Introduction

Parkinson’s disease (PD) is a progressive neurodegenerative disorder that leads to a decline in fundamental abilities that are vital for activities of daily life1,2,3,4. This reduction in function has been reported to directly affect independence and quality of life5,6,7,8,9. Moreover, PD is one of the most common disorders affecting 3-5% of individuals over the age of 6510,11. While there is no known cure for PD, research has shown that via appropriate rehabilitation and medication therapy, one can alleviate symptom progression12,13,14,15,16,17. Thus, the early detection of PD and the ongoing monitoring of rehabilitation status are crucial for optimizing outcomes.

Current diagnosis and monitoring of PD rely on the comprehensive assessment of medical history and clinical examination18,19. Prodromal features, such as rapid eye movement sleep behavior disorder, hyposmia, constipation, tremor, stiffness, and slowness, serve as key elements in the diagnostic process20,21,22,23,24,25. However, these examinations are performed during prescheduled medical visits. This makes it difficult for neurologists to capture variations in the patient’s disability level and plan appropriate therapeutical adjustments especially in the early stages of PD. Furthermore, misclassification of PD in early diagnosis is not uncommon, with error rates reported to be in the range of 15% to 24%26,27. Therefore, imaging tests such as dopamine transporter single-photon emission computed tomography (DaT-SPECT) and FP-CIT PET/CT are widely utilized and known as the gold standards for PD diagnosis28,29,30,31. However, these methods still face challenges related to accessibility and cost. In recent years, several studies have been proposed for using convolutional neural networks (CNN) models to detect PD. Tarjni et al. proposed a CNN model for detecting PD using brain images obtained through magnetic resonance imaging (MRI)32. Ruilin et al. and Sue et al. utilized electroencephalography (EEG) signals33,34. Helber et al. proposed a combined model of CNN and gait recurrent units (GRU) using inertial measurement unit (IMU) signals during gait35. Imanne et al. utilized ground reaction force during gait36. Clayton et al. introduced a novel smart pen37 and Hankan et al. utilized vocal data38. Nevertheless, the early detection of PD still remains a very elusive area.

What is intriguing is that there have been reports of individuals diagnosed with PD, experiencing significant freezing of gait, that have demonstrated the ability to ride a bicycle39,40. Research suggests the unique phenomenon where PD patients maintain cycling ability despite significant gait impairment is related to several neural mechanisms. Bicycle’s rotating pedals may act as an external pacing cue39,40. Additionally, compared with walking, cycling does not require postural adjustments (lateral weight shift) to unload the swing leg to initiate a step40. Therefore, the motor-control mechanisms involved in cycling may be less affected by PD than in walking. This has inspired us to investigate whether a deep learning model using kinetic and kinematic data obtained during cycling can detect patients with PD and have the potential to serve as an auxiliary tool in monitoring the rehabilitation status. As is well known, cycling also has positive effects on cardiovascular fitness and motor skills41,42,43,44.

This study explores the feasibility of CNN-based PD detection using cycling data obtained from a specially designed bicycle platform which allows for lateral movement via steering in order to simulate outdoor cycling experiences (Fig. 1a). The bicycle used in the study was equipped with 6-axis force-torque sensors on the headset spacer and seat post respectively. The input data for the CNN model proposed in this study consist of time-series data (force, moment, speed, and lateral movement) recorded during cycling and tabular data comprising personal information (sex, age, height, and weight). With the participation of 29 PD patients and 36 healthy controls (HC), our CNN model showed an accuracy of approximately 86% in classification based on the 5-fold cross validation.

Materials and methods

Experimental design

Fig. 1
figure 1

Experimental design. (a) A bicycle and platform used in this study. (b) A real-time lateral position monitoring program. (c) A bicycle equipped with two force-torque sensors (indicated by red dotted circles).

Full size image

In this study, an indoor bicycle platform (Ultiracer, RealDesignTech Co., Ltd., Seongnam, Korea) was utilized to simulate outdoor cycling environments (Fig. 1a). Specifically designed for mounting conventional bicycles, this device consists of two free-spinning cylinders and a slidable pole that holds the body of the bicycle. Here, “slidable” means that the pole can move laterally. Like outdoor bicycles, one can move side to side by turning the handle. For safety, the range of lateral movement was restricted within ±100 mm. In addition, the sliding pole can adjust the lateral tilt within 3 degrees, but for patient safety, we limited this to 1 degree. When the bicycle wheels rotate, the cylinders spin due to friction. This can simulate the inertial effect of real cycling. The diameter of the cylinder is 110 mm, and the platform size is 520 mm in width and 1380 mm in length.

