Aberrant functional connectivity and activity in Parkinson’s disease and comorbidity with depression based on radiomic analysis

Abstract Introduction The current diagnosis of Parkinson's disease (PD) comorbidity with depression (DPD) largely depends on clinical evaluation. However, the modality may tend to lack precision in detecting PD with depression. A radiomic approach that combines functional connectivity and activity with clinical scores has the potential to achieve accurate and differential diagnosis between PD and DPD. Methods In this study, we aimed to employ the radiomic approach to extract large‐scale features of functional connectivity and activity for differentiating among DPD, PD with no depression (NDPD), and healthy controls (HC). We extracted 6,557 features of five types from all subjects including clinical characteristics, resting‐state functional connectivity (RSFC), amplitude of low‐frequency fluctuation (ALFF), regional homogeneity (ReHo), and voxel‐mirrored homotopic connectivity (VMHC). Lasso, random forest, and support vector machine (SVM) were implemented for feature selection and dimension reduction based on the training sets, and the prediction performance for different methods in the testing sets was compared. Results The results showed that nineteen features were selected for the group of DPD versus HC, 34 features were selected for the group of NDPD versus HC, and 17 features were retained for the group of DPD versus NDPD. In the testing sets, Lasso prediction achieved the accuracies of 0.95, 0.96, and 0.85 for distinguishing between DPD and HC, NDPD and HC, and DPD and NDPD, respectively. Random forest achieved the accuracies of 0.90, 0.82, and 0.90 for distinguishing between DPD and HC, NDPD and HC, and DPD and NDPD, respectively, while SVM yielded the accuracies of 1, 0.86 and 0.65 for distinguishing between DPD and HC, NDPD and HC, and DPD and NDPD, respectively. Conclusions By identifying aberrant functional connectivity and activity as potential biomarkers, the radiomic approach facilitates a deeper understanding and provides new insights into the pathophysiology of DPD to support the clinical diagnosis with high prediction accuracy.


