Using resting-state functional magnetic resonance imaging and contrastive learning to explore changes in the Parkinson’s disease brain network and correlations with gait impairment

Introduction

Parkinson’s disease (PD) is a progressive, degenerative disease that manifests with motor symptoms such as slow movement, static tremors, rigidity, and gait impairments, as well as autonomic nervous dysfunction, cognitive impairment, depression and anxiety (1,2). The diversity of clinical symptoms of PD may be related to the involvement of different neuronal pathways and alterations in neurotransmitter activity (3-6). The pathology of PD includes the loss of dopaminergic neurons in the midbrain substantia nigra and the aggregation of α-synuclein in multiple systems (7,8). Clinical pathological studies have revealed errors in clinical diagnoses, which mainly depend on symptoms in 7–35% of cases (9-12). The above findings demonstrate the importance of exploring neuroimaging changes in patients with PD to identify targets for early diagnosis and intervention.

Imaging biomarkers associated with neuroplasticity can be obtained from resting-state functional magnetic resonance imaging (rs-fMRI) data (13,14). Rs-fMRI can be used to detect early brain functional connectivity (FC) changes that may precede structural damage (15). Furthermore, fluctuations in brain blood flow and the similarity of blood oxygen level-dependent signals among different brain regions can be determined to calculate the strength of connections in resting-state networks (16). Currently, abnormal FC in brain networks of patients with PD with different symptoms has been identified through rs-fMRI studies (17,18). FC has been related to impairments of patients with PD, including motor and cognitive symptoms (16). The typical networks involved are the default mode network (DMN), sensorimotor network (SM), visual network (VN) and dorsal attention network (DAN), with FC abnormalities typically related to the severity of symptoms (19-23). However, neural networks constructed by traditional methods cannot comprehensively consider higher dimensional features in FC analyses, and the clinical symptoms of PD are complex. These factors increase the difficulty in classifying PD and establishing associations between brain changes and symptoms.

In recent years, with the development of artificial intelligence technology, many deep learning neural network models and multivariate machine learning methods have been developed based on magnetic resonance imaging (MRI) features to distinguish patients with PD from healthy controls (HCs) or atypical patients with PD. Compared with traditional machine learning models, deep learning neural network models have better diagnostic performance. The accuracy of such models reaches 78–96.8%, and the area under the curve (AUC) reaches 77.5–98% (24-30). However, these models have not been applied to rs-fMRI FC analyses of patients with PD, as existing approaches are typically based on structural images, mostly T1-weighted images or diffusion tensor images.

Therefore, the application of a neural network model to brain network FC analysis is both promising and innovative. As a deep learning model, a convolutional neural network (CNN) can learn based on raw high-dimensional data, reducing the complexity of the feature selection and extraction processes. Consequently, CNNs are superior to traditional methods and machine learning methods (31,32). In this work, we apply a CNN to brain network analysis, termed BrainNetCNN. The proposed model is a CNNs for brain networks, including edge-to-edge, edge-to-node, and node-to-graph convolutional layers, which are specific cases of more generalized convolutional filters with meaningful interpretations in terms of network topology (33). Additionally, CNNs alone cannot extract discriminative features for FC analyses, which are high dimensional with considerable noise. Therefore, we propose a feature extraction method based on contrastive learning (CL) to learn higher-dimensional self-supervised features by constructing positive and negative example pairs.

In addition, gait impairments in patients with PD often involve impaired step speed, step length, and rhythm. However, the low sensitivity of existing approaches is affected by objective factors, which limits their applicability in disease assessment (34). With the development of quantitative gait assessment methods, more objective and convenient gait monitoring methods with increased sensitivity have been proposed. Recently, the Zeno walkway (Proto Kinetics LLC, Havertown, PA, USA) has been developed, which measures spatiotemporal gait parameters by recording the positions of the feet over time and kinetic gait properties such as the centre-of-pressure. This device can detect subtle deviations in gait and has been used for PD; however, it has not been widely applied because of its high cost (35,36). One advantage of our study is the use of the Zeno walkway to extract gait features.

In the present study, we innovatively used a CNN model combined with a self-supervised CL method, named BrainNetCNN + CL. We hypothesized that the diagnostic performance of BrainNetCNN + CL for differentiating patients with PD and HCs would enable us to more precisely explore abnormally connected regions, with these regions serving as potential therapeutic targets. First, BrainNetCNN + CL was applied to analyse rs-fMRI data to evaluate its effectiveness in differentiating patients with PD and HCs and to compare its performance advantages with those of other machine learning methods, such as support vector machines (SVMs) and multilayer perceptrons (MLPs). Then, we investigated correlations between the significant connections identified by BrainNetCNN + CL and symptoms such as gait impairments and mental changes. We present this article in accordance with the STROBE reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1227/rc).