In order to measure the lateral position of the pole, an infrared time of flight (ToF) distance sensor VL53L0X (STMicroelectronics Inc., Geneva, Switzerland), which can measure distances of up to 2 m with 1 mm resolution, was placed on the right-side fixed frame of the sliding pole, facing the pole. An encoder was equipped on the front cylinder to measure its spinning speed. A simple game screen provides real-time feedback on the lateral position and speed of the bicycle (Fig. 1a and b). To measure the interaction between the user and bicycle, two 6-axis force-torque (FT) sensors Dynpick (WACOH-TECH Inc., Tokyo, Japan) were inserted into the headset spacer and the seat post of the bicycle (Fig. 1c). The sample rates of the signal acquired from the infrared ToF sensor and encoder were both 50 Hz. The original sample rate of the FT sensor was 2000 Hz, which was downsampled to 200 Hz in this study.

Participants

Twenty-nine elderly outpatients with PD from the Department of Neurolody at Korea University Medical Center participated in this study (Table 1). All PD patients received confirmation of their PD diagnosis through FP-CIT PET/CT scans. Additionally, in order to quantitatively assess the motor, cognitive, and emotional status of the PD patients, the Hoehn and Yahr (H&Y) stage and the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) were also employed. The H&Y stage is a classification system used for diagnosis and assessment of the progression of PD45. It consists of five stages, each indicating different levels of symptom severity and functional impact. Until Stage 2, there is minimal interference with daily activities. For the control group, 36 healthy elderly individuals were recruited. There were no statistically significant differences in sex ratio, age, height, and weight between the PD patients and the HC group by chi-square tests and Wilcoxon rank sum test (p > 0.05). All sessions were conducted under the supervision of a licensed medical doctor. All participants provided written informed consent before participating in the study, and the research ethics of human experiments were ensured by conducting the study in accordance with the approved contents by the Institutional Review Board of Korea University Medicine Anam Hospital (IRB No. 2022AN0420, and 2023AN0236). All experiments were carried out in accordance with the approved guidelines and with the Declaration of Helsinki.

Prior to any data acquisition, the participants underwent a familiarization period to adapt to cycling on the bicycle platform. Once they felt comfortable and ready to proceed with the experiment, they were asked to maintain a central position on the screen while cycling at self-selected speeds (Fig. 1a). The experiment consisted of three trials, each lasting 30 s, and only the 20-second data excluding the 5 s at the beginning and end of each trial was utilized as the input of the proposed CNN model.

Table 1 Participant information.
Full size table

Signal processing

Normalizing input data is known to enhance model stability, mitigate issues related to diverse feature scales, and prevent overfitting in deep learning. In this study, we primarily used the min-max normalization method, which scales the data to a range between 0 and 1, as shown below:

$$\:{x}_{normalized}=\:\frac{x-\text{m}\text{i}\text{n}\left(x\right)}{\text{max}\left(x\right)-\text{m}\text{i}\text{n}\left(x\right)}\:$$
(1)

The force and torque measurement ranges of the FT sensor we used were − 1000 to 1000 N and − 30 to + 30 Nm, respectively, so their minimum values ​​for normalization were − 1000 N and − 30 Nm, respectively. Spinning cylinder speed were similarly normalized with minimum and maximum values set at 0 rpm and 105 rpm, respectively, where the upper bound was determined based on the maximum instantaneous speed observed in all experiments, which was approximately 104 rpm. The position of the sliding pole measured by the ToF sensor was normalized with a minimum of 6 cm and a maximum of 206 cm. The personal information data were also normalized: age was scaled between 0 and 100 years, height between 0 and 200 cm, and weight between 0 and 100 kg. Gender was binary-encoded as 0 for male and 1 for female.