| INTRODUC TI ON
Depression is a frequent psychiatric symptom of Parkinson's disease (PD) and one of the earliest prodromal comorbidities that can significantly impact quality of life (Chagas et al., 2013). Nonmotor features including depression can appear in the earliest phase of the disease even before clinical motor impairment (Lix et al., 2010;Shearer et al., 2012;Tibar et al., 2018). The efficacy of medications and psychotherapies for treating depression in PD patients remains limited (Abós et al., 2017). Hence, advances in timely detection and concerted management of PD comorbidity with depression (DPD) become urgent. Motor symptoms were easily detected than nonmotor symptoms using the present diagnostic tools (Picillo et al. 2017).
According to the Unified Parkinson's Disease Rating Scale (UPDRS), over half DPD patients were not recognized by neurologists (Lachner et al. 2017), while the incidence of PD with depression was already substantially elevated recently (Kay et al., 2018). Clearly, physician recognition and current understanding for comorbidity of depression in PD are not enough.
Although knowledge of the neural and pathophysiologic mechanisms of DPD progression remains limited, many researchers are devoted to conduct research trying to understand the inner working mechanisms and discovering biomarkers of DPD. Clinical intervention is urgent around the early therapeutic windows (Tibar et al., 2018;Vu et al., 2012). Multimodal neuroimaging methods such as functional magnetic resonance imaging (MRI) and electroencephalography have aided the diagnosis of PD. Resting-state functional MRI (rs-fMRI) can provide more information on functional connections to assess the correlations among different networks. An intra-and internetwork functional connectivity study in DPD demonstrated aberrant functional connectivity (FC) in left frontoparietal network, basal ganglia network, salience network, and default-mode network (DMN) (Wei et al., 2017). Meanwhile, these connectivity anomalies were correlated with the depression severity in DPD. This may indicate the mechanism of progressive deterioration and compensation for integrative neural models in DPD (Wei et al., 2017;Zhu et al. 2016).
Structural MRI has also received research attention because of its stability and repeatability (Jacob et al., 2019;Remes et al., 2011).
Diffusion tensor imaging can discover microstructural changes in the brain white matter. Previous studies found abnormal white matter fiber characteristic (mainly located in the right arcuate fasciculus and bilateral middle cerebellar peduncles) in prodromal early stage of PD (Sanjari Moghaddam et al., 2019). Another microstructure difference was located in the bilateral white matter fiber of the mediodorsal thalamic regions between the DPD and NDPD groups, but the sample size was relatively small and the clinical score only included the Hamilton depression rating scale (HAMD) (Li et al., 2010).
In recent years, machine learning has been recognized as a promising and powerful algorithm method for prediction and medical diagnosis. Studies have been conducted to obtain voxel-based morphological biomarkers of PD by using machine learning such as support vector machine (SVM) or principal component analysis (PCA) that allowed individual differential diagnosis of PD (Lix et al., 2010;Palumbo et al., 2014;Salvatore et al., 2014). Another method (Peng et al., 2017;Peran et al., 2010) focusing on region of interest (ROI) has also been implemented where some specific regions of the brain such as gray matter and hippocampal volume were extracted based on prior knowledge regarding their effects on brain functionality and memory.
Recent progress in digital medical image analysis allows us to develop a novel feature extraction method called radiomics which converts large amounts of medical imaging characteristics into high-dimensional mineable data pool to build a predictive and descriptive model. The method has been applied to some neuropsychiatric diseases such as autism, schizophrenia, and Alzheimer disease (Feng et al., 2019;Salvatore et al., 2019). These findings demonstrate the validity of these radiomic approaches in improving the classification accuracy and discovering discriminative features that can reveal pathological information. A radiomic study on quantitative susceptibility mapping (QSM) achieved good performance for predicting PD (Xiao et al., 2019). The combination of radiomics features and convolutional neural networks (CNN) can increase the diagnostic accuracy (Ortiz et al., 2019). Other radiomic analysis focusing on longitudinal SPECT images and T2weighted MRI can also enhance the prediction accuracy of PD Rahmim et al., 2017). A radiomic study based on PET/ CT images extracted high-order features and trained a SVM model to classify PD and HC subjects, and the results demonstrated that the radiomic method combined with SVM could distinguish PD from HC . Cao et al. leveraging rs-fMRI radiomic features showed that machine learning methods including Lasso and SVM could significantly improve diagnostic accuracy of PD (Cao et al., 2020).
In the present study, we aimed to build and validate a radiomic method that can facilitate the individual diagnosis of patients with PD and the development of DPD by extracting whole-brain functional connectivity and activity using the radiomic approach. The proposed method can also identify brain regions of interest with aberrant functional activity between DPD and PD that were relevant to the disease onset, which may contribute to the early diagnosis and treatment for clinical practice.

depression, machine learning, Parkinson's disease, radiomics
This prospective study was approved by the institutional review board and followed the ethical guidelines of the Declaration of Helsinki, and written informed consent was acquired from each subject before inclusion.

| Participates and clinical evaluation
We used the same imaging data from the same recruited subjects as in our previously published issue (Cao et al., 2020). The only difference is that we further stratify the PD patients into two groups of DPD and NDPD to examine the aberrant functional connectivity and activity in DPD and to build machine learning models for predicting DPD and NDPD. Seventy PD patients including 21 DPD and 49 NDPD subjects were recruited, along with 50 matched healthy controls. The details regarding the diagnostic criteria and clinical evaluation of the NDPD and DPD groups are provided in Data S1.

2.2-2.5
Image data acquisition, preprocessing, extraction of radiomic features including regional homogeneity (ReHo), amplitude of low-frequency fluctuation (ALFF) and voxel-mirrored homotopic connectivity (VMHC), resting-state functional connectivity (RSFC), feature selection, and model validation are provided in Data S2. The flowchart of this study is shown in Figure 1.

| Differences in clinical characteristics
Clinical information from three groups was displayed in Table 1. No significant difference was observed among the three groups regarding age, gender, education, and MMSE score, while significant difference in HAMD score was detected among three groups. In particular, for the DPD group, the HAMD scores (20.2 ± 4.6) were significantly higher than those for other two groups (the same data from our aforementioned published study were used).

