neural network (CNN) technique based on a 6-min walk test (6MWT) using wearable sensors for distinguishing patients with early-stage PD (n = 78) from healthy controls (n = 50). Wearing six sensors, the participants performed the 6MWT, and the time-series data were converted into new images. The main results showed that the gyroscopic vertical component of the lumbar spine had the highest classification accuracy of 83.5%, followed by the thoracic spine (83.1%) and right thigh (79.5%) segment. These results suggest that the 6MWT and CNN models may pave the way for clinicians to diagnose and track PD symptoms earlier and thus provide timely treatment during the golden transition from geriatric to pathologic gait patterns. Health sciences/Biomarkers Health sciences/Neurology Parkinson’s disease detection artificial intelligence deep learning convolutional neural network six-minute walk test Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 INTRODUCTION Parkinson’s disease (PD) is a chronic, degenerative neurological disorder that affects movements [ 1 ], and the characteristic symptoms include increasing difficulty walking, speaking, or completing other simple tasks as well as a combination of motor and non-motor symptoms that can confer functional disability, loss of independence, and decreased quality of life [ 2 , 3 ]. The heterogeneity of PD generates significant challenges for accurate diagnosis, especially in early-stage PD that may include very subtle symptoms [ 4 ]. The severity of motor symptoms in PD is ascertained with the Hoehn and Yahr (H&Y) Scale, Movement Disorders Society (MDS)-sponsored revision of the Unified Parkinson’s Disease Rating Scale (UPDRS), MDS-UPDRS Part III, PD phenotype assessment, and discrete variable-based analysis of physical and fitness factors as well as postural and gait characteristics [ 5 , 6 ]. However, these subjective assessments are scored through clinical observation by neurospecialists [ 7 ] and, therefore, research on the objective, accurate body-part-specific motor-symptom assessment and symptom impact-based classification of disease severity is needed [ 8 ]. Gait in PD is a potential status and trait marker, because gait impairments appear in early-stage PD, precede the appearance of overt motor signs, and progress more rapidly than other motor characteristics of PD [ 2 , 9 ]. Objective quantification of individual gait characteristics may include three-dimensional (3D) motion capture, force plates, instrumented walkways, treadmills, and electromyography [ 10 ], and these tools are essential for accurately characterizing PD gait disorders in clinical populations. However, the application of these tools is primarily constrained to research environments owing to the cost and the need for specialized expertise for utilization [ 11 ]. Thus, monitoring gait outcomes with wearable sensors may be a useful, inexpensive alternative option for ascertaining the presence of impairment in a controlled daily living environment [ 12 – 14 ]. The 6-min walk test (6MWT) is a submaximal movement assessment that is routinely used to assess an individual’s aerobic capacity and physical mobility [ 15 , 16 ]. Originally developed for use in patients with chronic respiratory or cardiovascular disease, the 6MWT is currently used to assess diverse populations, including older adults and people who have had a stroke, lower extremity amputations, or PD [ 17 – 20 ]. The 6MWT constitutes a simple, inexpensive (minimal setup and space), safe, reproducible, and alternative functional test [ 21 ]. Despite the limitation that the standard clinical outcome of the traditional 6MWT only comprises the distance walked in 6 min, the distance walked shows a moderate correlation with more complex aerobic capacity tests, such as V̇O 2 max (maximal oxygen uptake) and, thus, can minimize the burden on patients and clinicians [22,23]. Furthermore, compared to other gait tests, the well-defined settings and longer duration of the 6MWT constitute favorable conditions for obtaining useful information from both turning and straight gaits, and this potentially increases the test efficiency [24]. Therefore, the 6MWT conducted using wearable sensors appears clinically applicable and comprehensive, and is a quick, easily applicable method that provides objective details of gait quality [25,26]. The recent trend of the increased use of artificial intelligence-based machine learning and deep learning models to classify and predict diseases [ 27 ] has advanced research to resolve the problem of motor-symptom severity classification with the comprehensive gait characteristics of people with PD, such as by using a machine learning algorithm for distinguishing three types of neurodegenerative diseases (PD, Huntington’s disease, and amyotrophic lateral sclerosis) [ 28 ] and diagnosing and assessing PD severity according to the H&Y scale [ 29 ]. Borzì et al. reported neural network-based freezing of gait detection from the 6MWT in PD by using the waist sensor-derived acceleration data [ 30 ]. Juneau and Daines et al. used a random forest model to classify the fall risk through accelerometer- and gyroscope-derived data that were recorded using a smartphone placed over the posterior pelvis during the 6MWT in patients with lower limb amputations [ 19 , 31 ]. However, these studies have limitations such as a small sample size; low accuracy due to the analysis of movement data of specific body segments based on a single wearable sensor without objective, specific identification of body segments according to disease severity; and the use of unproductive, inefficient, and less sensitive tools [ 13 ]. Comparative analysis of passive signal classification and gait parameter extraction from wearable sensors for the detection and prediction of gait-disorder signals in PD through gait assessment has limited utility in the identification of subtle motor symptomatic disorders in early- to mid-stage PD [ 12 ]. Thus, the interest in deep learning technologies, which are automated approaches for objective diagnosis, quantitative assessment, and accurate predictive disease progression modelling, to overcome these limitations has increased recently [ 32 , 34 ]. EI Maachi et al. and Yang et al. proposed an intelligent gait-based PD-detection system that used a convolutional neural network (CNN) [ 32 , 33 ]. In a wearable sensor-based study, Nguyen et al. classified abnormal gait in musculoskeletal disorders through a CNN-derived sensor-based spectrogram approach [ 35 ]. For objective gait-disorder signal detection, we believe that accuracy could be improved by using time-series (TS) gait patterns and classifying them into CNN architecture-converted visual images [ 34 ]. However, studies specifically related to CNN-based 6MWT identification to categorize early-stage PD are limited. Therefore, this study aimed to identify segments and components that exhibit optimal accuracy for classifying people with early-stage PD and healthy controls using a CNN algorithm for sensor-based 6MWT data. The research hypothesis is as follows: First, CNN algorithms can improve the classification of early-stage PD and healthy controls, wherein TS signals of accelerometer and gyroscope-based gait patterns obtained using six segmentally mounted wearable sensors are converted into images to determine the classifiability of the two groups. Second, based on the acceleration and gyroscope data obtained from the wearable sensors, the characteristics of early-stage PD that are prominent in specific components of the six segments can be identified. RESULTS CNN classification results for the early PD and control groups We visualized triaxial accelerations and gyroscope data using TS images during the 6MWT to identify the most prominent gait variables for classification into two groups according to early PDs and controls. The proposed algorithms, such as Recurrence Plot (Rec), are based on CNNs. The accuracy results for all variables are summarized in Fig. 1 , with an accuracy range of 49.5–83.5%. The variables with the highest accuracy in each body segment were the z -axis gyroscope data (Gyr_Z) on the left arm segment at 76.6%, the i-axis acceleration (Acc_Z) for the right arm segment at 70.4%, the acceleration resultant value of the x- , y- , and z -axes (Acc_XYZ) for the left thigh at 78.4%, and the y -axis acceleration (Acc_Y) for the right thigh at 79.5% using the DenseNet model. Using the SqueezeNet model, the x -axis gyroscope data (Gyr_X) for the thoracic spine segment demonstrated an accuracy of 83.1%, and the x -axis gyroscope data (Gyr_X) for the lumbar spine segment demonstrated an accuracy of 83.5%. Figure 2 shows the confusion matrices for the variables with the highest accuracy in body segments. We present representative images and corresponding frequency patterns for an early PDs and controls for the variables that showed the highest accuracy (Fig. 3 ). The representative images and graphs show the last phase of the 6MWT, including straight walking and turning. The mean accuracy, precision, recall, and F1 scores for all gait variables for classification between early PDs and controls are shown in Supplementary Table S1 . DISCUSSION The primary finding of the present study is that, in the CNN-algorithm and sensor-based 6MWT-derived classification of early PDs and healthy controls, the gyroscopic vertical component of the lumbar spine (Gyr_X) has the highest classification accuracy (83.5%), followed by the gyroscopic vertical component of the thoracic spine (Gyr_X; 83.1%), the acceleration mediolateral component of the right thigh (Acc_Y; 79.5%), the acceleration resultant component of the left thigh (Acc_XYZ; 78.4%), the gyroscopic anteroposterior component of the left arm (Gyr_Z; 76.6%), and the acceleration anteroposterior component of the right arm (Acc_Z; 70.4%). Recent studies have harnessed the power of the 6MWT method as a promising tool in gait analysis by using wearable-sensor technology and have explored its potential applicability for investigating the observability and classifiability of continuous and segment-specific gait patterns over time in diverse populations, including those with lung transplants [ 36 ], multiple sclerosis [ 37 ], lower-limb amputations [ 19 , 31 ], and PD [ 26 , 30 , 38 ]. Despite the variations in disease groups, the 6MWT has promising potential for detecting a decline in motor function and facilitating the management of prolonged gait. This opens exciting possibilities for further studies to determine the accuracy of the 6MWT in the early identification and severity determination of PD [ 26 , 30 ]. These studies are crucial to advance our understanding and application of the 6MWT. Borzì et al. used waist-wearable sensors and multi-head CNN techniques for the detection and prediction of freezing of gait (FOG) during a 6MWT in PD and achieved a specificity of 100% in healthy older adult participants [ 30 ]. Bailo et al. extracted gait domain-related metrics, such as rhythm and pace, regularity, intensity, dynamic instability, and asymmetry from lower truncal accelerations and angular velocities, during the 6MWT in participants with mild PD, and reported significant differences with healthy controls [ 38 ]. However, none of these studies have focused on classifying early-stage PD using CNN algorithms and components of sensor-based triaxial accelerations and gyroscopic data obtained from the 6MWT. The novelty of our study lies in the use of wearable sensors and CNN algorithms to classify early-stage PD using the 6MWT, which may have significant implications in clinical and daily-life applications. We fitted wearable sensors to each of the six segments during the 6MWT and used a CNN algorithm to determine the classification possibilities of the two groups. Axial rigidity, increased postural tone, en bloc turning, FOG, and loss of intersegmental flexibility may be evident during gait in PD, particularly in walking and turning gait [ 39 – 42 ]. Specifically, individuals with early-stage PD experience a reduced range of motion at the hip joint [ 43 ], asymmetric tremors, decreased coordination [ 44 ], and reduced postural stability control [ 45 ] during gait, which exacerbate with progression to moderate PD [ 46 ]. However, in early-stage PD, as the symptoms are not well differentiated and can easily be confused with other conditions, further investigation is needed [ 43 , 45 ]. Cai et al. suggested that asymmetry and axial features of the trunk and lumbar spine contribute to the early PD classification model for discriminating healthy controls from individuals with H&Y Stage 1 limitation [ 47 ]. During walking, people with early-stage PD can exhibit significantly impaired asymmetry because of unilateral parkinsonism-related symptoms, with consequent axial impairments [ 48 ]. In this study, early-stage PD was classified through six body segments at H&Y stages 1 and 2, with similar classification accuracy for the affected and unaffected limb segments, and the highest accuracy in the lumbar spine and truncal segments, and this is consistent with the results from previous studies. Nonetheless, it is difficult to generalize our results because of the limited evaluability of a single waist sensor [ 26 , 30 ]. Therefore, asymmetric and unilateral measurements, including of axial features, should be used whenever possible to detect early-stage PD in clinical and community settings. Despite the asymmetrical basal ganglia degeneration in early-stage PD that may lead to unilateral symptoms, measurement of gait variation between the left and right sides may not be reliable for quantifying motor asymmetry in patients [ 45 ]. Furthermore, as the side on which symptoms will occur is unclear, a comparative analysis of bilateral-limb sensor data is needed [ 44 ]. Buckley et al. proposed that upper-body acceleration may be a biomarker for gait impairment in early-stage PD [ 49 ]. In combination with spatiotemporal information, variables of upper body acceleration contribute to a better description of the PD gait [ 50 ] and could be associated with a reduced ability to regulate repetitive steps and strides during gait or control the rhythmic displacements of the upper body during walking [ 49 ]. We found that specific components of the limb segments, such as the mediolateral component for the right thigh position, had high classification accuracy. Research on symmetry in the walking domain indicated that measuring gait symmetry in the mediolateral and vertical directions was significantly associated with walking endurance (6MWT), whereas no relationship was found for balance scales; this suggests that gait symmetry may be primarily associated with the energy efficiency of locomotion, rather than balance [ 38 ]. Similarly, our results showed that the gyroscopic vertical (Gyr_X) component demonstrated the highest discrimination accuracy for the lumbar spine (83.5%), followed by the thoracic spine (83.1%). In particular, increased rigidity and signs of akinesia and/or bradykinesia were associated with smaller truncal sway, mainly in the anteroposterior and vertical directions [ 51 ]. Regarding gait and balance, the lower truncal acceleration amplitude was associated with lower walking endurance and poorer dynamic balance; therefore, reduced truncal acceleration may result from the minimizing of upper body movements during walking to compensate for lower-limb impairment and maintain balance [ 38 ]. The strengths of our work are that we identified the characteristics of early-stage PD and classification accuracy that were prominent in individual axial components, specifically in each of the six segments. We determined that considering the three-axis factors together may help identify and pre-screen for gait characteristics that indicate early-stage PD. Nevertheless, there are some limitations to this study. We conducted our study in the “medication ON state.” Considering both ON and OFF medication states may enable a better understanding of the characteristics of early-stage PD. For this, a larger dataset is needed to validate the effectiveness of the model. Future studies should aim to propose a measurement time for the 6MWT, considering measurement efficiency, such as for 2 or 3 min, by using specific intervals (e.g., 15 s) [ 36 , 52 ]. Our study involved a CNN analysis of the entire 6 minutes, but without distinguishing between straight and turning gaits in the 6MWT, and merely identified schematic frequency patterns for straight and turning gaits. Therefore, in the future, it is necessary to extract data specific to turning and straight gait sections and compare the classification performance using deep learning and machine learning techniques [ 53 ]. Furthermore, further analysis of specific biomarkers, such as amplitude and magnitude based on wearable sensors, could go further than the identification of the best-performing body segments [ 47 ]. To additionally compare detailed gait