Methods

Participants enrolment and assessment

We used a cross-sectional study. Ethics, informed consent and treatment of patients in this study were consistent with previous study (37). This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This study was approved by the ethics committees of Beijing Friendship Hospital, Capital Medical University (No. 2019-P2-283-02). Demographics and clinical parameters are shown in Tables 1,2. All participants provided written informed consent prior to participation. Patients with PD receiving optimal treatment with dopaminergic medications were recruited from the Movement Disorders Program at Beijing Friendship Hospital of the Capital Medical University. All PD participants were evaluated by a motor neurologist with 2 years of clinical neurology experience and another physician with 20 years of clinical neurology experience. All patients were right-handed and diagnosed with idiopathic PD according to the British Parkinson’s Association Brain Bank criteria (11). The exclusion criteria were atypical or secondary parkinsonism, confounding medical or psychiatric condition(s), any condition that limits the ability to provide informed consent, or other neurological diseases leading to motor impairments. Community volunteers without neurological and/or psychiatric disorders were recruited as HCs. Patients with poor image scan quality were not included in the study. Twenty-nine patients and 38 HCs were enrolled. General indicators, including age, sex, education level, duration of disease and levodopa equivalent daily dose (LEDD), were recorded for all participants.

Gait parameter assessment

Patients with gait impairment were identified on the basis of their history and the International Parkinson and Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) (38) part II and III (a score ≥1 item on 2.12 and 2.13 on the MDS-UPDRS-II or a score ≥1 on item 3.10 or 3.11 on the MDS-UPDRS-III). Participants with PD were assessed with the following tests to characterise the disease status and gait impairment: the New Freezing of Gait Questionnaire (nFOGQ) (39), the MDS-UPDRS, the Timed-Up and Go (TUG) tests (40), the Berg Balance Scale (BBS) (41) and the Hohen-Yahr stage (H-Y stage) (42). For all participants, objective gait assessments quantifying velocity, cadence, stride time and stride length at a self-selected pace (SSP) and fast pace (FP) were performed using a 20-foot-long computerized Zeno Walkway Gait Analysis System (Proto Kinetics, Havertown, PA, USA). Notably, many PD participants experienced motor fluctuations. Therefore, clinical assessments were performed in the “OFF” state (12 hours after the patient had last taken PD medication) and in the “ON” state with a supra-ON dose (125% of the morning levodopa equivalent dose on the same day) (43).

Non-gait parameter assessment

All participants completed the Parkinson’s Disease Questionnaire-39 (PDQ-39) (44) to assess their overall quality of life and the Mini-Mental State Examination (MMSE) (45) to evaluate their cognitive status. As comorbid anxiety and depression are widely reported with PD, the Modified Apathy Evaluation Scale (MAES) (46), Beck Anxiety Inventory (BAI) score (47) and Beck Depression Inventory (BDI) score (48) were used to assess the presence of comorbid mood and anxiety disorders. In patients with PD, the data were acquired during the ON state.

MRI data acquisition and preprocessing

Images were obtained using a 3.0-T MRI system (Prisma, Siemens, Erlangen, Germany) with a 64-channel phased-array head coil. High-resolution three-dimensional (3D) structural T1-weighted images parameters are consistent with our previous study (49). Resting-state functional images were acquired using an echo-planar imaging (EPI) sequence. The acquisition parameters were as follows: repetition time (TR) =2,000 ms, echo time (TE) =30 ms, flip angle (FA) =90°, field of view (FOV) =224 mm × 224 mm, slice thickness =2 mm, number of slices =62, matrix =112×112, voxel size =2 mm × 2 mm × 2 mm, bandwidth =2,232 Hz/Px, number of measurements =240, and acquisition time =8 min 13 s.

MRI data preprocessing was performed by using SPM12 software (http://www.fil.ion.ucl.ac.uk/spm) and DPABI (http://rfmri.org/dpabi). The preprocessing steps were as follows: (I) the DICOM data were converted to NIFTI data; (II) the first 5 EPI volumes were removed; (III) slice timing correction was performed; (IV) head motion correction was performed, and subjects with translation >3 mm or rotation >3° in any direction were excluded; (V) the mean functional image and structural image were aligned, and the mean image was coregistered with the corresponding segmented structural T1-weighted images (resampled to a voxel size of 3×3×3 mm³); (VI) interference signal regression was performed [6 head parameters, white matter, cerebrospinal fluid (CSF) signals, and global signals]; and (VII) bandpass filtering (0.01< f <0.08 Hz) was performed.

Framework and indices of the self-supervised neural network

FC

After image preprocessing, 3D brain imaging data with temporal dimensions were obtained. We used an automated anatomical labelling (AAL) brain atlas to divide the brain regions in the functional magnetic resonance imaging (fMRI) data. This template divided the cerebral cortex into 90 regions of interest with similar sizes and related functional activities. The time series of each region of interest was obtained by averaging the time series of all voxel points within the region of interest. Finally, the average time series of each region of interest was used to calculate the Pearson correlation coefficient between pairs of regions of interest to obtain an FC matrix with a size of 90×90. Each element in the FC matrix represented the strength of the FC between two brain regions of interest.