Figure 2 illustrates the processing steps to prepare input data samples. In order to improve generalization of the CNN model, we performed the stratified 5-fold cross-validation. This technique involves splitting the dataset into five subsets, while maintaining the proportionate representation of PD patient and HC groups (Fig. 2a, b). Then, all the time-series signals were segmented into 2-second data samples with 90% overlap for data augmentation (Fig. 2c, d, and e). Figure 3 shows two examples of data samples from HC and PD patient groups. The signal processing was performed using a custom-built MATLAB (Mathworks Inc., Natick, MA, USA) code.

Fig. 2
figure 2

Stratified 5-fold splitting and signal segmentation process of time-series data. (a) The entire subject data labeled PD or HC. (b) Stratified 5-fold split data set. The red boxes represent the test set, while the others represent the training set. (c) One subject’s time-series data consisting of three trials with each trial lasting 20 s. The trial data was collected three times per individual. (d) The segmentation process involves dividing the 20-second data into 2-second segments with a 1.8-second overlap. (e) The outcome yielded 91 sets of 2-second time series data per trial.

Full size image
Fig. 3
figure 3

Example of time-series data sample. (a) The force, moment, lateral position, and speed data of a subject in the HC group. (Male, 65 years, 170 cm, and 71 kg). (b) The force, moment, lateral position, and speed data of a subject in the PD group (Female, 70 years, 153 cm, and 63 kg). (c) Definition of the force and moment components measured by the two FT sensors. The arrow direction indicates positive direction. The arrow colors correspond to the colors of the force and moment graphs shown in (a) and (b).

Full size image

Convolutional neural network model

Fig. 4
figure 4

CNN model architecture used in this study.

Full size image

The CNN model architecture proposed in this study (Fig. 4) was based on the model proposed by Sushravya et al.46. TensorFlow (version 2.10.0) was utilized to train the CNN model in Python (version 3.9.18). We designed a CNN model using 6 branches with distinct inputs (branch 1: three-axis force signal from headset spacer, branch 2: three-axis moment signal from headset spacer, branch 3: three-axis force signal from seat post, branch 4: three-axis moment signal from seat post, branch 5: lateral position, branch 6: speed). The numbers in the parentheses above the arrows in Fig. 4 indicate the output data size of each layer.

The first four branches and last two branches consist of three and two convolutional blocks, respectively. Each convolutional block included a one-dimensional convolutional layer, a rectified linear unit (ReLU) activation, and a max pooling layer. The final outputs of all convolutional blocks are merged into a concatenate layer with tabular inputs (gender, age, height and weight) and then connected to a series of three dense layers with a ReLU function and an output layer with a sigmoid function. For the binary classification of PD vs. HC, when the output of the sigmoid function was greater than a threshold of 0.5, it was classified to PD. We used the Adam optimizer with the batch size set to 8, and an early stopping technique with up to 50 epochs to prevent overfitting. And unless specified, we used the default hyperparameters of TensorFlow.

Performance evaluation

In order to evaluate the performance of the proposed classification model, we employed two evaluation methods. The first method functions at the subject-level, considering it to be correctly classified if over 50% of the classified data samples, gathered from an individual, match their true label. We tested 13 subjects for this method. The second approach functions at the individual dataset-level, where each 2-second data sample is individually evaluated and considered to be accurate if the data sample matches its true label. With this method, a total of 3,549 data samples were used for the test dataset. The evaluation metrics applied to these methods include the accuracy, precision, recall, specificity, and the F1 score, which were defined as follows.

$$\:Accuracy=\:\frac{TP+TN}{TP+FP+FN+TN}$$
(2)
$$\:Precision=\:\frac{TP}{TP+FP}$$
(3)
$$\:Recall=\:\frac{TP}{TP+FN}$$
(4)
$$\:Specificity=\:\frac{TN}{FP+TN}$$
(5)
$$\:F1=\:\frac{2\times\:Precision\times\:Recall}{Precision+Recall}$$
(6)

Here, PD is defined as the positive class and HC as the negative class. The true positive (TP) is the number of PD-labeled data correctly classified as positive. The true negative (TN) is the number of HC-labeled data correctly classified as negative. The false positive (FP) is the number of HC-labeled data misclassified as positive and false negative (FN) is the number of PD-labeled data misclassified as negative.