characteristics, we identified frequency patterns for the last phase of the 6MWT, including straight and turning gaits, for early PDs and controls (Fig. 3 ). We found the highest accuracy in the Gyr_X component, especially in the lumbar and thoracic spine, where characteristics such as the differences in frequency patterns on the graph during turning and straightforward walking between early PDs and controls (Fig. 3 ) on the gyroscope data of the x -axis may have influenced the image generation and affected the accuracy. In future work, it is important to extract relevant features that can identify straight and turning gait patterns in the 6MWT from the domains of bilateral coordination, asymmetry, truncal symptom severity, and variability by performing non-linear analysis, and then use machine learning techniques to analyze specific biomarkers associated with early-stage PD [ 54 ]. Finally, we only considered patients with early H&Y-stage PD; however, patients with severe symptoms may differ from those analyzed herein. Future studies should include more subjects with intermediate and tremor-dominant subtypes. CONCLUSION In conclusion, our study confirms the possibility of classifying two groups when converting TS data of 6MWT into images, using CNN algorithms, and identifying which segments and components exhibit optimal accuracy. We suggest that 2D imaging for analyzing infinite-length 1D TS data may provide an opportunity to screen the gait patterns of patients with early-stage PD for primary classification. We believe that this work will not only improve the accuracy of motor-symptom-severity assessment and classification but also contribute to the improvement of time-consuming, economical, and favorable experimental and analytical procedures, enable early disease detection in clinical settings, and the availability of objective indicators for disease-severity assessment [ 2 ]. METHODS Participants A total of 79 individuals with PD (early PDs) and 50 age-matched healthy controls (controls) participated in the study and were included in the final analysis. Figure 4 depicts a flowchart of the study process and participant details. Based on the United Kingdom’s Parkinson's Disease Society Brain Bank [ 55 ] criteria, a neurospecialist diagnosed patients with idiopathic PD. The inclusion criteria were as follows: (a) age 40–85 years, (b) ability to walk and move independently without assistive devices, with a modified H&Y stage of 1–2 [ 56 , 57 ], (c) Mini-Mental State Examination (MMSE) score ≥ 24 [ 58 ], and (d) stable response to anti-Parkinson’s medications. Only one participant who dropped out during the 6MWT was excluded from the analysis. Individuals with a history of cardiovascular, musculoskeletal, vestibular, or other neurological diseases, who required mobility-assistive devices, and with pharmacotherapy-refractory dyskinesia were excluded. The controls included individuals with no medical history of orthopedic surgery or cognitive impairment and gait disturbance in the past 6 months. The physical and clinical characteristics of the participants are shown in Table 1 . All experiments were performed in accordance with the relevant guidelines and regulations. The study protocol was approved by the Institutional Review Board (IRB) of Dong-A University Medical Center (IRB number: DAUHIRB-22-089). All participants provided written informed consent before participating in this study. The study is registered in the Clinical Research Information Service in the Republic of Korea (KCT0009353). Experimental procedures Participants performed two sessions. In the first session, the participants completed the informed consent form and were assessed by using the MMSE, Montreal Cognitive Assessment (MoCA) [ 59 ], UPDRS, and modified H&Y scale, and the physical characteristics such as sex, age, height, body mass, and body mass index were measured (Table 1 ). In the second session, before performing the main task, wearable sensors were attached to the left and right upper arms (5 cm above the lateral humeral epicondyle), thighs (10 cm above the lateral femoral epicondyle), thoracic spine (T10), and lumbar spine (center of the left and right posterior superior iliac spines) region using a stretchable belt (Fig. 5 a). Participants performed the 6MWT as the main task, wherein they were instructed to walk, without running, as far as possible in the preferred direction for 6 min. The 20-m course was demarcated by cones placed at each end (Fig. 5 b). The raters walked behind participants to ensure their safety, offer encouragement, and let them know how much time they had left. All experiments were performed in the “On” medication state approximately 2 h after taking the medication. Framework for classifying early PDs and controls The framework for sensor-based classification of early PDs and controls in this study consists of four major process steps (Fig. 6 ). The overall flow was designed to acquire raw data from wearable sensors, perform data pre-processing, categorize the sample data into the training and test sets, and then train the CNN to evaluate the model's performance. Data acquisition Data were collected using an Xsens DOT Wearable Sensor platform (Movella Technologies, Enschede, the Netherlands) with a package size of 36.30 × 30.35 × 10.80 mm and a weight of 11.2 g. Signal-based TS gait characteristics were extracted from each wearable sensor (accelerometer and gyroscope) in each axis ( x -axis: vertical, y -axis: mediolateral, z -axis: anteroposterior; Fig. 5 a). We collected the data on 3D-acceleration (m/s 2 ; up to ± 16 g) and 3D angular velocity (°/s; up to ± 2000°/s) via the accelerometer and gyroscope, respectively, in the wearable sensor. By default, the local Earth-fixed reference coordinate system (L) used is defined as a right-handed Cartesian coordinate system (East-North-Up [ENU]) with X positive to the East (E), Y positive to the North (N), and Z positive when pointing up (U) and is the standard reference in inertial navigation for aviation and geodetic applications. With the default ENU (L) coordinate system, the X-sens yaw output is defined as the angle between East (X) and the horizontal projection of the sensor x -axis, positive about the local vertical axis (Z) following the righthand rule. The orientation calculated by X-sens DOT is the orientation of the sensor coordinate system (S) with regard to L. The orientation output is in the ENU frame: accordingly, the roll, pitch, and yaw will equal 0 degrees when the x -, y -, and z -axes of the sensor are aligned with the east, north, and upward directions, respectively. Thus, X, Y, and Z are positive when pointing to the local magnetic east, the local magnetic north, and upward, respectively. The data output rate was set to 60 Hz. Three-dimensional data were collected using an iPad (iOS 15.6.1; Apple, Cupertino, CA) with the MovellaDOT app using Bluetooth 5.0 communication and analyzed using MATLAB R2023a (MathWorks, Natick, MA). Data pre-processing Before the data pre-processing, we checked the graphs of the total 6 min of raw data from each of the six sensors. Next, the following data pre-processing steps were carried out. First, we categorized 6 min (360 s) of TS data from each participant’s six sensors into 15-s segments that represent one phase of the 24 data phases, including straight and turning gait. Second, we normalized the data to eliminate the effect of the difference in participants’ height and sensor-placement variations following reattachment. The normalization procedure involved centering the sensor data within each trial to achieve a median of 0 while retaining the original scale of the sensor data. The median value of the data was adjusted to align with the 0-reference point. Lastly, we set the length of the sample to 896 to create a single image. The length of the 896 samples is proportional to the pixel size. We multiplied the pixel size of the input images (224 pixels) by an even number factor (4) for the sample in each data phase for all participants to obtain a sample length of 224 pixels × 4 = 896 (approximately 9 s). This process constitutes the standardized approach to effectively construct a CNN model and train it with an optimized resolution image [ 60 ]. For trials lasting less than 8.96 s, we applied zero padding. The resulting values of the three axes, as well as the acceleration and gyroscope data for each of the x-, y- , and z- axes, were used to analyze the TS data and their components were calculated with resultant acceleration and gyroscope data according to the following formula: $$Resultant data=sqrt(x^2+y^2+z^2)$$ 1 Data pre-processing was performed using MATLAB. Data generation For training and testing CNN models using the accumulated imaging dataset of all phases, the dataset was randomly divided, per subject, into the training, validation, and test sets in a ratio of 50%, 20%, and 30%, respectively. The dataset initially consisted of 1776 samples for early PDs and 1056 samples for controls across six sensors for each segment (left arm, right arm, left thigh, right thigh, thoracic spine, and lumbar spine). Therefore, the unbalanced dataset was randomly oversampled using the imbalanced-learn (version 0.10.1) Python package to perform an unbiased analysis, and this generated a total of 3552 samples for each body segment, with a sample count of 1776 for both early PDs and controls, in the entire dataset. Training models and classification The deep learning classification analysis based on the CNN models and TS imaging methods was performed as follows. For the variance analysis, fivefold cross-validation was employed. The batch size used in this experiment was 128; Adam optimizers were used, and the learning rate was used for the optimizer of 1e-05. After 150 epochs, the proposed model reaches saturation, and training is stopped with the help of early stopping. To overcome overfitting, only the best model was saved; this meant that, during the training phase, if the validation accuracy of the epoch was higher than the highest accuracy, then the model was saved. TS imaging methods The perspective of the data was broadened from one-dimensional (1D) TS to two-dimensional (2D) images to analyze the temporal aspect of the gait data processed using TS imaging methods, specifically Recurrence Plot (Rec) [ 61 ]. This approach offers two distinct advantages. First, it augments the dimensionality of the feature space, thereby enabling a more comprehensive examination of the gait data. Second, it leverages established CNN-based deep learning architectures for classification. The acceleration and gyroscope data of 24 phases (6 minutes divided into 15-s intervals) were used to analyze the TS data. After converting TS data from each phase to a matrix form, the converted TS data were generated in a new imaging, as described further in the subsections on the TS imaging methods. Rec plot The Rec is a method employed to transform 1D TS into 2D images by representing recurring states of the TS [ 61 ]. Given a TS \(X=({x}_{1}, {x}_{2}, \cdots , {x}_{n})\) with a length of n , the trajectory at a discrete time point i is define d as follows: $${\dot{x}}_{i}=\left({x}_{i},{x}_{i+r},\cdots ,{x}_{i+\left(m-1\right)r}\right),\forall i\in \left\{1,\cdots ,n-\left(m-1\right)\tau \right\}$$ 2 Here, m represents the dimension of the trajectories, and τ is the time delay. Using these trajectories, the binarized recurrence matrix \({\widehat{R}}_{i,j}\) , with i and j as discrete timepoints, is calculated as the pairwise distance between the trajectories, as shown below: $${\widehat{R}}_{i,j}=\theta \left(\epsilon -‖{\dot{x}}_{i}-{\dot{x}}_{j}‖\right),\forall i,j\in \left\{1,\cdots ,n-(m-1)\tau \right\}$$ 3 Where \(‖ \bullet ‖\) represents a norm operation, θ is the Heaviside function, and ε is the recurrence threshold. One limitation of this approach is that the calculated values are discretized. To calculate the continuous matrix elements of the recurrence matrix, the Heaviside function and recurrence threshold are excluded from the equation, resulting in the continuous-valued recurrence matrix \({\widehat{R}}_{i,j}\) , which is expressed as: $${R}_{i,j}=‖{\dot{x}}_{i}-{\dot{x}}_{j}‖,\forall i,j\in \left\{1,\cdots ,n-(m-1)\tau \right\}$$ 4 In our data analysis, we utilized the continuous recurrence matrix. We simplified the calculation by setting the dimension of the trajectories m to 1 and the time delay τ to 1 when converting the TS series data of gait into 2D images. CNNs For the classification of early PDs and controls using the processed TS images, three CNN architectures were employed in the analysis: Residual Neural Network (ResNet) [ 62 ], Dense Convolutional Network (DenseNet) [ 63 ], and SqueezeNet [ 64 ]. In this study, the input image size for the CNN models was configured as 224 × 224 pixels (Fig. 6 ). We employed a ResNet model with 18 layers. We utilized a DenseNet-121 model consisting of four dense blocks with 6, 12, 24, and 16 channels, respectively, and employed SqueezeNet v1.0. These CNN architectures were chosen based on the processed TS images for classification tasks. ResNet ResNet is a well-known convolution-based deep neural network that classifies 2D images. It introduced the concept of a residual block, which addresses the challenge of training deeper neural networks by creating a shortcut connection between the input and output layers. This shortcut connection helps prevent issues such as vanishing or exploding gradient problems during training, and thus facilitate the training of deep neural networks. In ResNet, this skip connection is implemented by summing the output, represented as \(H( \bullet )\) , with the input, denoted as x : $$H\left(x\right)=F\left(x\right)+x$$ 5 Here, F represents an activation function with weight parameter w and bias parameter b , similar to the output of a standard neural network. Adding x in this equation effectively performs identity mapping, transferring the input directly to the output. Achieving this identity mapping within the residual block, especially when dealing with cases where the dimensions of the input and output layers differ, requires special consideration in two distinct cases: 1) when the dimensions of input and output layers are equivalent and 2) when they differ. In the former case, element-wise summation suffices for identity mapping. In the latter case, Tank et al. have proposed two solutions: zero padding and the utilization of 1 × 1 convolutional layers [ 62 ]. Zero padding aligns the dimensions by adding zero-valued elements to the input tensor. Conversely, 1 × 1 convolutional layers enable dimension projection, ensuring compatibility between input and output dimensions. These techniques, along with the concept of residual blocks, contribute to the effectiveness of ResNet in training very deep neural networks for image-classification tasks. DenseNet DenseNet, which was introduced as an enhancement over ResNet [ 63 ], was devised to improve performance whilst utilizing fewer network parameters. Its primary innovation revolves around dense connectivity, where all preceding feature maps are utilized for identity mapping. The authors’ core motivation stemmed from concerns that the summation operation employed in residual blocks might hinder the efficient flow of information through the network. To address this limitation, they introduced dense connectivity, concatenating all preceding feature maps for identity mapping. Here, \({x}_{l}\) represents the output of the \(l\) th layer in the network, and \({x}_{0}\) denotes the input. The concatenation operation is expressed as follows: $${x}_{l}={H}_{l}\left(\left[{x}_{0},{x}_{1},\cdots ,{x}_{l-1}\right]\right)$$ 6 Here, \({H}_{l}\) is a composite function involving batch normalization (BN), Rectified Linear Unit (ReLU) activation, and a 3 × 3 convolution. \({H}_{l}\) takes the feature maps of all preceding layers, \({x}_{0},{x}_{1},\cdots ,{x}_{l-1}\) , as input, and \(\left[ \bullet \right]\) denotes concatenation. Besides dense connectivity, the authors introduced two operations to reduce the number of feature maps: bottleneck layers and compression. Bottleneck layers consist of BN-ReLU-1 × 1 convolution, followed by BN-ReLU-3 × 3 convolution. The inclusion of a 1 × 1 convolution in the bottleneck layer aims to reduce the parameter count of the input-feature maps. Compression is employed to reduce feature map dimensions using a compression factor θ within a transition layer, which acts as a down-sampling layer. The transition layer comprises a 1 × 1 convolution layer followed by a 2 × 2 average pooling layer, and it is positioned between the dense blocks. DenseNet mitigates vanishing gradient problems by leveraging dense connectivity, ensuring that information from earlier layers is preserved and reused throughout the network. This advantageous property enables DenseNet to surpass the performance of ResNet whilst maintaining a more efficient