Algorithm framework and prediction model

We first introduce the overall framework for diagnosing and analysing PD based on a self-supervised neural network. The model consists of three main parts. The first part is the encoder, in which a CNN is used to extract features from rs-fMRI data. The encoder consists of three different convolutional layers: an edge-to-edge layer, an edge-to-node layer and a node-to-graph layer. The processing was performed according to previous methods (33). The second part is CL methods in self-supervised learning. The methods take the brain FC features and FC features after masking transformations as inputs, and the hidden layer features are obtained by the encoder. Then, positive and negative sample pairs are constructed, and the neural network parameters are jointly optimized via a comparison recognition task in the self-supervised learning network. Thus, the model can learn self-supervised feature representations with higher dimensionality and reduce the distance between similar samples. The third part is the classification module, which consists of several fully connected layers. The hidden features are fused with phenotypic information such as the BAI, BDI, MAES, BBS, TUG, and MMSE scores and objective gait parameters. The fused features are used to predict the brain FC category labels for PD classification. During the training process, the classification task and the self-supervised auxiliary task were jointly performed for parameter optimization.

Additional notes: in the fusion feature step of the classification task, we included different symptom categories, with the BAI, BDI, MAES, and MMSE scores as emotional and mental category features, as well as gait parameters for objective gait assessment, with the BBS and TUG scores as gait impairment category features. Because different types of phenotypic information have various units, we normalized and standardized this information separately to ensure that the features at different scales had the same numerical magnitude.

Evaluation indices

In this study, we use Adam’s gradient descent algorithm to minimize the loss function during the training process. The batch size of each iteration is 96 and the size of the learning rate at update is 1×10⁻⁴. The experiments were randomly divided into training set, validation set and test set in the ratio of 8:1:1, and the experimental performance is evaluated using 10-fold cross-validation. We also compare the performance of BrainNetCNN + CL with other machine learning methods. Six widely used evaluation metrics, the accuracy (ACC), sensitivity (SEN), specificity (SPE), balanced accuracy (BA), positive predictive value (PPV) and negative predictive value (NPV), were used to evaluate the model classification performance.

Brain network analysis

The brain regions associated with the top 20 functional connections that contributed the most to the classification task were summarized. The brain regions were mapped to the Yeo7 network template (50) to determine the brain networks that contributed most to the classification performance.

Statistical analysis

Continuous variables that followed a normal distribution were presented as the average ± standard deviation and were compared with a two-sample t-test. Continuous variables that followed a nonnormal distribution were presented as the medians and interquartile ranges and were compared with the Wilcoxon rank sum test. Differences between classified variables were compared with the Chi-squared test. P<0.05 indicated statistical significance. Pearson’s correlation coefficients were calculated to assess the relationship between the FC values and patient-related rating scale scores, using gender, age, education, and head movement as covariates.

Results

Demographics and behavioural characteristics of the study participants

The demographic and clinical characteristics of the study participants are shown in Table 1, and the objective gait analysis results are shown in Table 2. Normality testing revealed that not all the measures met the normality assumptions (Tables 1,2). There were no significant differences in sex or MAES score between the two groups. However, there were significant differences in age, education level, and MMSE, BAI, BDI, TUG and BBS scores between the two groups. The MMSE, BAI, BDI, and TUG scores in the PD group were significantly higher than those in the control group (Table 1), and the BBS scores in the PD group were significantly lower than those in the control group. Disease severity, as measured by the MDS-UPDRS-III score, MDS-UPDRS-Total score, TUG score and BBS score, was significantly lower in the OFF state compared to that in the ON state in patients with PD (Table 1). According to our objective gait analysis, patients with PD exhibited lower velocities and shorter stride lengths than did HCs when moving at the SSP and FP (Table 2).

Classification performance based on the different features of various methods

As shown in Table 3, we compared the classification performance of BrainNetCNN + CL with that of different methods and across various features. Among the other classification methods, BrainNetCNN + CL had the highest BA and ACC values. Furthermore, BrainNetCNN + CL obtained the best classification performance when using FC features and all clinical features, with the highest ACC value of 0.80 and highest BA value of 0.78 (Table 3).

FC and brain network analysis

We visualized the 20 FCs that contributed most significantly to the classification performance when applying BrainNetCNN + CL (Figure 1, Table 4). According to the Yeo7 brain network template (50), the 20 FCs were mainly associated with the DMN, ventral attention network (VAN), and limbic network (LN).

Figure 1 Visualization of the 20 connections with the greatest contributions to the classification results. The thickness of each line indicates the contribution to the classification results. (A) A colourful circular representation altered functional connectivity between patients with PD and HCs. (B) Full view (frontal, profile, top) of altered functional connectivity between patients with PD and HCs. L, left; R, right; PD, Parkinson’s disease; HCs, healthy controls.

Correlation analysis

In the PD group, the strength of the FC between the left superior temporal gyrus and the right inferior frontal gyrus of the triangular part was negatively correlated with the patients’ velocity at their SSP and the patients’ stride length in the OFF state (P=0.021, r=−0.433; P=0.015, r=−0.456). Furthermore, the strength of the FC between the right inferior occipital gyrus and the left postcentral gyrus in the PD group was negatively correlated with the patients’ stride length at their SSP in the ON state (P=0.001, r=−0.589). In addition, the strength of the FCs between the right fusiform gyrus and the right calcarine fissure and surrounding cortex were negatively correlated with the BAI score (P=0.032, r=−0.406) and positively correlated with the BBS-ON score (P=0.037, r=0.395). Finaly, the strength of the FC between the right supplementary motor area and the left parahippocampal gyrus was negatively correlated with the patients’ stride time at their FP in the ON state (P=0.036, r=−0.398) (Figure 2).