Effect of excluding input data

As the number of sensors increases, the system becomes more complex, and the computational cost tends to rise. Therefore, it is desirable to minimize the input data while maintaining the classification performance. We investigated changes in the model’s performance by systematically excluding each individual input data. The CNN model was trained and evaluated under the same conditions as when all input data was included (baseline), except for the excluded layer or node. The excluded input data was selected at the sensor level as follows: force and moment signals from the headset spacer, force and moment signals from the seat post, lateral position, speed, and personal information. In order to compare the performance across 5 different conditions for each fold, we compared the difference in accuracy against the baseline, representing it in percentages as shown below:

$$\:\varDelta\:\:{Accuracy}_{Rel}=\:\frac{{Accuracy}_{condition}-\:{Accuracy}_{baseline}}{{Accuracy}_{baseline}}\:\times\:\:100\%$$
(7)

Results

The total number of data samples collected from 29 PD patients and 36 healthy elderly individuals was 17,745. Table 2 presents the results for subject-level evaluation for each fold and the average performance obtained from cross-validation. Figure 5 shows the aforementioned classification results in confusion matrices: the average accuracy is 0.861, precision is 0.900, recall is 0.833, specificity is 0.886, and the F1 score is 0.836.

Table 2 The results of evaluation metrices according to subject-level evaluation.
Full size table
Fig. 5
figure 5

Confusion matrices for each fold obtained from subject-level evaluation. Numbers in matrices indicate the number of subjects in each case.

Full size image

Table 3 presents the results for dataset-level evaluation for each fold and the average performance obtained from cross-validation. Figure 6 shows the aforementioned results in confusion matrices: the average accuracy is 0.807, precision is 0.837, recall is 0.779, specificity is 0.828, and the F1 score is 0.777.

Table 3 The results of evaluation metrices according to dataset-level evaluation.
Full size table
Fig. 6
figure 6

Confusion matrices for each fold obtained by dataset-level evaluation. Numbers in matrices indicate the number of samples in each case.

Full size image

Figure 7 depicts the individual results of the CNN model for each data sample in trial 1 of the test set of fold 1. For subject #1, labeled as HC, all data samples from trial 1 were accurately classified as HC (Fig. 7a). Likewise, for subject #8, all data samples from trial 1 were accurately classified as PD (Fig. 7b). Conversely, for subject #11, labeled as PD, all data samples from trial 1 were misclassified as HC (Fig. 7b). Subject #12, labeled as PD, had 37 correctly classified data samples and 54 misclassified data samples (Fig. 7b). This would be considered misclassified for the subject-level evaluation method.

Fig. 7
figure 7

Results of the CNN model’s classification for data samples in trial 1 of the test set of fold 1. Each box represents an individual data sample, with gray boxes denoting those classified as HC and black boxes denoting those classified as PD. (a) Data samples with a true label of HC. (b) Data samples with a true label of PD.

Full size image

Figure 8 shows variations in classification performance resulting from an exclusion of the specific input data. The exclusion of lateral position data resulted in the most significant reduction in accuracy, 21.56% and 27.26%, for subject-level and individual dataset-level evaluation methods, respectively.

Fig. 8
figure 8

Changes in classification performance due to the exclusion of specific input data. The left bar graph and the right bar graph depict the differences in accuracy compared to the baseline using subject-level and individual dataset-level evaluation methods, respectively. Each bar represents the difference in average accuracy across the folds when excluding input data corresponding to the legend on the right. The black dots aligned with the bars on the x-axis represent the accuracy differences for each fold.

Full size image

Discussion

In this study we proposed a 1-D CNN-based deep learning model designed to classify PD patients and HC groups using kinematic and kinetic signals obtained during indoor cycling. The cycling experiments were conducted on a platform, Ultiracer, which simulates outdoor environments by enabling steering and lateral movement of the bicycle. The proposed model achieved an accuracy of 86.1% and 80.7% for subject-level and individual dataset-level evaluation methods, respectively.