parameter utilization, and thereby results in a reduction of computational resource requirements. SqueezeNet SqueezeNet aims to create a more compact neural network with a reduced number of parameters whilst maintaining high performance, which makes it well-suited for resource-constrained computational environments [ 64 ]. The authors noted that AlexNet had a parameter count of 240 MB, whereas SqueezeNet achieved a remarkable reduction to merely 4.8 MB. The authors introduced three key strategies to achieve this substantial reduction in model size. First, instead of using 3 × 3 convolutions, SqueezeNet relies on 1 × 1 convolutions in certain layers. This change significantly reduces the number of parameters in the network whilst preserving its capacity to learn important features. Second, another strategy involves reducing the number of input channels. The model's overall parameter count is effectively reduced by decreasing the number of input channels. Third, later in the network architecture, SqueezeNet adopts a down-sampling approach that is strategically placed to ensure that the network generates large activation maps. This design choice optimizes the model's performance whilst keeping the parameter count low. Central to the architecture of SqueezeNet is the fire module, which consists of a squeeze layer (comprising only 1 × 1 convolutions) followed by an expand layer (comprising a combination of 1 × 1 and 3 × 3 convolutions). Notably, the authors designed the squeeze layer to have fewer hyperparameters than the total number in the expanding layer, which aligns with the abovementioned second strategy. Evaluation index The performance of the two-group classification model was evaluated based on its accuracy, which represents the ratio of the correctly predicted number of samples to the total number of samples. Particularly, accuracy is a feasible evaluation metric for classification problems, which are well-balanced and not skewed or feature no class imbalance. To evaluate the performance of classifiers on datasets, we use the confusion matrix for the binary classification problem under analysis. For a confusion matrix, four measures, namely “true positive” (TP), “true negative” (TN), “false positive” (FP), and “false negative” (FN), have been reported. TP is an outcome wherein the model correctly predicts the positive class, which are the real early PDs labeled data in the test dataset and are classified as early PDs by training. TN is an outcome wherein the model correctly predicts the negative class, which are the real controls-labeled data in the test dataset and are classified as controls through training. FP is an outcome wherein the model incorrectly predicts the positive class, which are the real early PDs-labeled data in the test dataset and are classified as controls via training. FN is an outcome wherein the model incorrectly predicts the negative class, which are the real controls-labeled data in the test dataset and are classified as early PDs through training. The precision represents the proportion of identifications that are correct in practice. The recall is a measure of how many of the positive cases the classifier correctly predicted, over all the positive cases in the data. It is sometimes also referred to as the sensitivity. The F1-Score is a measure encompassing precision and recall. This combined measure is generally described as the harmonic mean of the two-constituent metrics. In essence, the harmonic mean is an alternative approach for calculating the average of values, which is more suitable for ratios (e.g., precision and recall) than the traditional arithmetic mean [ 65 , 66 ]. Statistical analysis The Shapiro–Wilk test was used for testing multivariate normality. Fisher’s exact test, the Mann–Whitney U test (for non-normal data), or the independent t -test (for normal data) was conducted to assess the differences in physical and clinical characteristics between early PDs and controls. Statistical analyses were performed using SPSS 21.0 (IBM, Armonk, NY), MATLAB, and Python (Python 3.10, Python Software Foundation). The statistical significance level was set at 0.05. Declarations DATA AVAILABILITY 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 code 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 code used for the analysis may be requested from the corresponding author. ACKNOWLEDGEMENTS This work was supported by a Dong-A University Foundation Grant. The authors thank all the participants of this study. The authors also thank Editage (www.editage.co.kr) for English language editing. This work was supported by a grant [no. 2022R1A2C100933711; Changhong Youm] from the National Research Foundation of Korea (NRF), funded by the Korean government (MSIT). This research was also supported [grant no. 2022R1A6A3A0108756411; Hwayoung Park] by the Basic Science Research Program through the NRF, funded by the Ministry of Education. The funders had no role in the study design, collection, analysis, and interpretation of the data and in writing the manuscript. AUTHOR CONTRIBUTIONS HC, CY, HP, BK, SC, and SS conceived and designed the study. HC, HP, BK, and SC. recruited the participants. HC, CY, HP, BK, SC, and SS performed the data acquisition. HC, CY, HP, BK, SC, and SS analyzed and interpreted the data. HC, CY, HP, BK, SC, and SS drafted the article. All authors read and approved the final version of the manuscript submitted. ETHICS DECLARATIONS 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 (IRB number: DAUHIRB-22-089) (see ethics approval letter in the supplementary file). All patients provided written informed consent before data collection. The study is registered in the Clinical Research Information Service in the Republic of Korea (KCT0009353). COMPETING INTERESTS The authors declare that they have no conflict of interests. CONSENT FOR PUBLICATION Not applicable. ADDITIONAL INFORMATION Supplementary information References Aversano, L., Bernardi, M. L., Cimitile, M., & Pecori, R. Early detection of Parkinson disease using deep neural networks on gait dynamics. In 2020 International Joint Conference on Neural Networks (IJCNN) (pp. 1-8). IEEE. (2020). Rehman, R. Z. U. et al . Accelerometry-based digital gait characteristics for classification of Parkinson's disease: what counts?. IEEE Open J. Eng. Med. Bio. 1 , 65-73 (2020). Jankovic, J. Parkinson's disease: clinical features and diagnosis. J. Neurol. Neurosurg. Psychiatry 79 , 368-376 (2008). De Lau, L. M. & Breteler, M. M. Epidemiology of Parkinson's disease. Lancet Neurol . 5 , 525-535 (2006). Gao, C. et al . Model-based and model-free machine learning techniques for diagnostic prediction and classification of clinical outcomes in Parkinson’s disease. Sci. Rep . 8 , 7129 (2018). Mirelman, A. et al . Detecting sensitive mobility features for Parkinson's disease stages via machine learning. Mov. Disord . 36 , 2144-2155 (2021). Rizzo, G. et al . Accuracy of clinical diagnosis of Parkinson disease: a systematic review and meta-analysis. Neurology 86 , 566-576 (2016). Brognara, L., Palumbo, P., Grimm, B. & Palmerini, L. Assessing gait in Parkinson’s disease using wearable motion sensors: a systematic review. Diseases 7 , 18 (2019). Del Din, S. et al . Gait analysis with wearables predicts conversion to parkinson disease. Ann. Neurol. 86 , 357-367 (2019). Bridenbaugh, S. A. & Kressig, R. W. Laboratory review: the role of gait analysis in seniors’ mobility and fall prevention. Gerontology 57 , 256-264 (2011). Buckley, C. et al . The role of movement analysis in diagnosing and monitoring neurodegenerative conditions: insights from gait and postural control. Brain Sci. 9 , 34 (2019). Daneault, J. F. et al . Accelerometer data collected with a minimum set of wearable sensors from subjects with Parkinson’s disease. Sci. Data 8 , 48 (2021). Evers, L. J. et al . Real-life gait performance as a digital biomarker for motor fluctuations: the Parkinson@ Home validation study. J. Med. Internet Res . 22 , e19068 (2020). Gadaleta, M. et al . Deep learning techniques for improving digital gait segmentation. Annu. Int. Conf. IEEE. Eng. Med. Biol. Soc. 2019 , 1834-1837 (2019). Juneau, P. et al . Amputee fall risk classification using machine learning and smartphone sensor data from 2-minute and 6-minute walk tests. Sensors 22 , 1749 (2022). Balke, B. A simple field test for the assessment of physical fitness. Rep Civ Aeromed Res Inst US . 63 , (1963). Fan, S. et al . Digital health technology combining wearable gait sensors and machine learning improve the accuracy in prediction of frailty. Front. Public Health 11 , 1169083 (2023). Mudge, S. & Stott, N. S. Timed walking tests correlate with daily step activity in persons with stroke. Arch. Phys. Med. Rehabil . 90 , 296-301 (2009). Daines, K. et al . Fall risk classification for people with lower extremity amputations using random forests and smartphone sensor features from a 6-minute walk test. PLoS One 16 , e0247574 (2021). Ireland, D. et al . Classification of movement of people with parkinsons disease using wearable inertial movement units and machine learning. Stud. Health Technol. Infor. 227 , 61-6 (2016). Fuentes-Abolafio, I. J. et al . Kinematic parameters related to functional capacity, fatigue, and breathlessness during the 6-min walk test in older adults with heart failure with preserved ejection fraction. Eur. J. Cardiovasc. Nurs. 23 , 69-80 (2023). Mänttäri, A. et al . Six‐minute walk test: a tool for predicting maximal aerobic power (VO 2 max) in healthy adults. Clin. Physiol. Funct. Imaging 38 , 1038-1045 (2018). Sperandio, E. F. et al . Reference values for the 6-min walk test in healthy middle-aged and older adults: from the total distance traveled to physiological responses. Fisioterapia em Movimento , 32 , (2019). Storm, F. A. et al . Wearable inertial sensors to assess gait during the 6-minute walk test: A systematic review. Sensors (Basel). 20 , 2660 (2020). Atri, R. et al . Deep learning for daily monitoring of Parkinson’s disease outside the clinic using wearable sensors. Sensors (Basel). 22 , 6831 (2022). Angelini, L. et al . Wearable sensors can reliably quantify gait alterations associated with disability in people with progressive multiple sclerosis in a clinical setting. J. Neurol . 267 , 2897-2909 (2020). Mei, J., Desrosiers, C. & Frasnelli, J. Machine learning for the diagnosis of Parkinson's disease: a review of literature. Front. Aging Neurosci . 13 , 633752 (2021). Alaskar, H. et al . A data science approach for reliable classification of neuro-degenerative diseases using gait patterns. J. Reliab. Intell. Environ . 6 , 233-247 (2020). Balaji, E., Brindha, D., Elumalai, V. K. & Umesh, K. Data-driven gait analysis for diagnosis and severity rating of Parkinson’s disease. Med. Eng. Phys . 91 , 54-64 (2021). Borzì, L., Sigcha, L., Rodríguez-Martín, D. & Olmo, G. Real-time detection of freezing of gait in Parkinson’s disease using multi-head convolutional neural networks and a single inertial sensor. Artif. Intell. Med. 135 , 102459 (2023). Juneau, P. Automated foot strike identification and fall risk classification for people with lower limb amputations using smartphone sensor signals from 2 and 6-minute walk tests (Doctoral dissertation, Université d'Ottawa/University of Ottawa). (2022). EI Maachi, I., Bilodeau, G. A. & Bouachir, W. Deep 1D-Convnet for accurate Parkinson disease detection and severity prediction from gait. Expert Syst. Appl. 143 , 113075 (2020). Yang, X., Ye, Q., Cai, G., Wang, Y. & Cai, G. PD-ResNet for classification of Parkinson’s disease from gait. IEE. J. Transl. Eng. Health Med. 10, 2200111 (2022). Ramanujam, E., Perumal, T., & Padmavathi, S. Human activity recognition with smartphone and wearable sensors using deep learning techniques: a review. IEEE Sens. J. 21 , 13029-13040 (2021). Nguyen, M. D. et al . IMU-based spectrogram approach with deep convolutional neural networks for gait classification. In 2020 IEEE International Conference on Consumer Electronics (ICCE) (pp. 1-6). IEEE. (2020). Güney, P. A. et al. Comparison of 1-, 3-, and 6-minute walk tests in the post-lung transplant period. South Clin. Ist. Euras. 33 , (2022). Hadouiri, N., Monnet, E., Gouelle, A., Sagawa Jr, Y. & Decavel, P. Locomotor strategy to perform 6-minute walk test in people with multiple sclerosis: a prospective observational study. Sensors (Basel). 23 , 3407 (2023). Bailo, G. et al . Characterization of walking in mild Parkinson’s disease: reliability, validity and discriminant ability of the six-minute walk test instrumented with a single inertial sensor. Sensors (Basel). 24 , 662 (2024). Chou, P. Y. & Lee, S. C. Turning deficits in people with Parkinson's disease. Tzu Chi Med. J. 25 , 200-202. (2013). Akram, S., Frank, J. S., & Jog, M. Parkinson's disease and segmental coordination during turning: I. Standing turns. Can. J. Neurol. Sci. 40 , 512-519 (2013). Ramesh, V., & Bilal, E. Detecting motor symptom fluctuations in Parkinson’s disease with generative adversarial networks. NPJ Digit. Med. 5 , 138 (2022). Choi, H. et al . Association between severity of freezing of gait and turning characteristics in people with Parkinson’s disease. Int. J. Environ. Res. Public Health 19 , 12131 (2022). Wu, Z. et al . Mild gait impairment and its potential diagnostic value in patients with early-stage Parkinson’s disease. Behav. Neurol. 2021 , 6696454 (2021). Lin, C. H. et al . Early detection of Parkinson’s disease by neural network models. IEEE Access 10 , 19033-19044 (2022). Meng, L. et al . Inertial-based gait metrics during turning improve the detection of early-stage parkinson’s disease patients. IEEE Trans. Neural. Syst. Rehabil. Eng. 31 , 1472-1482 (2023). Shcherbak, A., Kovalenko, E., & Somov, A. Detection and classification of early stages of Parkinson’s disease through wearable sensors and machine learning. IEEE Trans. Instrum. Meas. 72 , 1-9 (2023). Cai, G. et al . Specific distribution of digital gait biomarkers in Parkinson’s disease using body-worn sensors and machine learning. Gerontol. A. Biol. Sci. Med. Sci . 78, 1348-1354 (2023). Sveinbjornsdottir, S. The clinical symptoms of Parkinson's disease. J. Neurochem. 139 , 318-324 (2016). Buckley, C. et al . Upper body accelerations as a biomarker of gait impairment in the early stages of Parkinson’s disease. Gait Posture 71 , 289-295 (2019). Chandrabhatla, A. S., Pomeraniec, I. J. & Ksendzovsky, A. Co-evolution of machine learning and digital technologies to improve monitoring of Parkinson’s disease motor symptoms. NPJ Digit. Med. 5 , 32 (2022). Palmerini, L., Mellone, S., Avanzolini, G., Valzania, F. & Chiari, L. Quantification of motor impairment in Parkinson's disease using an instrumented timed up and go test. IEEE Trans. Neural. Syst. Rehabil. Eng. 21 , 664-673 (2013). Bohannon, R. W. 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 . 7 , 269 (2014). Zhang, H., Deng, K., Li, H., Albin, R. L. & Guan, Y. Deep learning identifies digital biomarkers for self-reported Parkinson's disease. Patterns. (N. Y.) 1 :100042 (2020). Vanmechelen, I. et al . Assessment of movement disorders using wearable sensors during upper limb tasks: a scoping review. Front. Robot. AI 9 , 1068413 (2023). Hughes, A. J., Daniel, S. E. & Lees, A. J. Improved accuracy of clinical diagnosis of Lewy body Parkinson’s disease. Neurology 57 , 1497-1499 (2001). Hoehn, M.M. & Yahr, M.D. Parkinsonism: onset, progression, and mortality. Neurology 17 , 427-442 (1967). Goetz, C. G. et al . Movement Disorder Society Task Force report on the Hoehn and Yahr staging scale: status and recommendations the Movement Disorder Society Task Force on rating scales for Parkinson's disease. Mov. Disord . 19 , 1020-1028 (2004). Folstein, M. F., Folstein, S. E. & McHugh, P. R. “Mini-mental state”: a practical method for grading the cognitive state of patients for the clinician. J. Psychiatr. Res . 12 , 189-198 (1975). Nasreddine, Z. S. et al . The Montreal Cognitive Assessment, MoCA: a brief screening tool for mild cognitive impairment. J. Am. Geriatr. Soc . 53 , 695-699 (2005). Wang, Z., Qin, M., & Chen, Y. K. Learning from the cnn-based compressed domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (pp. 3582-3590). (2022). Eckmann, J. P., Kamphorst, S. O. & Ruelle, D. Recurrence plots of dynamical systems. EPL , 4 , 973-977 (1987). Tank, V. H. et al . Drug eluting stents versus bare metal stents for the treatment of extracranial vertebral artery disease: a meta-analysis. J. Neurointerv. Surg . 8 , 770-774 (2016). Huang, G. et al . Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4700-4708). (2017). Iandola, F. N. et al . SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and< 0.5 MB model size. arXiv preprint arXiv:1602.07360 (2016). 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 . 