Figure 2 Correlation analyses revealed that the strength of the FCs between the left superior temporal gyrus and the right inferior frontal gyrus of the triangular part (A,B), the right inferior occipital gyrus and the left postcentral gyrus (C), the right fusiform gyrus and the right calcarine fissure and surrounding cortex (D,E), and the right supplementary motor area and the left parahippocampal gyrus (F) were strongly correlated with clinical scale scores and gait parameters. The strength of the FC between the left superior temporal gyrus and the right inferior frontal gyrus of the triangular part was negatively correlated with the patient’s velocity at their SSP and stride length in the OFF state (P=0.021, r=−0.433; P=0.015, r=−0.456); the strength of the FC between the right inferior occipital gyrus and the left postcentral gyrus in the PD group was negatively correlated with the stride length at the SSP in the ON state (P=0.001, r=−0.589); the strength of the FC between the right fusiform gyrus and the right calcarine fissure and surrounding cortex was negatively correlated with the BAI score (P=0.032, r=−0.406) and positively correlated with the BBS-ON score (P=0.037, r=0.395) and the strength of the FC between the right supplementary motor area and the left parahippocampal gyrus was negatively correlated with the patients’ stride time at the FP in the ON state (P=0.036, r=−0.398). SSP, self-selected pace; BBS-ON, Berg Balance Scale during the “ON” state; BAI, Beck Anxiety Inventory; FP, fast pace; L, left; R, right; FC, functional connectivity; PD, Parkinson’s disease.

Discussion

We screened for significant FC changes in patients with PD by using BrainNetCNN + CL to our knowledge, this is the first work to identify resting-state FCs in patients with PD based on a self-supervised neural network. In this study, BrainNetCNN + CL achieved better diagnostic performance than machine learning methods such as SVMs and MLPs; furthermore, BrainNetCNN + CL showed better performance than BrainNetCNN alone. Moreover, the FCs that contributed substantially to the classification performance were significantly correlated with gait and nonmotor symptoms. In the PD and HC groups, the BAI score, BDI score, BBS score and TUG score were significantly different, as were the velocity and stride length. Our study results demonstrated that the FC between the inferior occipital gyrus and postcentral gyrus and the FC between the supplementary motor area and parahippocampal gyrus in patients with PD are correlated with gait impairment. These networks and abnormal FC may be potential mechanisms underlying gait impairments in patients.

Classification performance of BrainNetCNN + CL

Deep learning has been increasingly used in the study of neurological disorders such as Alzheimer’s disease, PD, and autism spectrum disorder and has shown advantages over machine learning classification models (51). However, no studies have used deep learning to investigate neuroimaging-based diagnosis of PD based on the correlated brain network features within or between patients with PD and HCs. Moreover, the heterogeneity of symptoms and the high-dimensional intranetwork and internetwork association features of the variables complicate the classification and diagnosis of PD, demonstrating the need for more sensitive and efficient methods. In BrainNetCNN, the input FC matrices are passed through unique convolutional filters, which make full use of the topological locations within the brain network to obtain feature information. The proposed deep learning framework is specifically designed for brain networks. CL is used to improve the effectiveness of classifying different features by constructing pairs of positive and negative samples and reducing the distance between similar samples (52). Our introduction of CL into BrainNetCNN resulted in a 3% increase in diagnostic accuracy. Moreover, our BrainNetCNN + CL model has the highest BA value in classifying patients with PD and HCs, which is 3–6% greater than the BA values of the other three methods. This finding suggested that our proposed model is an optimal method for capturing multivariate features, including FC data.

Alterations in brain networks in patients with PD compared to those of HCs

In the present study, the DMN was the brain network that contributed most to the classification of patients with PD and HCs, and among the 20 FCs that contributed most to the classification results, four FCs were intranetwork connections in the DMN. Previous studies have shown significant differences in the DMN between patients with PD and HCs. Patients with a greater FC in the DMN have increased default mode states, which may account for reduced self-awareness in patients with PD (53,54). This difference is thought to be related to cognitive impairment in patients with PD. However, previous research has focused on the relationship between nonmotor symptoms in patients with PD and changes in the strength of intranetwork connectivity in the DMN (53,55), while few studies have investigated the correlations between the DMN and motor symptoms or gait impairments in patients with PD. We found an association among enhanced FC between the left superior temporal gyrus and the right inferior frontal gyrus of the triangular part, the intranetwork FC of the DMN, and reduced velocity and step length at the SSP. This finding suggested that the DMN and gait were correlated in patients with PD. Previous works have suggested that the DMN may act as a hub for multiple networks in response to patient movement (56,57). Deep brain stimulation treatment induces changes in brain networks, including the superior temporal gyrus, suggesting that it is also associated with post-treatment response (58). The mechanism through which the DMN influences gait in patients with PD through connections with other networks needs to be further explored.

In addition to the DMN, the VAN and the limbic network (LN) contributed considerably to the classification results. Our results suggest that the VAN and LN are altered in patients with PD compared with those in HCs. Existing models of visual misperception and hallucinations in patients with PD suggest that Lewy vesicles are clustered in the LN and that the LN leads to activation of the VAN. In addition, dopamine deficiency in patients with PD exacerbates ganglion damage within the VAN (59,60). Our results confirmed that the VAN and LN are altered in patients with PD and that these networks may be associated with symptoms such as visual misperception and hallucinations in patients with PD.