It can be postulated that our system’s performance is comparable to previous approaches. Tarjni et al. proposed a CNN-based classifier using MRI images with an accuracy of 88.9%32. However, this method presents challenges as obtaining MRI images is costly. Ruilin et al. and Sue et al. developed deep learning models for classifying PD using EEG data, achieving accuracies of 99.9% and 88.3%, respectively33,34. However, the demerits of utilizing EEG are that experts are needed to place the sensors, and it is common to apply gel to the head in order to obtain good signal quality. This could present as a burden for continuous and consistent PD monitoring. Clayton et al. introduced an approach using a smart pen, and Hankan et al. introduced an approach utilizing vocal data, achieving accuracies of 95.7% and 88.8%, respectively37,38. Helber et al. introduced an approach using signals from IMU sensors during gait, achieving an accuracy of 83.7%35, and Imanne et al. introduced an approach using vertical ground reaction force signals during gait, achieving an accuracy of 98.7%36.

Our proposed approach has several advantages in the following aspects. Our system may possibly be used for early detection of PD which is important for patient care. We attribute this to the fact that the data used to train and test the CNN model were obtained from PD subjects who exhibited only mild symptoms and received a H&Y scale of approximately 2 (stage 1: 5, stage 2: 21, stage 2.5: 1, and stage 3: 2 among a total of 29). Also, the CNN model proposed in this study may be able to serve as a quantitative assessment tool by not simply identifying the presence of PD but also providing an estimated likelihood. Furthermore, the subjects do not need to attach or wear sensors on their bodies, and there is no need for a pre-calibration process. Additionally, our method is safe because there is a low fall risk due to the pole on the bicycle platform that ensures an upright position while encompassing the advantages of outdoor cycling such as balance control and situational awareness.

Meanwhile, we segmented all data into 2-second windows in this study. This was based on a comparative analysis across various window sizes, which showed that the 2-second segments provided relatively better classification performance (Fig. S1). Considering that a single pedaling cycle typically takes approximately one second, a 2-second window can effectively capture the complete cycle pattern while maximizing the number of training samples. We also compared the proposed CNN-based model with deep learning models based on long short-term memory (LSTM) and Transformer architecture to verify the validity of our model (Tables S1, S2, and Fig. S2). As a result, the CNN-based model achieved the highest performance.

While deep learning models such as CNNs have long been criticized for their ‘black box’ nature47, recent advances in explainable AI introduced robust methods to elucidate their decision-making processes48. In this study, we first tried Gradient-weighted Class Activation Mapping (Grad-CAM) to identify temporal regions and signal components that most strongly influence classification outcomes. However, significant heterogeneity was observed in importance score maps across individual samples, which made it difficult to extract consistent patterns. Figure S3 presents representative examples of Grad-CAM visualizations. Next, we conducted ablation analysis by systematically excluding individual input modalities. Although this approach reduced temporal resolution, it offered two valuable insights: (1) finding the minimal sensor configuration that maintains adequate classification performance, and (2) identifying which input signals have the highest diagnostic value for PD detection. Notably, classification accuracy decreased by 13-28% in all modality-exclusion conditions as shown in Fig. 8, suggesting that all input modalities were utilized in our CNN model. Among them, the most influential input was the lateral position. PD patients are known to rely more on continuous visual information than healthy individuals49. In our system, only the lateral position data are visually fed back to the patient (Fig. 1a and b), which may make it more influential. Additionally, maintaining lateral stability during cycling may engage mechanisms similar to those involved in postural control, which are known to be compromised even in early-stage PD50. Taken together, our findings suggest that the lateral position signal may reflect subtle motor control deficits preceding overt clinical symptoms, underscoring its potential utility for early PD detection.

We believe that our system holds promise not only for monitoring the likelihood of PD, but also for delivering rehabilitation exercises. Prior studies have highlighted the therapeutic value of cycling for PD patients, showing improvements across a range of domains including motor function, balance, gait, bradykinesia, executive function, and overall quality of life41,42,43,44. By integrating both assessment and rehabilitation capabilities, the system could address limitations in portability and size, ultimately offering a more comprehensive and clinically practical solution for PD management.

Conclusion

In summary, we used a dataset comprised of 17,745 data samples collected from 65 subjects (29 PD and 36 HC) to demonstrate the potential for a CNN model to classify PD patients from HC groups using sensors attached to a bicycle and the Ultiracer. This study presents several important findings. We have shown that our CNN model has the potential for early detection of PD. Not only does our system provide PD classification, but also a quantitative likelihood by assessing the individual dataset-level results. This is available without having to attach any sensors to the subject in a safe environment.