13 , 12638 (2023). Bernardo, L. S., Damaševičius, R., Ling, S. H., de Albuquerque, V. H. C. & Tavares, J. M. R. Modified squeezenet architecture for parkinson’s disease detection based on keypress data. Biomedicines 10 , 2746 (2022). Table Table 1. Physical and clinical characteristics of all 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 - - 6MWT (m) 414.24 ± 85.16 494.23 ± 51.42 <0.001 c The data presented are 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; 6MWT: 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. Additional Declarations No competing interests reported. Supplementary Files SupplementaryTable1.docx Cite Share Download PDF Status: Published Journal Publication published 30 Sep, 2024 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 17 Jul, 2024 Reviews received at journal 09 Jul, 2024 Reviewers agreed at journal 08 Jul, 2024 Reviews received at journal 15 Jun, 2024 Reviewers agreed at journal 06 Jun, 2024 Reviewers agreed at journal 06 Jun, 2024 Reviewers invited by journal 03 Jun, 2024 Editor assigned by journal 03 Jun, 2024 Editor invited by journal 28 May, 2024 Submission checks completed at journal 28 May, 2024 First submitted to journal 27 May, 2024 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-4482534","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":312171995,"identity":"72c40cf8-9264-4c46-bca3-71ad7f4a428a","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":312171996,"identity":"9c5fb84d-122a-4cf4-afaf-5d01843d0bca","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":312171997,"identity":"aa35c2d3-3503-436a-9e06-3ec01e19675b","order_by":2,"name":"Hwayoung Park","email":"","orcid":"","institution":"Dong-A University","correspondingAuthor":false,"prefix":"","firstName":"Hwayoung","middleName":"","lastName":"Park","suffix":""},{"id":312171998,"identity":"bbc20296-8794-462b-a324-7e08aec4ed90","order_by":3,"name":"Bohyun Kim","email":"","orcid":"","institution":"The Graduate School of Dong-A University","correspondingAuthor":false,"prefix":"","firstName":"Bohyun","middleName":"","lastName":"Kim","suffix":""},{"id":312171999,"identity":"485aa816-eb8b-48b9-b0f0-76bbd9efa5c4","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":312172000,"identity":"cc92d571-4c9b-4bb8-b787-5be72f38c865","order_by":5,"name":"Sang-Myung Cheon","email":"","orcid":"","institution":"Dong-A University","correspondingAuthor":false,"prefix":"","firstName":"Sang-Myung","middleName":"","lastName":"Cheon","suffix":""},{"id":312172001,"identity":"60c228b3-952f-43e3-86a6-0865ed4e709b","order_by":6,"name":"Sungtae Shin","email":"","orcid":"","institution":"Dong-A University","correspondingAuthor":false,"prefix":"","firstName":"Sungtae","middleName":"","lastName":"Shin","suffix":""}],"badges":[],"createdAt":"2024-05-27 05:51:08","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4482534/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4482534/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-024-72648-w","type":"published","date":"2024-09-30T15:57:16+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":58229009,"identity":"5f9d093b-867c-4824-9c88-457be15173da","added_by":"auto","created_at":"2024-06-12 19:13:39","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1698687,"visible":true,"origin":"","legend":"\u003cp\u003eThe accuracy of CNN models for classifying individuals with early PDs and controls using time-series images of acceleration and gyroscope data of wearable sensors (referred to as the highest accuracy for each body segment, bold). The rows in each table represent the three CNN architectures (ResNet, DenseNet, SqueezeNet) and the columns represent the variables. CNN: convolutional neural network; PD: Parkinson’s disease; Acc: acceleration; Gyr: gyroscope; XYZ: resultant values of acceleration and gyroscope data; ResNet: residual neural network; DenseNet: dense convolutional network.\u003c/p\u003e","description":"","filename":"floatimage1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4482534/v1/3dc7ba9f2c303c11f648f07a.jpg"},{"id":58230717,"identity":"b4f75b4e-cf15-4f9f-afae-95b5cc32dcc2","added_by":"auto","created_at":"2024-06-12 19:21:39","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":331292,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrices for the variables with the highest accuracy in body segments. Acc: acceleration; Gyr: gyroscope; PD: Parkinson’s disease; Cons: controls.\u003c/p\u003e","description":"","filename":"floatimage2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4482534/v1/1b55fb844d7ed07899ed77e6.jpg"},{"id":58229007,"identity":"146b3ba6-b596-428a-86d0-776e63fcc8b9","added_by":"auto","created_at":"2024-06-12 19:13:39","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":706936,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative images and corresponding frequency patterns of early PD and control participants. The images and graphs depict the last phase of the 6MWT, including straight walking and turning for the variables that showed the highest accuracy in body segments. 6MWT: 6-min walk test; PD: Parkinson’s disease; Con: control; Acc: acceleration; Gyr: gyroscope; XYZ: resultant values of acceleration and gyroscope data.\u003c/p\u003e","description":"","filename":"floatimage3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4482534/v1/e912602dbde81656202c1caa.jpg"},{"id":58229008,"identity":"632e9583-126c-44b7-bce3-d3fba25f3027","added_by":"auto","created_at":"2024-06-12 19:13:39","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":628275,"visible":true,"origin":"","legend":"\u003cp\u003eStudy flowchart. PD: Parkinson’s disease.\u003c/p\u003e","description":"","filename":"floatimage4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4482534/v1/aac9b37c61664278740108b2.jpg"},{"id":58229006,"identity":"637eecf4-3c60-4d60-a3cd-cd0ff9b26171","added_by":"auto","created_at":"2024-06-12 19:13:39","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":472975,"visible":true,"origin":"","legend":"\u003cp\u003eExperimental setup.\u003cstrong\u003e \u003c/strong\u003e(\u003cstrong\u003ea\u003c/strong\u003e) Placement of wearable sensors and coordinate system. A 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). The orientation coordinate system is represented in the lower right. (\u003cstrong\u003eb\u003c/strong\u003e) Six-minute walk test. The main task of walking back and forth on a 20-m course without running in the preferred direction for 6 minutes. VT: vertical; ML: mediolateral; AP: anteroposterior.\u003c/p\u003e","description":"","filename":"floatimage5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4482534/v1/006ebb74d8e48a878a64abc6.jpg"},{"id":58229011,"identity":"1c231c62-048e-4918-a2e5-41674c0f7b0b","added_by":"auto","created_at":"2024-06-12 19:13:39","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1297534,"visible":true,"origin":"","legend":"\u003cp\u003eFramework for classifying early PDs and controls based on wearable sensors used in this work. PD: Parkinson’s disease; Cons: controls; Rec: Recurrence Plot; CV: cross-validation; ResNet: residual neural network; DenseNet: dense convolutional network.\u003c/p\u003e","description":"","filename":"floatimage6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4482534/v1/a6a8e546da0fdcb839dcd0a6.jpg"},{"id":66096765,"identity":"ed5d846a-bbed-4138-976b-95734b9aca28","added_by":"auto","created_at":"2024-10-07 16:09:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5843424,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4482534/v1/c40848d2-9ee1-4aa8-a45a-c7d92be604c4.pdf"},{"id":58229012,"identity":"ec5c82b7-a5a7-4939-b21c-8f539f81b0a8","added_by":"auto","created_at":"2024-06-12 19:13:39","extension":"docx","order_by":9,"title":"","display":"","copyAsset":false,"role":"supplement","size":48958,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTable1.docx","url":"https://assets-eu.researchsquare.com/files/rs-4482534/v1/0f81c4787306b2dc452fb338.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Convolutional neural network-based detection of early-stage Parkinson’s disease using the six-minute walk test","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eParkinson\u0026rsquo;s disease (PD) is a chronic, degenerative neurological disorder that affects movements [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], and the characteristic symptoms include increasing difficulty walking, speaking, or completing other simple tasks as well as a combination of motor and non-motor symptoms that can confer functional disability, loss of independence, and decreased quality of life [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. The heterogeneity of PD generates significant challenges for accurate diagnosis, especially in early-stage PD that may include very subtle symptoms [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe severity of motor symptoms in PD is ascertained with the Hoehn and Yahr (H\u0026amp;Y) Scale, Movement Disorders Society (MDS)-sponsored revision of the Unified Parkinson\u0026rsquo;s Disease Rating Scale (UPDRS), MDS-UPDRS Part III, PD phenotype assessment, and discrete variable-based analysis of physical and fitness factors as well as postural and gait characteristics [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. However, these subjective assessments are scored through clinical observation by neurospecialists [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e] and, therefore, research on the objective, accurate body-part-specific motor-symptom assessment and symptom impact-based classification of disease severity is needed [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eGait in PD is a potential status and trait marker, because gait impairments appear in early-stage PD, precede the appearance of overt motor signs, and progress more rapidly than other motor characteristics of PD [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Objective quantification of individual gait characteristics may include three-dimensional (3D) motion capture, force plates, instrumented walkways, treadmills, and electromyography [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], and these tools are essential for accurately characterizing PD gait disorders in clinical populations. However, the application of these tools is primarily constrained to research environments owing to the cost and the need for specialized expertise for utilization [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Thus, monitoring gait outcomes with wearable sensors may be a useful, inexpensive alternative option for ascertaining the presence of impairment in a controlled daily living environment [\u003cspan additionalcitationids=\"CR13\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe 6-min walk test (6MWT) is a submaximal movement assessment that is routinely used to assess an individual\u0026rsquo;s aerobic capacity and physical mobility [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Originally developed for use in patients with chronic respiratory or cardiovascular disease, the 6MWT is currently used to assess diverse populations, including older adults and people who have had a stroke, lower extremity amputations, or PD [\u003cspan additionalcitationids=\"CR18 CR19\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e]. The 6MWT constitutes a simple, inexpensive (minimal setup and space), safe, reproducible, and alternative functional test [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. Despite the limitation that the standard clinical outcome of the traditional 6MWT only comprises the distance walked in 6 min, the distance walked shows a moderate correlation with more complex aerobic capacity tests, such as V̇O\u003csub\u003e2\u003c/sub\u003e max (maximal oxygen uptake) and, thus, can minimize the burden on patients and clinicians [22,23]. Furthermore, compared to other gait tests, the well-defined settings and longer duration of the 6MWT constitute favorable conditions for obtaining useful information from both turning and straight gaits, and this potentially increases the test efficiency [24]. Therefore, the 6MWT conducted using wearable sensors appears clinically applicable and comprehensive, and is a quick, easily applicable method that provides objective details of gait quality [25,26].\u003c/p\u003e \u003cp\u003eThe recent trend of the increased use of artificial intelligence-based machine learning and deep learning models to classify and predict diseases [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] has advanced research to resolve the problem of motor-symptom severity classification with the comprehensive gait characteristics of people with PD, such as by using a machine learning algorithm for distinguishing three types of neurodegenerative diseases (PD, Huntington\u0026rsquo;s disease, and amyotrophic lateral sclerosis) [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e] and diagnosing and assessing PD severity according to the H\u0026amp;Y scale [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e]. Borz\u0026igrave; et al. reported neural network-based freezing of gait detection from the 6MWT in PD by using the waist sensor-derived acceleration data [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Juneau and Daines et al. used a random forest model to classify the fall risk through accelerometer- and gyroscope-derived data that were recorded using a smartphone placed over the posterior pelvis during the 6MWT in patients with lower limb amputations [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. However, these studies have limitations such as a small sample size; low accuracy due to the analysis of movement data of specific body segments based on a single wearable sensor without objective, specific identification of body segments according to disease severity; and the use of unproductive, inefficient, and less sensitive tools [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eComparative analysis of passive signal classification and gait parameter extraction from wearable sensors for the detection and prediction of gait-disorder signals in PD through gait assessment has limited utility in the identification of subtle motor symptomatic disorders in early- to mid-stage PD [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Thus, the interest in deep learning technologies, which are automated approaches for objective diagnosis, quantitative assessment, and accurate predictive disease progression modelling, to overcome these limitations has increased recently [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. EI Maachi et al. and Yang et al. proposed an intelligent gait-based PD-detection system that used a convolutional neural network (CNN) [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. In a wearable sensor-based study, Nguyen et al. classified abnormal gait in musculoskeletal disorders through a CNN-derived sensor-based spectrogram approach [\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. For objective gait-disorder signal detection, we believe that accuracy could be improved by using time-series (TS) gait patterns and classifying them into CNN architecture-converted visual images [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e]. However, studies specifically related to CNN-based 6MWT identification to categorize early-stage PD are limited.\u003c/p\u003e \u003cp\u003eTherefore, this study aimed to identify segments and components that exhibit optimal accuracy for classifying people with early-stage PD and healthy controls using a CNN algorithm for sensor-based 6MWT data. The research hypothesis is as follows: First, CNN algorithms can improve the classification of early-stage PD and healthy controls, wherein TS signals of accelerometer and gyroscope-based gait patterns obtained using six segmentally mounted wearable sensors are converted into images to determine the classifiability of the two groups. Second, based on the acceleration and gyroscope data obtained from the wearable sensors, the characteristics of early-stage PD that are prominent in specific components of the six segments can be identified.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eCNN classification results for the early PD and control groups\u003c/h2\u003e \u003cp\u003eWe visualized triaxial accelerations and gyroscope data using TS images during the 6MWT to identify the most prominent gait variables for classification into two groups according to early PDs and controls. The proposed algorithms, such as Recurrence Plot (Rec), are based on CNNs.\u003c/p\u003e \u003cp\u003eThe accuracy results for all variables are summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, with an accuracy range of 49.5\u0026ndash;83.5%. The variables with the highest accuracy in each body segment were the \u003cem\u003ez\u003c/em\u003e-axis gyroscope data (Gyr_Z) on the left arm segment at 76.6%, the i-axis acceleration (Acc_Z) for the right arm segment at 70.4%, the acceleration resultant value of the \u003cem\u003ex-\u003c/em\u003e, \u003cem\u003ey-\u003c/em\u003e, and \u003cem\u003ez\u003c/em\u003e-axes (Acc_XYZ) for the left thigh at 78.4%, and the \u003cem\u003ey\u003c/em\u003e-axis acceleration (Acc_Y) for the right thigh at 79.5% using the DenseNet model. Using the SqueezeNet model, the \u003cem\u003ex\u003c/em\u003e-axis gyroscope data (Gyr_X) for the thoracic spine segment demonstrated an accuracy of 83.1%, and the \u003cem\u003ex\u003c/em\u003e-axis gyroscope data (Gyr_X) for the lumbar spine segment demonstrated an accuracy of 83.5%. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the confusion matrices for the variables with the highest accuracy in body segments. We present representative images and corresponding frequency patterns for an early PDs and controls for the variables that showed the highest accuracy (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The representative images and graphs show the last phase of the 6MWT, including straight walking and turning. The mean accuracy, precision, recall, and F1 scores for all gait variables for classification between early PDs and controls are shown in Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003eThe primary finding of the present study is that, in the CNN-algorithm and sensor-based 6MWT-derived classification of early PDs and healthy controls, the gyroscopic vertical component of the lumbar spine (Gyr_X) has the highest classification accuracy (83.5%), followed by the gyroscopic vertical component of the thoracic spine (Gyr_X; 83.1%), the acceleration mediolateral component of the right thigh (Acc_Y; 79.5%), the acceleration resultant component of the left thigh (Acc_XYZ; 78.4%), the gyroscopic anteroposterior component of the left arm (Gyr_Z; 76.6%), and the acceleration anteroposterior component of the right arm (Acc_Z; 70.4%).\u003c/p\u003e \u003cp\u003eRecent studies have harnessed the power of the 6MWT method as a promising tool in gait analysis by using wearable-sensor technology and have explored its potential applicability for investigating the observability and classifiability of continuous and segment-specific gait patterns over time in diverse populations, including those with lung transplants [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e], multiple sclerosis [\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e], lower-limb amputations [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], and PD [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. Despite the variations in disease groups, the 6MWT has promising potential for detecting a decline in motor function and facilitating the management of prolonged gait. This opens exciting possibilities for further studies to determine the accuracy of the 6MWT in the early identification and severity determination of PD [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. These studies are crucial to advance our understanding and application of the 6MWT.\u003c/p\u003e \u003cp\u003eBorz\u0026igrave; et al. used waist-wearable sensors and multi-head CNN techniques for the detection and prediction of freezing of gait (FOG) during a 6MWT in PD and achieved a specificity of 100% in healthy older adult participants [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Bailo et al. extracted gait domain-related metrics, such as rhythm and pace, regularity, intensity, dynamic instability, and asymmetry from lower truncal accelerations and angular velocities, during the 6MWT in participants with mild PD, and reported significant differences with healthy controls [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. However, none of these studies have focused on classifying early-stage PD using CNN algorithms and components of sensor-based triaxial accelerations and gyroscopic data obtained from the 6MWT. The novelty of our study lies in the use of wearable sensors and CNN algorithms to classify early-stage PD using the 6MWT, which may have significant implications in clinical and daily-life applications. We fitted wearable sensors to each of the six segments during the 6MWT and used a CNN algorithm to determine the classification possibilities of the two groups.\u003c/p\u003e \u003cp\u003eAxial rigidity, increased postural tone, \u003cem\u003een bloc\u003c/em\u003e turning, FOG, and loss of intersegmental flexibility may be evident during gait in PD, particularly in walking and turning gait [\u003cspan additionalcitationids=\"CR40 CR41\" citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]. Specifically, individuals with early-stage PD experience a reduced range of motion at the hip joint [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e], asymmetric tremors, decreased coordination [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e], and reduced postural stability control [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e] during gait, which exacerbate with progression to moderate PD [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. However, in early-stage PD, as the symptoms are not well differentiated and can easily be confused with other conditions, further investigation is needed [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e, \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eCai et al. suggested that asymmetry and axial features of the trunk and lumbar spine contribute to the early PD classification model for discriminating healthy controls from individuals with H\u0026amp;Y Stage 1 limitation [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. During walking, people with early-stage PD can exhibit significantly impaired asymmetry because of unilateral parkinsonism-related symptoms, with consequent axial impairments [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e]. In this study, early-stage PD was classified through six body segments at H\u0026amp;Y stages 1 and 2, with similar classification accuracy for the affected and unaffected limb segments, and the highest accuracy in the lumbar spine and truncal segments, and this is consistent with the results from previous studies. Nonetheless, it is difficult to generalize our results because of the limited evaluability of a single waist sensor [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Therefore, asymmetric and unilateral measurements, including of axial features, should be used whenever possible to detect early-stage PD in clinical and community settings.\u003c/p\u003e \u003cp\u003eDespite the asymmetrical basal ganglia degeneration in early-stage PD that may lead to unilateral symptoms, measurement of gait variation between the left and right sides may not be reliable for quantifying motor asymmetry in patients [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. Furthermore, as the side on which symptoms will occur is unclear, a comparative analysis of bilateral-limb sensor data is needed [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. Buckley et al. proposed that upper-body acceleration may be a biomarker for gait impairment in early-stage PD [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. In combination with spatiotemporal information, variables of upper body acceleration contribute to a better description of the PD gait [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e] and could be associated with a reduced ability to regulate repetitive steps and strides during gait or control the rhythmic displacements of the upper body during walking [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eWe found that specific components of the limb segments, such as the mediolateral component for the right thigh position, had high classification accuracy. Research on symmetry in the walking domain indicated that measuring gait symmetry in the mediolateral and vertical directions was significantly associated with walking endurance (6MWT), whereas no relationship was found for balance scales; this suggests that gait symmetry may be primarily associated with the energy efficiency of locomotion, rather than balance [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSimilarly, our results showed that the gyroscopic vertical (Gyr_X) component demonstrated the highest discrimination accuracy for the lumbar spine (83.5%), followed by the thoracic spine (83.1%). In particular, increased rigidity and signs of akinesia and/or bradykinesia were associated with smaller truncal sway, mainly in the anteroposterior and vertical directions [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e]. Regarding gait and balance, the lower truncal acceleration amplitude was associated with lower walking endurance and poorer dynamic balance; therefore, reduced truncal acceleration may result from the minimizing of upper body movements during walking to compensate for lower-limb impairment and maintain balance [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe strengths of our work are that we identified the characteristics of early-stage PD and classification accuracy that were prominent in individual axial components, specifically in each of the six segments. We determined that considering the three-axis factors together may help identify and pre-screen for gait characteristics that indicate early-stage PD.\u003c/p\u003e \u003cp\u003eNevertheless, there are some limitations to this study. We conducted our study in the \u0026ldquo;medication ON state.\u0026rdquo; Considering both ON and OFF medication states may enable a better understanding of the characteristics of early-stage PD. For this, a larger dataset is needed to validate the effectiveness of the model. Future studies should aim to propose a measurement time for the 6MWT, considering measurement efficiency, such as for 2 or 3 min, by using specific intervals (e.g., 15 s) [\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e, \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e]. Our study involved a CNN analysis of the entire 6 minutes, but without distinguishing between straight and turning gaits in the 6MWT, and merely identified schematic frequency patterns for straight and turning gaits. Therefore, in the future, it is necessary to extract data specific to turning and straight gait sections and compare the classification performance using deep learning and machine learning techniques [\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e]. Furthermore, further analysis of specific biomarkers, such as amplitude and magnitude based on wearable sensors, could go further than the identification of the best-performing body segments [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTo additionally compare detailed gait characteristics, we identified frequency patterns for the last phase of the 6MWT, including straight and turning gaits, for early PDs and controls (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). We found the highest accuracy in the Gyr_X component, especially in the lumbar and thoracic spine, where characteristics such as the differences in frequency patterns on the graph during turning and straightforward walking between early PDs and controls (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) on the gyroscope data of the \u003cem\u003ex\u003c/em\u003e-axis may have influenced the image generation and affected the accuracy. In future work, it is important to extract relevant features that can identify straight and turning gait patterns in the 6MWT from the domains of bilateral coordination, asymmetry, truncal symptom severity, and variability by performing non-linear analysis, and then use machine learning techniques to analyze specific biomarkers associated with early-stage PD [\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e]. Finally, we only considered patients with early H\u0026amp;Y-stage PD; however, patients with severe symptoms may differ from those analyzed herein. Future studies should include more subjects with intermediate and tremor-dominant subtypes.\u003c/p\u003e"},{"header":"CONCLUSION","content":"\u003cp\u003eIn conclusion, our study confirms the possibility of classifying two groups when converting TS data of 6MWT into images, using CNN algorithms, and identifying which segments and components exhibit optimal accuracy. We suggest that 2D imaging for analyzing infinite-length 1D TS data may provide an opportunity to screen the gait patterns of patients with early-stage PD for primary classification. We believe that this work will not only improve the accuracy of motor-symptom-severity assessment and classification but also contribute to the improvement of time-consuming, economical, and favorable experimental and analytical procedures, enable early disease detection in clinical settings, and the availability of objective indicators for disease-severity assessment [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e"},{"header":"METHODS","content":"\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n \u003ch2\u003eParticipants\u003c/h2\u003e\n \u003cp\u003eA total of 79 individuals with PD (early PDs) and 50 age-matched healthy controls (controls) participated in the study and were included in the final analysis. Figure \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e depicts a flowchart of the study process and participant details. Based on the United Kingdom\u0026rsquo;s Parkinson\u0026apos;s Disease Society Brain Bank [\u003cspan class=\"CitationRef\"\u003e55\u003c/span\u003e] criteria, a neurospecialist diagnosed patients with idiopathic PD. The inclusion criteria were as follows: (a) age 40\u0026ndash;85 years, (b) ability to walk and move independently without assistive devices, with a modified H\u0026amp;Y stage of 1\u0026ndash;2 [\u003cspan class=\"CitationRef\"\u003e56\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e57\u003c/span\u003e], (c) Mini-Mental State Examination (MMSE) score\u0026thinsp;\u0026ge;\u0026thinsp;24 [\u003cspan class=\"CitationRef\"\u003e58\u003c/span\u003e], and (d) stable response to anti-Parkinson\u0026rsquo;s medications. Only one participant who dropped out during the 6MWT was excluded from the analysis. Individuals with a history of cardiovascular, musculoskeletal, vestibular, or other neurological diseases, who required mobility-assistive devices, and with pharmacotherapy-refractory dyskinesia were excluded. The controls included individuals with no medical history of orthopedic surgery or cognitive impairment and gait disturbance in the past 6 months. The physical and clinical characteristics of the participants are shown in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab1\" border=\"1\"\u003e\u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eAll experiments were performed in accordance with the relevant guidelines and regulations. The study protocol was approved by the Institutional Review Board (IRB) of Dong-A University Medical Center (IRB number: DAUHIRB-22-089). All participants provided written informed consent before participating in this study. The study is registered in the Clinical Research Information Service in the Republic of Korea (KCT0009353).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003eExperimental procedures\u003c/h2\u003e\n \u003cp\u003eParticipants performed two sessions. In the first session, the participants completed the informed consent form and were assessed by using the MMSE, Montreal Cognitive Assessment (MoCA) [\u003cspan class=\"CitationRef\"\u003e59\u003c/span\u003e], UPDRS, and modified H\u0026amp;Y scale, and the physical characteristics such as sex, age, height, body mass, and body mass index were measured (Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e). In the second session, before performing the main task, wearable sensors were attached to the left and right upper arms (5 cm above the lateral humeral epicondyle), thighs (10 cm above the lateral femoral epicondyle), thoracic spine (T10), and lumbar spine (center of the left and right posterior superior iliac spines) region using a stretchable belt (Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003ea). Participants performed the 6MWT as the main task, wherein they were instructed to walk, without running, as far as possible in the preferred direction for 6 min. The 20-m course was demarcated by cones placed at each end (Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003eb). The raters walked behind participants to ensure their safety, offer encouragement, and let them know how much time they had left. All experiments were performed in the \u0026ldquo;On\u0026rdquo; medication state approximately 2 h after taking the medication.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n \u003ch2\u003eFramework for classifying early PDs and controls\u003c/h2\u003e\n \u003cp\u003eThe framework for sensor-based classification of early PDs and controls in this study consists of four major process steps (Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e). The overall flow was designed to acquire raw data from wearable sensors, perform data pre-processing, categorize the sample data into the training and test sets, and then train the CNN to evaluate the model\u0026apos;s performance.