FC correlates with patient clinical characteristics and gait parameters

To assess the relationships between motor and nonmotor symptoms and brain networks in patients with PD, the FC values of the brain networks were correlated with gait parameters and clinical characteristics. We found that as the strength of the FC between the right inferior occipital gyrus and the left postcentral gyrus increased, the step length of patients at the SSP decreased in the ON state. Moreover, the weaker the FC between the right fusiform gyrus and the right calcarine fissure and surrounding cortex was, the greater the BAI scores and the lower the BBS scores of the patients. Furthermore, weaker FC between the right supplementary motor area and the right parahippocampal gyrus was correlated with longer stride time at the FP in the ON state. These correlations suggest that intra- and internetwork FC of brain networks involved in cognition and complex motor execution are associated with gait impairments, motor impairments, and clinical features in patients with PD. Gait impairments are complex and involve cognitive, visual, and motor-related brain regions, and gait impairments in patients with PD are accompanied by altered FC in brain regions involved in cognition and complex motor execution (23,61,62).

Clinical features of patients with PD and HCs

The clinical manifestations of PD can be divided into motor and nonmotor symptoms (63) . The MMSE, BAI, and BDI scores were significantly higher in the PD group than in the HC group. This finding suggested that patients with PD had more anxiety problems than HCs, which is a manifestation of nonmotor symptoms in patients with PD. We assessed the gait function of the patients with PD and HCs using objective gait parameters and clinical scales. The TUG test scores were significantly higher for patients with PD than for HCs, the BBS was significantly lower than in HCs, and the stride speed and stride length at the SSP and FP were significantly lower in patients with PD than in HCs. These findings indicate that patients with PD have more movement and balance problems than do HCs (64). Levodopa is the most commonly used drug for treating PD (1). The patients with PD had lower MDS-UPDRS-III scores, MDS-UPDRS-Total scores, and TUG scores and higher BBS scores in the ON state than in the OFF state, suggesting that levodopa ameliorated the patients’ motor symptoms.

One advantage of this study is the proposal of the first deep learning model that can differentiate patients with PD and HCs based on brain networks. We also investigated the correlations between the FC strength and objective gait parameters in patients with PD. There are some limitations in this study. First, due to the age limitation of the patients, only 29 patients with PD and 38 HCs were included in this study, which is a small sample. Additionally, there was a statistically significant difference in age between the patients with PD and the HCs in this study, so we added covariates during the correlation analysis. A larger sample size and multicentre validation are needed in future work. In addition, although deep learning models have advantages with high-dimensional data, the consistency between the deep learning results and the results of traditional methods remains to be verified.

Conclusions

In conclusion, this study showed that BrainNetCNN + CL shows good performance in differentiating patients with PD. According to the BrainNetCNN + CL results, patients with PD presented alterations in the DMN, VAN, and LN. Moreover, the results showed that the intranetwork FC of the above brain networks was correlated with the gait parameters and clinical features of the patients. These results suggest that BrainNetCNN + CL is useful for identifying regions with abnormal connectivity, which may serve as potential therapeutic targets to address gait impairment associated with PD.

Acknowledgments

Funding: This work was supported by the National Natural Science Foundation of China (Nos. 82202097 and 82373436), Beijing Scholars Program (No. [2015] 160 to Z.W.), Beijing Key Clinical Discipline Funding (No. 2021-135), Beijing Hospitals Authority Youth Programme (Nos. QML20230113 and QML20230114), Training Fund for Open Projects at Clinical Institutes and Departments of Capital Medical University (No. CCMU2022ZKYXY010), and Tongzhou District Health Development Program (No. KJ2024CX023).

Footnote

Reporting Checklist: The authors have completed the STROBE reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-24-1227/rc

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-24-1227/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This study was approved by the ethics committees of Beijing Friendship Hospital, Capital Medical University (No. 2019-P2-283-02). All participants provided written informed consent prior to participation.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bloem BR, Okun MS, Klein C. Parkinson's disease. Lancet 2021;397:2284-303. [Crossref] [PubMed]
2. Samii A, Nutt JG, Ransom BR. Parkinson's disease. Lancet 2004;363:1783-93. [Crossref] [PubMed]
3. Mirelman A, Bonato P, Camicioli R, Ellis TD, Giladi N, Hamilton JL, Hass CJ, Hausdorff JM, Pelosin E, Almeida QJ. Gait impairments in Parkinson's disease. Lancet Neurol 2019;18:697-708. [Crossref] [PubMed]
4. Monaghan AS, Gordon E, Graham L, Hughes E, Peterson DS, Morris R. Cognition and freezing of gait in Parkinson's disease: A systematic review and meta-analysis. Neurosci Biobehav Rev 2023;147:105068. [Crossref] [PubMed]
5. Dalrymple WA, Huss DS, Blair J, Flanigan JL, Patrie J, Sperling SA, Shah BB, Harrison MB, Druzgal TJ, Barrett MJ. Cholinergic nucleus 4 atrophy and gait impairment in Parkinson's disease. J Neurol 2021;268:95-101. [Crossref] [PubMed]
6. Kamble N, Yadav R, Lenka A, Kumar K, Nagaraju BC, Pal PK. Impaired sleep quality and cognition in patients of Parkinson's disease with REM sleep behavior disorder: a comparative study. Sleep Med 2019;62:1-5. [Crossref] [PubMed]
7. Dickson DW. Neuropathology of Parkinson disease. Parkinsonism Relat Disord 2018;46:S30-3. [Crossref] [PubMed]
8. Cao Q, Han X, Tang D, Qian H, Yan K, Shi X, Li Y, Zhang J. Diagnostic value of combined magnetic resonance imaging techniques in the evaluation of Parkinson disease. Quant Imaging Med Surg 2023;13:6503-16. [Crossref] [PubMed]
9. Tolosa E, Garrido A, Scholz SW, Poewe W. Challenges in the diagnosis of Parkinson's disease. Lancet Neurol 2021;20:385-97. [Crossref] [PubMed]
10. Adler CH, Beach TG, Hentz JG, Shill HA, Caviness JN, Driver-Dunckley E, Sabbagh MN, Sue LI, Jacobson SA, Belden CM, Dugger BN. Low clinical diagnostic accuracy of early vs advanced Parkinson disease: clinicopathologic study. Neurology 2014;83:406-12. [Crossref] [PubMed]
11. Hughes AJ, Daniel SE, Ben-Shlomo Y, Lees AJ. The accuracy of diagnosis of parkinsonian syndromes in a specialist movement disorder service. Brain 2002;125:861-70. [Crossref] [PubMed]
12. Koga S, Aoki N, Uitti RJ, van Gerpen JA, Cheshire WP, Josephs KA, Wszolek ZK, Langston JW, Dickson DW. When DLB, PD, and PSP masquerade as MSA: an autopsy study of 134 patients. Neurology 2015;85:404-12. [Crossref] [PubMed]
13. Claeys EHI, Mantingh T, Morrens M, Yalin N, Stokes PRA. Resting-state fMRI in depressive and (hypo)manic mood states in bipolar disorders: A systematic review. Prog Neuropsychopharmacol Biol Psychiatry 2022;113:110465. [Crossref] [PubMed]
14. Filippi M, Basaia S, Sarasso E, Stojkovic T, Stankovic I, Fontana A, Tomic A, Piramide N, Stefanova E, Markovic V, Kostic VS, Agosta F. Longitudinal brain connectivity changes and clinical evolution in Parkinson's disease. Mol Psychiatry 2021;26:5429-40. [Crossref] [PubMed]
15. Raimondo L, Oliveira ĹAF, Heij J, Priovoulos N, Kundu P, Leoni RF, van der Zwaag W. Advances in resting state fMRI acquisitions for functional connectomics. Neuroimage 2021;243:118503. [Crossref] [PubMed]
16. Filippi M, Sarasso E, Agosta F. Resting-state Functional MRI in Parkinsonian Syndromes. Mov Disord Clin Pract 2019;6:104-17. [Crossref] [PubMed]
17. Imperiale F, Agosta F, Canu E, Markovic V, Inuggi A, Jecmenica-Lukic M, Tomic A, Copetti M, Basaia S, Kostic VS, Filippi M. Brain structural and functional signatures of impulsive-compulsive behaviours in Parkinson's disease. Mol Psychiatry 2018;23:459-66. [Crossref] [PubMed]
18. Tessitore A, Cirillo M, De Micco R. Functional Connectivity Signatures of Parkinson's Disease. J Parkinsons Dis 2019;9:637-52. [Crossref] [PubMed]
19. Hepp DH, Foncke EMJ, Olde Dubbelink KTE, van de Berg WDJ, Berendse HW, Schoonheim MM. Loss of Functional Connectivity in Patients with Parkinson Disease and Visual Hallucinations. Radiology 2017;285:896-903. [Crossref] [PubMed]
20. Devignes Q, Bordier C, Viard R, Defebvre L, Kuchcinski G, Leentjens AFG, Lopes R, Dujardin K. Resting-State Functional Connectivity in Frontostriatal and Posterior Cortical Subtypes in Parkinson's Disease-Mild Cognitive Impairment. Mov Disord 2022;37:502-12. [Crossref] [PubMed]