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n \u003ch2\u003eData acquisition\u003c/h2\u003e\n \u003cp\u003eData were collected using an Xsens DOT Wearable Sensor platform (Movella Technologies, Enschede, the Netherlands) with a package size of 36.30 \u0026times; 30.35 \u0026times; 10.80 mm and a weight of 11.2 g. Signal-based TS gait characteristics were extracted from each wearable sensor (accelerometer and gyroscope) in each axis (\u003cem\u003ex\u003c/em\u003e-axis: vertical, \u003cem\u003ey\u003c/em\u003e-axis: mediolateral, \u003cem\u003ez\u003c/em\u003e-axis: anteroposterior; Fig. \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003ea). We collected the data on 3D-acceleration (m/s\u003csup\u003e2\u003c/sup\u003e; up to \u0026plusmn;\u0026thinsp;16 g) and 3D angular velocity (\u0026deg;/s; up to \u0026plusmn;\u0026thinsp;2000\u0026deg;/s) via the accelerometer and gyroscope, respectively, in the wearable sensor. By default, the local Earth-fixed reference coordinate system (L) used is defined as a right-handed Cartesian coordinate system (East-North-Up [ENU]) with X positive to the East (E), Y positive to the North (N), and Z positive when pointing up (U) and is the standard reference in inertial navigation for aviation and geodetic applications. With the default ENU (L) coordinate system, the X-sens yaw output is defined as the angle between East (X) and the horizontal projection of the sensor \u003cem\u003ex\u003c/em\u003e-axis, positive about the local vertical axis (Z) following the righthand rule. The orientation calculated by X-sens DOT is the orientation of the sensor coordinate system (S) with regard to L. The orientation output is in the ENU frame: accordingly, the roll, pitch, and yaw will equal 0 degrees when the \u003cem\u003ex\u003c/em\u003e-, \u003cem\u003ey\u003c/em\u003e-, and \u003cem\u003ez\u003c/em\u003e-axes of the sensor are aligned with the east, north, and upward directions, respectively. Thus, X, Y, and Z are positive when pointing to the local magnetic east, the local magnetic north, and upward, respectively. The data output rate was set to 60 Hz. Three-dimensional data were collected using an iPad (iOS 15.6.1; Apple, Cupertino, CA) with the MovellaDOT app using Bluetooth 5.0 communication and analyzed using MATLAB R2023a (MathWorks, Natick, MA).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003ch2\u003eData pre-processing\u003c/h2\u003e\n \u003cp\u003eBefore the data pre-processing, we checked the graphs of the total 6 min of raw data from each of the six sensors. Next, the following data pre-processing steps were carried out. First, we categorized 6 min (360 s) of TS data from each participant\u0026rsquo;s six sensors into 15-s segments that represent one phase of the 24 data phases, including straight and turning gait. Second, we normalized the data to eliminate the effect of the difference in participants\u0026rsquo; height and sensor-placement variations following reattachment. The normalization procedure involved centering the sensor data within each trial to achieve a median of 0 while retaining the original scale of the sensor data. The median value of the data was adjusted to align with the 0-reference point. Lastly, we set the length of the sample to 896 to create a single image. The length of the 896 samples is proportional to the pixel size. We multiplied the pixel size of the input images (224 pixels) by an even number factor (4) for the sample in each data phase for all participants to obtain a sample length of 224 pixels \u0026times; 4\u0026thinsp;=\u0026thinsp;896 (approximately 9 s). This process constitutes the standardized approach to effectively construct a CNN model and train it with an optimized resolution image [\u003cspan class=\"CitationRef\"\u003e60\u003c/span\u003e]. For trials lasting less than 8.96 s, we applied zero padding.\u003c/p\u003e\n \u003cp\u003eThe resulting values of the three axes, as well as the acceleration and gyroscope data for each of the \u003cem\u003ex-, y-\u003c/em\u003e, and \u003cem\u003ez-\u003c/em\u003eaxes, were used to analyze the TS data and their components were calculated with resultant acceleration and gyroscope data according to the following formula:\u003c/p\u003e\n \u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e$$Resultant data=sqrt(x^2+y^2+z^2)$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eData pre-processing was performed using MATLAB.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n \u003ch2\u003eData generation\u003c/h2\u003e\n \u003cp\u003eFor training and testing CNN models using the accumulated imaging dataset of all phases, the dataset was randomly divided, per subject, into the training, validation, and test sets in a ratio of 50%, 20%, and 30%, respectively. The dataset initially consisted of 1776 samples for early PDs and 1056 samples for controls across six sensors for each segment (left arm, right arm, left thigh, right thigh, thoracic spine, and lumbar spine). Therefore, the unbalanced dataset was randomly oversampled using the imbalanced-learn (version 0.10.1) Python package to perform an unbiased analysis, and this generated a total of 3552 samples for each body segment, with a sample count of 1776 for both early PDs and controls, in the entire dataset.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\n \u003ch2\u003eTraining models and classification\u003c/h2\u003e\n \u003cp\u003eThe deep learning classification analysis based on the CNN models and TS imaging methods was performed as follows. For the variance analysis, fivefold cross-validation was employed. The batch size used in this experiment was 128; Adam optimizers were used, and the learning rate was used for the optimizer of 1e-05. After 150 epochs, the proposed model reaches saturation, and training is stopped with the help of early stopping. To overcome overfitting, only the best model was saved; this meant that, during the training phase, if the validation accuracy of the epoch was higher than the highest accuracy, then the model was saved.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n \u003ch2\u003eTS imaging methods\u003c/h2\u003e\n \u003cp\u003eThe perspective of the data was broadened from one-dimensional (1D) TS to two-dimensional (2D) images to analyze the temporal aspect of the gait data processed using TS imaging methods, specifically Recurrence Plot (Rec) [\u003cspan class=\"CitationRef\"\u003e61\u003c/span\u003e]. This approach offers two distinct advantages. First, it augments the dimensionality of the feature space, thereby enabling a more comprehensive examination of the gait data. Second, it leverages established CNN-based deep learning architectures for classification. The acceleration and gyroscope data of 24 phases (6 minutes divided into 15-s intervals) were used to analyze the TS data. After converting TS data from each phase to a matrix form, the converted TS data were generated in a new imaging, as described further in the subsections on the TS imaging methods.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\n \u003ch2\u003eRec plot\u003c/h2\u003e\n \u003cp\u003eThe Rec is a method employed to transform 1D TS into 2D images by representing recurring states of the TS [\u003cspan class=\"CitationRef\"\u003e61\u003c/span\u003e]. Given a TS \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(X=({x}_{1}, {x}_{2}, \\cdots , {x}_{n})\\)\u003c/span\u003e\u003c/span\u003e with a length of \u003cem\u003en\u003c/em\u003e, the trajectory at a discrete time point \u003cem\u003ei\u003c/em\u003e is define d as follows:\u003c/p\u003e\n \u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e$${\\dot{x}}_{i}=\\left({x}_{i},{x}_{i+r},\\cdots ,{x}_{i+\\left(m-1\\right)r}\\right),\\forall i\\in \\left\\{1,\\cdots ,n-\\left(m-1\\right)\\tau \\right\\}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eHere, \u003cem\u003em\u003c/em\u003e represents the dimension of the trajectories, and \u003cem\u003e\u0026tau;\u003c/em\u003e is the time delay. Using these trajectories, the binarized recurrence matrix \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\widehat{R}}_{i,j}\\)\u003c/span\u003e\u003c/span\u003e, with \u003cem\u003ei\u003c/em\u003e and \u003cem\u003ej\u003c/em\u003e as discrete timepoints, is calculated as the pairwise distance between the trajectories, as shown below:\u003c/p\u003e\n \u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e$${\\widehat{R}}_{i,j}=\\theta \\left(\\epsilon -‖{\\dot{x}}_{i}-{\\dot{x}}_{j}‖\\right),\\forall i,j\\in \\left\\{1,\\cdots ,n-(m-1)\\tau \\right\\}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eWhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(‖ \\bullet ‖\\)\u003c/span\u003e\u003c/span\u003erepresents a norm operation, \u003cem\u003e\u0026theta;\u003c/em\u003e is the Heaviside function, and \u003cem\u003e\u0026epsilon;\u003c/em\u003e is the recurrence threshold. One limitation of this approach is that the calculated values are discretized. To calculate the continuous matrix elements of the recurrence matrix, the Heaviside function and recurrence threshold are excluded from the equation, resulting in the continuous-valued recurrence matrix \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\widehat{R}}_{i,j}\\)\u003c/span\u003e\u003c/span\u003e, which is expressed as:\u003c/p\u003e\n \u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e$${R}_{i,j}=‖{\\dot{x}}_{i}-{\\dot{x}}_{j}‖,\\forall i,j\\in \\left\\{1,\\cdots ,n-(m-1)\\tau \\right\\}$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eIn our data analysis, we utilized the continuous recurrence matrix. We simplified the calculation by setting the dimension of the trajectories m to 1 and the time delay \u003cem\u003e\u0026tau;\u003c/em\u003e to 1 when converting the TS series data of gait into 2D images.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\n \u003ch2\u003eCNNs\u003c/h2\u003e\n \u003cp\u003eFor the classification of early PDs and controls using the processed TS images, three CNN architectures were employed in the analysis: Residual Neural Network (ResNet) [\u003cspan class=\"CitationRef\"\u003e62\u003c/span\u003e], Dense Convolutional Network (DenseNet) [\u003cspan class=\"CitationRef\"\u003e63\u003c/span\u003e], and SqueezeNet [\u003cspan class=\"CitationRef\"\u003e64\u003c/span\u003e]. In this study, the input image size for the CNN models was configured as 224 \u0026times; 224 pixels (Fig. \u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e). We employed a ResNet model with 18 layers. We utilized a DenseNet-121 model consisting of four dense blocks with 6, 12, 24, and 16 channels, respectively, and employed SqueezeNet v1.0. These CNN architectures were chosen based on the processed TS images for classification tasks.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\n \u003ch2\u003eResNet\u003c/h2\u003e\n \u003cp\u003eResNet is a well-known convolution-based deep neural network that classifies 2D images. It introduced the concept of a residual block, which addresses the challenge of training deeper neural networks by creating a shortcut connection between the input and output layers. This shortcut connection helps prevent issues such as vanishing or exploding gradient problems during training, and thus facilitate the training of deep neural networks. In ResNet, this skip connection is implemented by summing the output, represented as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(H( \\bullet )\\)\u003c/span\u003e\u003c/span\u003e, with the input, denoted as \u003cem\u003ex\u003c/em\u003e:\u003c/p\u003e\n \u003cdiv id=\"Equ5\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ5\" name=\"EquationSource\"\u003e$$H\\left(x\\right)=F\\left(x\\right)+x$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eHere, \u003cem\u003eF\u003c/em\u003e represents an activation function with weight parameter \u003cem\u003ew\u003c/em\u003e and bias parameter \u003cem\u003eb\u003c/em\u003e, similar to the output of a standard neural network. Adding \u003cem\u003ex\u003c/em\u003e in this equation effectively performs identity mapping, transferring the input directly to the output. Achieving this identity mapping within the residual block, especially when dealing with cases where the dimensions of the input and output layers differ, requires special consideration in two distinct cases: 1) when the dimensions of input and output layers are equivalent and 2) when they differ. In the former case, element-wise summation suffices for identity mapping. In the latter case, Tank et al. have proposed two solutions: zero padding and the utilization of 1 \u0026times; 1 convolutional layers [\u003cspan class=\"CitationRef\"\u003e62\u003c/span\u003e]. Zero padding aligns the dimensions by adding zero-valued elements to the input tensor. Conversely, 1 \u0026times; 1 convolutional layers enable dimension projection, ensuring compatibility between input and output dimensions. These techniques, along with the concept of residual blocks, contribute to the effectiveness of ResNet in training very deep neural networks for image-classification tasks.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\n \u003ch2\u003eDenseNet\u003c/h2\u003e\n \u003cp\u003eDenseNet, which was introduced as an enhancement over ResNet [\u003cspan class=\"CitationRef\"\u003e63\u003c/span\u003e], was devised to improve performance whilst utilizing fewer network parameters. Its primary innovation revolves around dense connectivity, where all preceding feature maps are utilized for identity mapping. The authors\u0026rsquo; core motivation stemmed from concerns that the summation operation employed in residual blocks might hinder the efficient flow of information through the network. To address this limitation, they introduced dense connectivity, concatenating all preceding feature maps for identity mapping. Here, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({x}_{l}\\)\u003c/span\u003e\u003c/span\u003e represents the output of the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(l\\)\u003c/span\u003e\u003c/span\u003eth layer in the network, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({x}_{0}\\)\u003c/span\u003e\u003c/span\u003e denotes the input. The concatenation operation is expressed as follows:\u003c/p\u003e\n \u003cdiv id=\"Equ6\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equ6\" name=\"EquationSource\"\u003e$${x}_{l}={H}_{l}\\left(\\left[{x}_{0},{x}_{1},\\cdots ,{x}_{l-1}\\right]\\right)$$\u003c/div\u003e\n \u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003eHere, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({H}_{l}\\)\u003c/span\u003e\u003c/span\u003e is a composite function involving batch normalization (BN), Rectified Linear Unit (ReLU) activation, and a 3 \u0026times; 3 convolution. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({H}_{l}\\)\u003c/span\u003e\u003c/span\u003e takes the feature maps of all preceding layers, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({x}_{0},{x}_{1},\\cdots ,{x}_{l-1}\\)\u003c/span\u003e\u003c/span\u003e, as input, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\left[ \\bullet \\right]\\)\u003c/span\u003e\u003c/span\u003e denotes concatenation. Besides dense connectivity, the authors introduced two operations to reduce the number of feature maps: bottleneck layers and compression. Bottleneck layers consist of BN-ReLU-1 \u0026times; 1 convolution, followed by BN-ReLU-3 \u0026times; 3 convolution. The inclusion of a 1 \u0026times; 1 convolution in the bottleneck layer aims to reduce the parameter count of the input-feature maps. Compression is employed to reduce feature map dimensions using a compression factor \u003cem\u003e\u0026theta;\u003c/em\u003e within a transition layer, which acts as a down-sampling layer. The transition layer comprises a 1 \u0026times; 1 convolution layer followed by a 2 \u0026times; 2 average pooling layer, and it is positioned between the dense blocks.\u003c/p\u003e\n \u003cp\u003eDenseNet mitigates vanishing gradient problems by leveraging dense connectivity, ensuring that information from earlier layers is preserved and reused throughout the network. This advantageous property enables DenseNet to surpass the performance of ResNet whilst maintaining a more efficient parameter utilization, and thereby results in a reduction of computational resource requirements.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\n \u003ch2\u003eSqueezeNet\u003c/h2\u003e\n \u003cp\u003eSqueezeNet aims to create a more compact neural network with a reduced number of parameters whilst maintaining high performance, which makes it well-suited for resource-constrained computational environments [\u003cspan class=\"CitationRef\"\u003e64\u003c/span\u003e]. The authors noted that AlexNet had a parameter count of 240 MB, whereas SqueezeNet achieved a remarkable reduction to merely 4.8 MB. The authors introduced three key strategies to achieve this substantial reduction in model size. First, instead of using 3 \u0026times; 3 convolutions, SqueezeNet relies on 1 \u0026times; 1 convolutions in certain layers. This change significantly reduces the number of parameters in the network whilst preserving its capacity to learn important features. Second, another strategy involves reducing the number of input channels. The model\u0026apos;s overall parameter count is effectively reduced by decreasing the number of input channels. Third, later in the network architecture, SqueezeNet adopts a down-sampling approach that is strategically placed to ensure that the network generates large activation maps. This design choice optimizes the model\u0026apos;s performance whilst keeping the parameter count low.