21. Tessitore A, Esposito F, Vitale C, Santangelo G, Amboni M, Russo A, Corbo D, Cirillo G, Barone P, Tedeschi G. Default-mode network connectivity in cognitively unimpaired patients with Parkinson disease. Neurology 2012;79:2226-32. [Crossref] [PubMed]
22. Tang S, Wang Y, Liu Y, Chau SW, Chan JW, Chu WC, Abrigo JM, Mok VC, Wing YK. Large-scale network dysfunction in α-Synucleinopathy: A meta-analysis of resting-state functional connectivity. EBioMedicine 2022;77:103915. [Crossref] [PubMed]
23. Wu H, Wu C, Qin J, Zhou C, Tan S. DuanMu X, Guan X, Bai X, Guo T, Wu J, Chen J, Wen J, Cao Z, Gao T, Gu L, Huang P, Zhang B, Xu X, Zhang M. Functional connectome predicting individual gait function and its relationship with molecular architecture in Parkinson's disease. Neurobiol Dis 2023;184:106216. [Crossref] [PubMed]
24. Pahuja G, Prasad B. Deep learning architectures for Parkinson's disease detection by using multi-modal features. Comput Biol Med 2022;146:105610. [Crossref] [PubMed]
25. Chakraborty S, Aich S, Kim HC. Detection of Parkinson's Disease from 3T T1 Weighted MRI Scans Using 3D Convolutional Neural Network. Diagnostics (Basel) 2020;10:402. [Crossref] [PubMed]
26. Shinde S, Prasad S, Saboo Y, Kaushick R, Saini J, Pal PK, Ingalhalikar M. Predictive markers for Parkinson's disease using deep neural nets on neuromelanin sensitive MRI. Neuroimage Clin 2019;22:101748. [Crossref] [PubMed]
27. Zhang X, He L, Chen K, Luo Y, Zhou J, Wang F. Multi-View Graph Convolutional Network and Its Applications on Neuroimage Analysis for Parkinson's Disease. AMIA Annu Symp Proc 2018;2018:1147-56. [PubMed]
28. Souza R, Wilms M, Camacho M, Pike GB, Camicioli R, Monchi O, Forkert ND. Image-encoded biological and non-biological variables may be used as shortcuts in deep learning models trained on multisite neuroimaging data. J Am Med Inform Assoc 2023;30:1925-33. [Crossref] [PubMed]
29. Tsuda M, Asano S, Kato Y, Murai K, Miyazaki M. Differential diagnosis of multiple system atrophy with predominant parkinsonism and Parkinson's disease using neural networks. J Neurol Sci 2019;401:19-26. [Crossref] [PubMed]
30. Zhao H, Tsai CC, Zhou M, Liu Y, Chen YL, Huang F, Lin YC, Wang JJ. Deep learning based diagnosis of Parkinson's Disease using diffusion magnetic resonance imaging. Brain Imaging Behav 2022;16:1749-60. [Crossref] [PubMed]
31. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:436-44. [Crossref] [PubMed]
32. Meszlényi RJ, Buza K, Vidnyánszky Z. Resting State fMRI Functional Connectivity-Based Classification Using a Convolutional Neural Network Architecture. Front Neuroinform 2017;11:61. [Crossref] [PubMed]
33. Kawahara J, Brown CJ, Miller SP, Booth BG, Chau V, Grunau RE, Zwicker JG, Hamarneh G. BrainNetCNN: Convolutional neural networks for brain networks; towards predicting neurodevelopment. Neuroimage 2017;146:1038-49. [Crossref] [PubMed]
34. Li F, Harmer P, Fitzgerald K, Eckstrom E, Stock R, Galver J, Maddalozzo G, Batya SS. Tai chi and postural stability in patients with Parkinson's disease. N Engl J Med 2012;366:511-9. [Crossref] [PubMed]
35. Vallabhajosula S, Humphrey SK, Cook AJ, Freund JE. Concurrent Validity of the Zeno Walkway for Measuring Spatiotemporal Gait Parameters in Older Adults. J Geriatr Phys Ther 2019;42:E42-50. [Crossref] [PubMed]
36. Sabo A, Gorodetsky C, Fasano A, Iaboni A, Taati B. Concurrent Validity of Zeno Instrumented Walkway and Video-Based Gait Features in Adults With Parkinson's Disease. IEEE J Transl Eng Health Med 2022;10:2100511. [Crossref] [PubMed]
37. Yan Y, Wang Z, Wei W, Yang Z, Guo L, Wang Z, Wei X. Correlation of brain iron deposition and freezing of gait in Parkinson's disease: a cross-sectional study. Quant Imaging Med Surg 2023;13:7961-72. [Crossref] [PubMed]
38. Goetz CG, Tilley BC, Shaftman SR, Stebbins GT, Fahn S, Martinez-Martin P, et al. Movement Disorder Society-sponsored revision of the Unified Parkinson's Disease Rating Scale (MDS-UPDRS): scale presentation and clinimetric testing results. Mov Disord 2008;23:2129-70. [Crossref] [PubMed]
39. Nieuwboer A, Rochester L, Herman T, Vandenberghe W, Emil GE, Thomaes T, Giladi N. Reliability of the new freezing of gait questionnaire: agreement between patients with Parkinson's disease and their carers. Gait Posture 2009;30:459-63. [Crossref] [PubMed]
40. Podsiadlo D, Richardson S. The timed "Up & Go": a test of basic functional mobility for frail elderly persons. J Am Geriatr Soc 1991;39:142-8. [Crossref] [PubMed]
41. Berg KO, Wood-Dauphinee SL, Williams JI, Maki B. Measuring balance in the elderly: validation of an instrument. Can J Public Health 1992;83:S7-11. [PubMed]
42. Hoehn MM, Yahr MD. Parkinsonism: onset, progression and mortality. Neurology 1967;17:427-42. [Crossref] [PubMed]
43. Tomlinson CL, Stowe R, Patel S, Rick C, Gray R, Clarke CE. Systematic review of levodopa dose equivalency reporting in Parkinson's disease. Mov Disord 2010;25:2649-53. [Crossref] [PubMed]
44. Zhang JL, Chan P. Reliability and validity of PDQ-39: a quality-of-life measure for patients with PD in China. Qual Life Res 2012;21:1217-21. [Crossref] [PubMed]
45. Folstein MF, Folstein SE, McHugh PR. "Mini-mental state". A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res 1975;12:189-98. [Crossref] [PubMed]
46. Marin RS, Biedrzycki RC, Firinciogullari S. Reliability and validity of the Apathy Evaluation Scale. Psychiatry Res 1991;38:143-62. [Crossref] [PubMed]
47. Beck AT, Epstein N, Brown G, Steer RA. An inventory for measuring clinical anxiety: psychometric properties. J Consult Clin Psychol 1988;56:893-7. [Crossref] [PubMed]
48. BECK AT. WARD CH, MENDELSON M, MOCK J, ERBAUGH J. An inventory for measuring depression. Arch Gen Psychiatry 1961;4:561-71. [Crossref]
49. Wei X, Wang Z, Zhang M, Li M, Chen YC, Lv H, Tuo H, Yang Z, Wang Z, Ba F. Brain Surface Area Alterations Correlate With Gait Impairments in Parkinson's Disease. Front Aging Neurosci 2022;14:806026. [Crossref] [PubMed]
50. Yeo BT, Krienen FM, Sepulcre J, Sabuncu MR, Lashkari D, Hollinshead M, Roffman JL, Smoller JW, Zöllei L, Polimeni JR, Fischl B, Liu H, Buckner RL. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J Neurophysiol 2011;106:1125-65. [Crossref] [PubMed]
51. Zhang L, Wang M, Liu M, Zhang D. A Survey on Deep Learning for Neuroimaging-Based Brain Disorder Analysis. Front Neurosci 2020;14:779. [Crossref] [PubMed]
52. Albelwi S. Survey on Self-Supervised Learning: Auxiliary Pretext Tasks and Contrastive Learning Methods in Imaging. Entropy (Basel) 2022;24:551. [Crossref] [PubMed]
53. Chen L, Huang T, Ma D, Chen YC. Altered Default Mode Network Functional Connectivity in Parkinson's Disease: A Resting-State Functional Magnetic Resonance Imaging Study. Front Neurosci 2022;16:905121. [Crossref] [PubMed]
54. Smallwood J, Bernhardt BC, Leech R, Bzdok D, Jefferies E, Margulies DS. The default mode network in cognition: a topographical perspective. Nat Rev Neurosci 2021;22:503-13. [Crossref] [PubMed]
55. Ruppert MC, Greuel A, Freigang J, Tahmasian M, Maier F, Hammes J, van Eimeren T, Timmermann L, Tittgemeyer M, Drzezga A, Eggers C. The default mode network and cognition in Parkinson's disease: A multimodal resting-state network approach. Hum Brain Mapp 2021;42:2623-41. [Crossref] [PubMed]
56. Wang X, Yoo K, Chen H, Zou T, Wang H, Gao Q, Meng L, Hu X, Li R. Antagonistic network signature of motor function in Parkinson's disease revealed by connectome-based predictive modeling. NPJ Parkinsons Dis 2022;8:49. [Crossref] [PubMed]
57. Karunanayaka PR, Lee EY, Lewis MM, Sen S, Eslinger PJ, Yang QX, Huang X. Default mode network differences between rigidity- and tremor-predominant Parkinson's disease. Cortex 2016;81:239-50. [Crossref] [PubMed]
58. Chu C, He N, Zeljic K, Zhang Z, Wang J, Li J, Liu Y, Zhang Y, Sun B, Li D, Yan F, Zhang C, Liu C. Subthalamic and pallidal stimulation in Parkinson's disease induce distinct brain topological reconstruction. Neuroimage 2022;255:119196. [Crossref] [PubMed]
59. Shine JM, Halliday GM, Naismith SL, Lewis SJ. Visual misperceptions and hallucinations in Parkinson's disease: dysfunction of attentional control networks? Mov Disord 2011;26:2154-9. [Crossref] [PubMed]
60. Stuart S, Lord S, Hill E, Rochester L. Gait in Parkinson's disease: A visuo-cognitive challenge. Neurosci Biobehav Rev 2016;62:76-88. [Crossref] [PubMed]
61. Song W, Raza HK, Lu L, Zhang Z, Zu J, Zhang W, Dong L, Xu C, Gong X, Lv B, Cui G. Functional MRI in Parkinson's disease with freezing of gait: a systematic review of the literature. Neurol Sci 2021;42:1759-71. [Crossref] [PubMed]
62. Gan L, Yan R, Su D, Liu Z, Miao G, Wang Z, Wang X, Ma H, Bai Y, Zhou J, Feng T. Alterations of structure and functional connectivity of visual brain network in patients with freezing of gait in Parkinson's disease. Front Aging Neurosci 2022;14:978976. [Crossref] [PubMed]
63. Reich SG, Savitt JM. Parkinson's Disease. Med Clin North Am 2019;103:337-50. [Crossref] [PubMed]
64. Lew M. Overview of Parkinson's disease. Pharmacotherapy 2007;27:155S-160S. [Crossref] [PubMed]