\u003c/p\u003e\n \u003cp\u003eCentral to the architecture of SqueezeNet is the fire module, which consists of a squeeze layer (comprising only 1 \u0026times; 1 convolutions) followed by an expand layer (comprising a combination of 1 \u0026times; 1 and 3 \u0026times; 3 convolutions). Notably, the authors designed the squeeze layer to have fewer hyperparameters than the total number in the expanding layer, which aligns with the abovementioned second strategy.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec20\" class=\"Section2\"\u003e\n \u003ch2\u003eEvaluation index\u003c/h2\u003e\n \u003cp\u003eThe performance of the two-group classification model was evaluated based on its accuracy, which represents the ratio of the correctly predicted number of samples to the total number of samples. Particularly, accuracy is a feasible evaluation metric for classification problems, which are well-balanced and not skewed or feature no class imbalance.\u003c/p\u003e\n \u003cp\u003eTo evaluate the performance of classifiers on datasets, we use the confusion matrix for the binary classification problem under analysis. For a confusion matrix, four measures, namely \u0026ldquo;true positive\u0026rdquo; (TP), \u0026ldquo;true negative\u0026rdquo; (TN), \u0026ldquo;false positive\u0026rdquo; (FP), and \u0026ldquo;false negative\u0026rdquo; (FN), have been reported. TP is an outcome wherein the model correctly predicts the positive class, which are the real early PDs labeled data in the test dataset and are classified as early PDs by training. TN is an outcome wherein the model correctly predicts the negative class, which are the real controls-labeled data in the test dataset and are classified as controls through training. FP is an outcome wherein the model incorrectly predicts the positive class, which are the real early PDs-labeled data in the test dataset and are classified as controls via training. FN is an outcome wherein the model incorrectly predicts the negative class, which are the real controls-labeled data in the test dataset and are classified as early PDs through training.\u003c/p\u003e\n \u003cp\u003eThe precision represents the proportion of identifications that are correct in practice. The recall is a measure of how many of the positive cases the classifier correctly predicted, over all the positive cases in the data. It is sometimes also referred to as the sensitivity. The F1-Score is a measure encompassing precision and recall. This combined measure is generally described as the harmonic mean of the two-constituent metrics. In essence, the harmonic mean is an alternative approach for calculating the average of values, which is more suitable for ratios (e.g., precision and recall) than the traditional arithmetic mean [\u003cspan class=\"CitationRef\"\u003e65\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e66\u003c/span\u003e].\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\n \u003ch2\u003eStatistical analysis\u003c/h2\u003e\n \u003cp\u003eThe Shapiro\u0026ndash;Wilk test was used for testing multivariate normality. Fisher\u0026rsquo;s exact test, the Mann\u0026ndash;Whitney \u003cem\u003eU\u003c/em\u003e test (for non-normal data), or the independent \u003cem\u003et\u003c/em\u003e-test (for normal data) was conducted to assess the differences in physical and clinical characteristics between early PDs and controls. Statistical analyses were performed using SPSS 21.0 (IBM, Armonk, NY), MATLAB, and Python (Python 3.10, Python Software Foundation). The statistical significance level was set at 0.05.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eDATA AVAILABILITY\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets supporting this study’s findings are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCODE AVAILABILITY\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe do not have an open-source code available. The code 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 code used for the analysis may be requested from the corresponding author.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eACKNOWLEDGEMENTS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by a Dong-A University Foundation Grant. The authors thank all the participants of this study. The authors also thank Editage (www.editage.co.kr) for English language editing. This work was supported by a grant [no. 2022R1A2C100933711; Changhong Youm] from the National Research Foundation of Korea (NRF), funded by the Korean government (MSIT). This research was also supported [grant no. 2022R1A6A3A0108756411; Hwayoung Park] by the Basic Science Research Program through the NRF, funded by the Ministry of Education. The funders had no role in the study design, collection, analysis, and interpretation of the data and in writing the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAUTHOR CONTRIBUTIONS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHC, CY, HP, BK, SC, and SS conceived and designed the study. HC, HP, BK, and SC. recruited the participants. HC, CY, HP, BK, SC, and SS performed the data acquisition. HC, CY, HP, BK, SC, and SS analyzed and interpreted the data. HC, CY, HP, BK, SC, and SS drafted the article. All authors read and approved the final version of the manuscript submitted.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eETHICS DECLARATIONS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll 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 (IRB number: DAUHIRB-22-089) (see ethics approval letter in the supplementary file). All patients provided written informed consent before data collection. The study is registered in the Clinical Research Information Service in the Republic of Korea (KCT0009353).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCOMPETING INTERESTS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no conflict of interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCONSENT FOR PUBLICATION\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eADDITIONAL INFORMATION\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSupplementary information\u003c/strong\u003e\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAversano, L., Bernardi, M. L., Cimitile, M., \u0026amp; Pecori, R. Early detection of Parkinson disease using deep neural networks on gait dynamics. In 2020 International Joint Conference on Neural Networks (IJCNN) (pp. 1-8). \u003cem\u003eIEEE.\u003c/em\u003e (2020).\u003c/li\u003e\n\u003cli\u003eRehman, R. Z. U. \u003cem\u003eet al\u003c/em\u003e. Accelerometry-based digital gait characteristics for classification of Parkinson\u0026apos;s disease: what counts?. \u003cem\u003eIEEE Open J. Eng. Med. Bio.\u003c/em\u003e\u003cstrong\u003e1\u003c/strong\u003e, 65-73 (2020).\u003c/li\u003e\n\u003cli\u003eJankovic, J. Parkinson\u0026apos;s disease: clinical features and diagnosis. \u003cem\u003eJ. Neurol. Neurosurg. Psychiatry\u003c/em\u003e\u003cstrong\u003e79\u003c/strong\u003e, 368-376 (2008).\u003c/li\u003e\n\u003cli\u003eDe Lau, L. M. \u0026amp; Breteler, M. M. Epidemiology of Parkinson\u0026apos;s disease. \u003cem\u003eLancet Neurol\u003c/em\u003e. \u003cstrong\u003e5\u003c/strong\u003e, 525-535 (2006).\u003c/li\u003e\n\u003cli\u003eGao, C. \u003cem\u003eet al\u003c/em\u003e. Model-based and model-free machine learning techniques for diagnostic prediction and classification of clinical outcomes in Parkinson\u0026rsquo;s disease. \u003cem\u003eSci. Rep\u003c/em\u003e. \u003cstrong\u003e8\u003c/strong\u003e, 7129 (2018).\u003c/li\u003e\n\u003cli\u003eMirelman, A. \u003cem\u003eet al\u003c/em\u003e. Detecting sensitive mobility features for Parkinson\u0026apos;s disease stages via machine learning. \u003cem\u003eMov. Disord\u003c/em\u003e. \u003cstrong\u003e36\u003c/strong\u003e, 2144-2155 (2021).\u003c/li\u003e\n\u003cli\u003eRizzo, G. \u003cem\u003eet al\u003c/em\u003e. Accuracy of clinical diagnosis of Parkinson disease: a systematic review and meta-analysis. \u003cem\u003eNeurology\u003c/em\u003e\u003cstrong\u003e86\u003c/strong\u003e, 566-576 (2016).\u003c/li\u003e\n\u003cli\u003eBrognara, L., Palumbo, P., Grimm, B. \u0026amp; Palmerini, L. Assessing gait in Parkinson\u0026rsquo;s disease using wearable motion sensors: a systematic review. \u003cem\u003eDiseases\u003c/em\u003e\u003cstrong\u003e7\u003c/strong\u003e, 18 (2019).\u003c/li\u003e\n\u003cli\u003eDel Din, S. \u003cem\u003eet al\u003c/em\u003e. Gait analysis with wearables predicts conversion to parkinson disease. \u003cem\u003eAnn. Neurol.\u003c/em\u003e\u003cstrong\u003e86\u003c/strong\u003e, 357-367 (2019).\u003c/li\u003e\n\u003cli\u003eBridenbaugh, S. A. \u0026amp; Kressig, R. W. Laboratory review: the role of gait analysis in seniors\u0026rsquo; mobility and fall prevention. \u003cem\u003eGerontology\u003c/em\u003e\u003cstrong\u003e57\u003c/strong\u003e, 256-264 (2011).\u003c/li\u003e\n\u003cli\u003eBuckley, C. \u003cem\u003eet al\u003c/em\u003e. The role of movement analysis in diagnosing and monitoring neurodegenerative conditions: insights from gait and postural control. \u003cem\u003eBrain Sci.\u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, 34 (2019).\u003c/li\u003e\n\u003cli\u003eDaneault, J. F. \u003cem\u003eet al\u003c/em\u003e. Accelerometer data collected with a minimum set of wearable sensors from subjects with Parkinson\u0026rsquo;s disease. \u003cem\u003eSci. Data\u003c/em\u003e\u003cstrong\u003e8\u003c/strong\u003e, 48 (2021).\u003c/li\u003e\n\u003cli\u003eEvers, L. J. \u003cem\u003eet al\u003c/em\u003e. Real-life gait performance as a digital biomarker for motor fluctuations: the Parkinson@ Home validation study. \u003cem\u003eJ. Med. Internet Res\u003c/em\u003e. \u003cstrong\u003e22\u003c/strong\u003e, e19068 (2020).\u003c/li\u003e\n\u003cli\u003eGadaleta, M. \u003cem\u003eet al\u003c/em\u003e. Deep learning techniques for improving digital gait segmentation. \u003cem\u003eAnnu. Int. Conf. IEEE. Eng. Med. Biol. Soc.\u003c/em\u003e\u003cstrong\u003e2019\u003c/strong\u003e, 1834-1837 (2019).\u003c/li\u003e\n\u003cli\u003eJuneau, P. \u003cem\u003eet al\u003c/em\u003e. Amputee fall risk classification using machine learning and smartphone sensor data from 2-minute and 6-minute walk tests. \u003cem\u003eSensors\u003c/em\u003e\u003cstrong\u003e22\u003c/strong\u003e, 1749 (2022).\u003c/li\u003e\n\u003cli\u003eBalke, B. A simple field test for the assessment of physical fitness. \u003cem\u003eRep Civ Aeromed Res Inst US\u003c/em\u003e. \u003cstrong\u003e63\u003c/strong\u003e, (1963).\u003c/li\u003e\n\u003cli\u003eFan, S. \u003cem\u003eet al\u003c/em\u003e. Digital health technology combining wearable gait sensors and machine learning improve the accuracy in prediction of frailty. \u003cem\u003eFront. Public Health\u003c/em\u003e\u003cstrong\u003e11\u003c/strong\u003e, 1169083 (2023).\u003c/li\u003e\n\u003cli\u003eMudge, S. \u0026amp; Stott, N. S. Timed walking tests correlate with daily step activity in persons with stroke. \u003cem\u003eArch. Phys. Med. Rehabil\u003c/em\u003e. \u003cstrong\u003e90\u003c/strong\u003e, 296-301 (2009).\u003c/li\u003e\n\u003cli\u003eDaines, K. \u003cem\u003eet al\u003c/em\u003e. Fall risk classification for people with lower extremity amputations using random forests and smartphone sensor features from a 6-minute walk test. \u003cem\u003ePLoS One\u003c/em\u003e\u003cstrong\u003e16\u003c/strong\u003e, e0247574 (2021).\u003c/li\u003e\n\u003cli\u003eIreland, D. \u003cem\u003eet al\u003c/em\u003e. Classification of movement of people with parkinsons disease using wearable inertial movement units and machine learning. \u003cem\u003eStud. Health Technol. Infor.\u003c/em\u003e\u003cstrong\u003e227\u003c/strong\u003e, 61-6 (2016).\u003c/li\u003e\n\u003cli\u003eFuentes-Abolafio, I. J. \u003cem\u003eet al\u003c/em\u003e. Kinematic parameters related to functional capacity, fatigue, and breathlessness during the 6-min walk test in older adults with heart failure with preserved ejection fraction. \u003cem\u003eEur. J. Cardiovasc. Nurs.\u003c/em\u003e\u003cstrong\u003e23\u003c/strong\u003e, 69-80 (2023).\u003c/li\u003e\n\u003cli\u003eM\u0026auml;ntt\u0026auml;ri, A. \u003cem\u003eet al\u003c/em\u003e. Six‐minute walk test: a tool for predicting maximal aerobic power (VO 2 max) in healthy adults. \u003cem\u003eClin. Physiol. Funct. Imaging\u003c/em\u003e\u003cstrong\u003e38\u003c/strong\u003e, 1038-1045 (2018).\u003c/li\u003e\n\u003cli\u003eSperandio, E. F. \u003cem\u003eet al\u003c/em\u003e. Reference values for the 6-min walk test in healthy middle-aged and older adults: from the total distance traveled to physiological responses. \u003cem\u003eFisioterapia em Movimento\u003c/em\u003e, \u003cstrong\u003e32\u003c/strong\u003e, (2019).\u003c/li\u003e\n\u003cli\u003eStorm, F. A. \u003cem\u003eet al\u003c/em\u003e. Wearable inertial sensors to assess gait during the 6-minute walk test: A systematic review. \u003cem\u003eSensors (Basel).\u003c/em\u003e\u003cstrong\u003e20\u003c/strong\u003e, 2660 (2020).\u003c/li\u003e\n\u003cli\u003eAtri, R. \u003cem\u003eet al\u003c/em\u003e. Deep learning for daily monitoring of Parkinson\u0026rsquo;s disease outside the clinic using wearable sensors. \u003cem\u003eSensors (Basel).\u003c/em\u003e\u003cstrong\u003e22\u003c/strong\u003e, 6831 (2022).\u003c/li\u003e\n\u003cli\u003eAngelini, L. \u003cem\u003eet al\u003c/em\u003e. Wearable sensors can reliably quantify gait alterations associated with disability in people with progressive multiple sclerosis in a clinical setting. \u003cem\u003eJ. Neurol\u003c/em\u003e. \u003cstrong\u003e267\u003c/strong\u003e, 2897-2909 (2020).\u003c/li\u003e\n\u003cli\u003eMei, J., Desrosiers, C. \u0026amp; Frasnelli, J. Machine learning for the diagnosis of Parkinson\u0026apos;s disease: a review of literature. \u003cem\u003eFront. Aging Neurosci\u003c/em\u003e. \u003cstrong\u003e13\u003c/strong\u003e, 633752 (2021).\u003c/li\u003e\n\u003cli\u003eAlaskar, H. \u003cem\u003eet al\u003c/em\u003e. A data science approach for reliable classification of neuro-degenerative diseases using gait patterns. \u003cem\u003eJ. Reliab. Intell. Environ\u003c/em\u003e. \u003cstrong\u003e6\u003c/strong\u003e, 233-247 (2020).\u003c/li\u003e\n\u003cli\u003eBalaji, E., Brindha, D., Elumalai, V. K. \u0026amp; Umesh, K. Data-driven gait analysis for diagnosis and severity rating of Parkinson\u0026rsquo;s disease. \u003cem\u003eMed. Eng. Phys\u003c/em\u003e. \u003cstrong\u003e91\u003c/strong\u003e, 54-64 (2021).\u003c/li\u003e\n\u003cli\u003eBorz\u0026igrave;, L., Sigcha, L., Rodr\u0026iacute;guez-Mart\u0026iacute;n, D. \u0026amp; Olmo, G. Real-time detection of freezing of gait in Parkinson\u0026rsquo;s disease using multi-head convolutional neural networks and a single inertial sensor. \u003cem\u003eArtif. Intell. Med.\u003c/em\u003e\u003cstrong\u003e135\u003c/strong\u003e, 102459 (2023).\u003c/li\u003e\n\u003cli\u003eJuneau, P. Automated foot strike identification and fall risk classification for people with lower limb amputations using smartphone sensor signals from 2 and 6-minute walk tests (Doctoral dissertation, Universit\u0026eacute; d\u0026apos;Ottawa/University of Ottawa). (2022).\u003c/li\u003e\n\u003cli\u003eEI Maachi, I., Bilodeau, G. A. \u0026amp; Bouachir, W. Deep 1D-Convnet for accurate Parkinson disease detection and severity prediction from gait. \u003cem\u003eExpert Syst. Appl.\u003c/em\u003e\u003cstrong\u003e143\u003c/strong\u003e, 113075 (2020).\u003c/li\u003e\n\u003cli\u003eYang, X., Ye, Q., Cai, G., Wang, Y. \u0026amp; Cai, G. PD-ResNet for classification of Parkinson\u0026rsquo;s disease from gait. \u003cem\u003eIEE.\u003c/em\u003e\u003cem\u003eJ. Transl. Eng. Health Med.\u003c/em\u003e\u003cstrong\u003e10,\u003c/strong\u003e 2200111 (2022).\u003c/li\u003e\n\u003cli\u003eRamanujam, E., Perumal, T., \u0026amp; Padmavathi, S. Human activity recognition with smartphone and wearable sensors using deep learning techniques: a review. \u003cem\u003eIEEE Sens. J.\u003c/em\u003e\u003cstrong\u003e21\u003c/strong\u003e, 13029-13040 (2021).\u003c/li\u003e\n\u003cli\u003eNguyen, M. D.\u003cem\u003e et al\u003c/em\u003e. IMU-based spectrogram approach with deep convolutional neural networks for gait classification. In 2020 IEEE International Conference on Consumer Electronics (ICCE) (pp. 1-6). IEEE. (2020).\u003c/li\u003e\n\u003cli\u003eG\u0026uuml;ney, P. A. et al. Comparison of 1-, 3-, and 6-minute walk tests in the post-lung transplant period. \u003cem\u003eSouth Clin. Ist. Euras.\u003c/em\u003e\u003cstrong\u003e33\u003c/strong\u003e, (2022).\u003c/li\u003e\n\u003cli\u003eHadouiri, N., Monnet, E., Gouelle, A., Sagawa Jr, Y. \u0026amp; Decavel, P. Locomotor strategy to perform 6-minute walk test in people with multiple sclerosis: a prospective observational study. \u003cem\u003eSensors (Basel).\u003c/em\u003e\u003cstrong\u003e23\u003c/strong\u003e, 3407 (2023).\u003c/li\u003e\n\u003cli\u003eBailo, G. \u003cem\u003eet al\u003c/em\u003e. Characterization of walking in mild Parkinson\u0026rsquo;s disease: reliability, validity and discriminant ability of the six-minute walk test instrumented with a single inertial sensor. \u003cem\u003eSensors (Basel).\u003c/em\u003e\u003cstrong\u003e24\u003c/strong\u003e, 662 (2024).\u003c/li\u003e\n\u003cli\u003eChou, P. Y. \u0026amp; Lee, S. C. Turning deficits in people with Parkinson\u0026apos;s disease. \u003cem\u003eTzu Chi Med. J.\u003c/em\u003e\u003cstrong\u003e25\u003c/strong\u003e, 200-202. (2013).\u003c/li\u003e\n\u003cli\u003eAkram, S., Frank, J. S., \u0026amp; Jog, M. Parkinson\u0026apos;s disease and segmental coordination during turning: I. Standing turns. \u003cem\u003eCan. J. Neurol. Sci.\u003c/em\u003e\u003cstrong\u003e40\u003c/strong\u003e, 512-519 (2013).\u003c/li\u003e\n\u003cli\u003eRamesh, V., \u0026amp; Bilal, E. Detecting motor symptom fluctuations in Parkinson\u0026rsquo;s disease with generative adversarial networks. \u003cem\u003eNPJ Digit. Med.\u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, 138 (2022).\u003c/li\u003e\n\u003cli\u003eChoi, H. \u003cem\u003eet al\u003c/em\u003e. Association between severity of freezing of gait and turning characteristics in people with Parkinson\u0026rsquo;s disease. \u003cem\u003eInt. J. Environ. Res. Public Health\u003c/em\u003e\u003cstrong\u003e19\u003c/strong\u003e, 12131 (2022).\u003c/li\u003e\n\u003cli\u003eWu, Z. \u003cem\u003eet al\u003c/em\u003e. Mild gait impairment and its potential diagnostic value in patients with early-stage Parkinson\u0026rsquo;s disease. \u003cem\u003eBehav. Neurol.\u003c/em\u003e\u003cstrong\u003e2021\u003c/strong\u003e, 6696454 (2021).\u003c/li\u003e\n\u003cli\u003eLin, C. H. \u003cem\u003eet al\u003c/em\u003e. Early detection of Parkinson\u0026rsquo;s disease by neural network models. \u003cem\u003eIEEE Access\u003c/em\u003e\u003cstrong\u003e10\u003c/strong\u003e, 19033-19044 (2022).\u003c/li\u003e\n\u003cli\u003eMeng, L. \u003cem\u003eet al\u003c/em\u003e. Inertial-based gait metrics during turning improve the detection of early-stage parkinson\u0026rsquo;s disease patients. \u003cem\u003eIEEE Trans. Neural. Syst. Rehabil. Eng.\u003c/em\u003e\u003cstrong\u003e31\u003c/strong\u003e, 1472-1482 (2023).\u003c/li\u003e\n\u003cli\u003eShcherbak, A., Kovalenko, E., \u0026amp; Somov, A. Detection and classification of early stages of Parkinson\u0026rsquo;s disease through wearable sensors and machine learning. \u003cem\u003eIEEE Trans. Instrum. Meas.\u003c/em\u003e\u003cstrong\u003e72\u003c/strong\u003e, 1-9 (2023).\u003c/li\u003e\n\u003cli\u003eCai, G. \u003cem\u003eet al\u003c/em\u003e. Specific distribution of digital gait biomarkers in Parkinson\u0026rsquo;s disease using body-worn sensors and machine learning. \u003cem\u003eGerontol. A. Biol. Sci. Med. Sci\u003c/em\u003e. \u003cstrong\u003e78,\u003c/strong\u003e 1348-1354 (2023).\u003c/li\u003e\n\u003cli\u003eSveinbjornsdottir, S. The clinical symptoms of Parkinson\u0026apos;s disease. \u003cem\u003eJ. Neurochem.\u003c/em\u003e\u003cstrong\u003e139\u003c/strong\u003e, 318-324 (2016).\u003c/li\u003e\n\u003cli\u003eBuckley, C. \u003cem\u003eet al\u003c/em\u003e. Upper body accelerations as a biomarker of gait impairment in the early stages of Parkinson\u0026rsquo;s disease. \u003cem\u003eGait Posture\u003c/em\u003e\u003cstrong\u003e71\u003c/strong\u003e, 289-295 (2019).\u003c/li\u003e\n\u003cli\u003eChandrabhatla, A. S., Pomeraniec, I. J. \u0026amp; Ksendzovsky, A. Co-evolution of machine learning and digital technologies to improve monitoring of Parkinson\u0026rsquo;s disease motor symptoms. \u003cem\u003eNPJ Digit. Med.\u003c/em\u003e\u003cstrong\u003e5\u003c/strong\u003e, 32 (2022).\u003c/li\u003e\n\u003cli\u003ePalmerini, L., Mellone, S., Avanzolini, G., Valzania, F. \u0026amp; Chiari, L. Quantification of motor impairment in Parkinson\u0026apos;s disease using an instrumented timed up and go test. \u003cem\u003eIEEE Trans. Neural. Syst. Rehabil. Eng.\u003c/em\u003e\u003cstrong\u003e21\u003c/strong\u003e, 664-673 (2013).\u003c/li\u003e\n\u003cli\u003eBohannon, R. W. \u003cem\u003eet al\u003c/em\u003e. Comparison of walking performance over the first 2 minutes and the full 6 minutes of the six-minute walk test. \u003cem\u003eBMC Res. Notes\u003c/em\u003e. \u003cstrong\u003e7\u003c/strong\u003e, 269 (2014).\u003c/li\u003e\n\u003cli\u003eZhang, H., Deng, K., Li, H., Albin, R. L. \u0026amp; Guan, Y. Deep learning identifies digital biomarkers for self-reported Parkinson\u0026apos;s disease. \u003cem\u003ePatterns. (N. Y.)\u003c/em\u003e\u003cstrong\u003e1\u003c/strong\u003e:100042 (2020).\u003c/li\u003e\n\u003cli\u003eVanmechelen, I. \u003cem\u003eet al\u003c/em\u003e. Assessment of movement disorders using wearable sensors during upper limb tasks: a scoping review. \u003cem\u003eFront. Robot. AI\u003c/em\u003e\u003cstrong\u003e9\u003c/strong\u003e, 1068413 (2023).\u003c/li\u003e\n\u003cli\u003eHughes, A. J., Daniel, S. E. \u0026amp; Lees, A. J. Improved accuracy of clinical diagnosis of Lewy body Parkinson\u0026rsquo;s disease. \u003cem\u003eNeurology\u003c/em\u003e\u003cstrong\u003e57\u003c/strong\u003e, 1497-1499 (2001).\u003c/li\u003e\n\u003cli\u003eHoehn, M.M. \u0026amp; Yahr, M.D. Parkinsonism: onset, progression, and mortality. \u003cem\u003eNeurology\u003c/em\u003e\u003cstrong\u003e17\u003c/strong\u003e, 427-442 (1967).\u003c/li\u003e\n\u003cli\u003eGoetz, C. G. \u003cem\u003eet al\u003c/em\u003e. Movement Disorder Society Task Force report on the Hoehn and Yahr staging scale: status and recommendations the Movement Disorder Society Task Force on rating scales for Parkinson\u0026apos;s disease. \u003cem\u003eMov. Disord\u003c/em\u003e. \u003cstrong\u003e19\u003c/strong\u003e, 1020-1028 (2004).\u003c/li\u003e\n\u003cli\u003eFolstein, M. F., Folstein, S. E. \u0026amp; McHugh, P. R. \u0026ldquo;Mini-mental state\u0026rdquo;: a practical method for grading the cognitive state of patients for the clinician. \u003cem\u003eJ. Psychiatr. Res\u003c/em\u003e. \u003cstrong\u003e12\u003c/strong\u003e, 189-198 (1975).\u003c/li\u003e\n\u003cli\u003eNasreddine, Z. S. \u003cem\u003eet al\u003c/em\u003e. The Montreal Cognitive Assessment, MoCA: a brief screening tool for mild cognitive impairment. \u003cem\u003eJ. Am. Geriatr. Soc\u003c/em\u003e. \u003cstrong\u003e53\u003c/strong\u003e, 695-699 (2005).\u003c/li\u003e\n\u003cli\u003eWang, Z., Qin, M., \u0026amp; Chen, Y. K. Learning from the cnn-based compressed domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (pp. 3582-3590). (2022).\u003c/li\u003e\n\u003cli\u003eEckmann, J. P., Kamphorst, S. O. \u0026amp; Ruelle, D. Recurrence plots of dynamical systems. \u003cem\u003eEPL\u003c/em\u003e, \u003cstrong\u003e4\u003c/strong\u003e, 973-977 (1987).\u003c/li\u003e\n\u003cli\u003eTank, V. H. \u003cem\u003eet al\u003c/em\u003e. Drug eluting stents versus bare metal stents for the treatment of extracranial vertebral artery disease: a meta-analysis. \u003cem\u003eJ. Neurointerv. Surg\u003c/em\u003e. \u003cstrong\u003e8\u003c/strong\u003e, 770-774 (2016).\u003c/li\u003e\n\u003cli\u003eHuang, G. \u003cem\u003eet al\u003c/em\u003e. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4700-4708). (2017).\u003c/li\u003e\n\u003cli\u003eIandola, F. N. \u003cem\u003eet al\u003c/em\u003e. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and\u0026lt; 0.5 MB model size. arXiv preprint arXiv:1602.07360 (2016).\u003c/li\u003e\n\u003cli\u003eUchitomi, H., Ming, X., Zhao, C., Ogata, T. \u0026amp; Miyake, Y. Classification of mild Parkinson\u0026rsquo;s disease: data augmentation of time-series gait data obtained via inertial measurement units. \u003cem\u003eSci. Rep\u003c/em\u003e\u003cem\u003e.\u003c/em\u003e\u003cstrong\u003e13\u003c/strong\u003e, 12638 (2023).\u003c/li\u003e\n\u003cli\u003eBernardo, L. S., Dama\u0026scaron;evičius, R., Ling, S. H., de Albuquerque, V. H. C. \u0026amp; Tavares, J. M. R. Modified squeezenet architecture for parkinson\u0026rsquo;s disease detection based on keypress data. \u003cem\u003eBiomedicines\u003c/em\u003e\u003cstrong\u003e10\u003c/strong\u003e, 2746 (2022).\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Table","content":"\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"567\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"100%\" colspan=\"4\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 1.\u003c/strong\u003e Physical and clinical characteristics of all participants.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003eEarly PDs (n = 78)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003eControls (n = 50)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e\u003cem\u003ep\u003c/em\u003e-value\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eSex (male/female)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e37 / 41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e21 / 29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e0.588\u003csup\u003ea\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eAge (years)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e67.51 \u0026plusmn; 7.08\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e65.80 \u0026plusmn; 5.51\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e0.149\u003csup\u003eb\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eHeight (cm)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e161.20 \u0026plusmn; 7.72\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e161.14 \u0026plusmn; 8.38\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e0.907\u003csup\u003eb\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eBody mass (kg)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e63.96 \u0026plusmn; 10.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e63.68 \u0026plusmn; 10.04\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e0.884\u003csup\u003eb\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eBMI (kg/m\u003csup\u003e2\u003c/sup\u003e)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e24.51 \u0026plusmn; 3.13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e24.40 \u0026plusmn; 2.43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e0.847\u003csup\u003eb\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eDisease duration (years)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e5.50 \u0026plusmn; 3.96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eTreatment duration (years)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e4.19 \u0026plusmn; 3.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eL-Dopa equivalent dose (mg/day)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e541.90 \u0026plusmn; 286.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eMMSE (scores)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e28.12 \u0026plusmn; 1.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e27.22 \u0026plusmn; 1.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e0.003\u003csup\u003ec\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eMoCA (scores)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e26.23 \u0026plusmn; 2.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e24.86 \u0026plusmn; 2.99\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e0.001\u003csup\u003ec\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eUPDRS Total (scores)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e47.68 \u0026plusmn; 19.54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eUPDRS III (scores)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e25.35 \u0026plusmn; 13.63\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003eH\u0026amp;Y Scale (stages 1 and 2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e28 / 50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"38.33922261484099%\"\u003e\n \u003cp\u003e6MWT (m)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e414.24 \u0026plusmn; 85.16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"23.32155477031802%\"\u003e\n \u003cp\u003e494.23 \u0026plusmn; 51.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"15.017667844522968%\"\u003e\n \u003cp\u003e\u0026lt;0.001\u003csup\u003ec\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"100%\" colspan=\"4\"\u003e\n \u003cp\u003eThe data presented are the mean \u0026plusmn; standard deviation. Significant difference: \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05; PD: Parkinson\u0026rsquo;s disease; BMI: Body mass index; L-Dopa: Levodopa; MMSE: Mini-Mental State Examination; MoCA: Montreal Cognitive Assessment; UPDRS: Unified Parkinson\u0026rsquo;s Disease Rating Scale; H\u0026amp;Y: Hoehn and Yahr; 6MWT: 6-min walk test.\u003c/p\u003e\n \u003cp\u003e\u003csup\u003ea\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026shy;value of Fisher\u0026rsquo;s exact test.\u003c/p\u003e\n \u003cp\u003e\u003csup\u003eb\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026shy;value of the independent t\u0026shy;test.\u003c/p\u003e\n \u003cp\u003e\u003csup\u003ec\u003c/sup\u003e \u003cem\u003ep\u003c/em\u003e\u0026shy;value of the Mann\u0026ndash;Whitney U test.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Parkinson’s disease, detection, artificial intelligence, deep learning, convolutional neural network, six-minute walk test","lastPublishedDoi":"10.21203/rs.3.rs-4482534/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4482534/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe heterogeneity of Parkinson\u0026rsquo;s disease (PD) generates significant challenges for accurate diagnosis, especially in early-stage disease, when symptoms may be very subtle. This study aimed to determine the accuracy of a convolutional neural network (CNN) technique based on a 6-min walk test (6MWT) using wearable sensors for distinguishing patients with early-stage PD (n\u0026thinsp;=\u0026thinsp;78) from healthy controls (n\u0026thinsp;=\u0026thinsp;50). Wearing six sensors, the participants performed the 6MWT, and the time-series data were converted into new images. The main results showed that the gyroscopic vertical component of the lumbar spine had the highest classification accuracy of 83.5%, followed by the thoracic spine (83.1%) and right thigh (79.5%) segment. These results suggest that the 6MWT and CNN models may pave the way for clinicians to diagnose and track PD symptoms earlier and thus provide timely treatment during the golden transition from geriatric to pathologic gait patterns.\u003c/p\u003e","manuscriptTitle":"Convolutional neural network-based detection of early-stage Parkinson’s disease using the six-minute walk test","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-06-12 19:13:34","doi":"10.21203/rs.3.rs-4482534/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-07-17T12:59:22+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-10T01:26:38+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"105286922833182132888450388560378223765","date":"2024-07-08T07:04:56+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-06-15T18:28:16+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"295764903418563087441187231789444799386","date":"2024-06-07T02:49:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"192217403655562616844604411333176351112","date":"2024-06-06T07:59:50+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-06-03T23:25:26+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-06-03T23:14:50+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2024-05-28T09:37:17+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-05-28T08:08:21+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2024-05-27T05:48:29+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"38fa7293-76e0-424b-8ec8-49b4201e4631","owner":[],"postedDate":"June 12th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":32996207,"name":"Health sciences/Biomarkers"},{"id":32996208,"name":"Health sciences/Neurology"}],"tags":[],"updatedAt":"2024-10-07T16:00:29+00:00","versionOfRecord":{"articleIdentity":"rs-4482534","link":"https://doi.org/10.1038/s41598-024-72648-w","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2024-09-30 15:57:16","publishedOnDateReadable":"September 30th, 2024"},"versionCreatedAt":"2024-06-12 19:13:34","video":"","vorDoi":"10.1038/s41598-024-72648-w","vorDoiUrl":"https://doi.org/10.1038/s41598-024-72648-w","workflowStages":[]},"version":"v1","identity":"rs-4482534","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4482534","identity":"rs-4482534","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}