| Feature selection
For the first classification of DPD versus (versus) HC, 19 features including (HAMD, 2 mALFFs, and 16 RSFCs) were retained for binary classification. The 16 RSFCs and corresponding brain regions using HOA template were presented in Table 2 The other two mALFF features were located at the left precentral F I G U R E 1 Flowchart of the study. We extracted the 6,557 metrics after the rs-fMRI images preprocessed. Then, Lasso regression was carried out to reduce the number of features. Last, Lasso prediction, random forest, and SVM were used to differentiate between different categories of subjects gyrus and the left planum polare. In Table 3, we reported the statistical characteristics of these features resulting from the dimension reduction step and illustrated the difference in these selected features between DPD and HC. The decreasing or increasing trend of these features between DPD and HC can also be discovered in Table 3.  Table 5.
For the third classification, DPD versus NDPD, 17 features including (15 RSFCs, HAMD, and 1 mALFF) were kept for binary classification. The 15 RSFCs and the corresponding brain regions using HOA template were presented in Table 6. The most aberrant networks associated with these RSFCs included TA B L E 1 Clinical and demographic data evaluation of DPD, NDPD, and HC  Figure 4). The remaining mALFF feature belonged to the region of left subcallosal cortex. In Table 7, we also listed the mean, standard deviation, and p value of these 17 radiomic features.

| Model fitting
After the screening process, for all three classifications, there were no more than 34 features left, and the ultrahigh dimensional situation was no longer present. Most of the commonly used

| Model validation
Although all the methods have achieved superior performance in the training set, the predictive result in the testing set is what really matters. We therefore tested the validity of Lasso, SVM, and random forest by assessing their classified performance in the testing set.
The area under the curve (AUC), accuracy, true positive rate (TPR), and true negative rate (TNR) were measured ( to determine whether these subjects should be classified as DPD.
To further evaluate the robustness of all three methods, the ROC curves were also plotted by varying thresholding values in Figure 5.
From Table 8 and Figure 5, we can tell that Lasso prediction performed better than random forest and SVM for differentiating NDPD from HC. Random forest outperformed Lasso and SVM for discriminating DPD from NDPD in terms of the overall accuracy.
SVM yielded a higher prediction accuracy compared with Lasso and random forest for distinguishing HC from DPD.

| D ISCUSS I ON
By conducting the radiomic analysis, our study presented a comprehensive framework for discovering predictive biomarkers of DPD and for classifying HC, DPD, and NDPD subjects using wholebrain rs-fMRI metrics including ReHo, mALFF, VHMC, and RSFC. In with an accuracy of 100% using a multimodal algorithm (Cherubini et al., 2014). Consistent with existing methods, the present study also found SVM with a multipredictor model was able to fully discriminate DPD from HC. with CNN features enhanced the diagnostic accuracy of PD (Ortiz et al., 2019). In our study, we also discovered discriminative RSFC features in NDPD and DPD, which supported the validity of the radiomic approach. With the emergence of data-driven approaches, radiomics have been shown to be trustworthy and practically useful to aid PD diagnosis and to reach precision medicine.

TA B L E 5
The mean, standard deviation (SD), and p value for all 34 selected features in the training sets for the group of NDPD versus HC  (Hu, Song, Li, et al., 2015), which facilitated the advancement of more detailed and integrative neural models of DPD Wei et al., 2017).
In our study, aberrant functional connectivity and activity of the emotion network and motor network were also identified in DPD patients. Abnormal directional connectivity between motor network and emotion network in DPD has been described in the existing literature. Compared with HC, DPD patients displayed significant gray matter volume abnormality in some limbic and subcortical regions in addition to the unique alterations of directional connectivity from the different brain regions, which may provide differential biomarkers for distinguishing DPD from HC and NDPD (Liang et al., 2016).
Rs-fMRI studies in depression have shown that antidepressant treatment could affect cortical connectivity. The corticolimbic network and amygdala play an important role in the development of DPD, even antidepressant effects also associated with the abnormal hypoconnectivity in DPD (Morgan et al., 2018). Compared to NDPD and HC, DPD showed abnormal functional connectivity in the left amygdala, right amygdala, and the bilateral mediodorsal thalamus.
The disturbed connectivity between limbic regions and corticolimbic networks in DPD patients may reflect impaired limbic areas in mood dysregulation of emotion-related regulatory effect (Hu, Song, Li, et al., 2015).
The group of DPD displayed altered spontaneous brain activity in the frontal, temporal, and limbic regions in our study. Compared with NDPD, DPD exhibited significantly increased regional activity in the superior temporal gyrus, middle temporal gyrus, inferior temporal gyrus, and frontal medial cortex. Decreased RSFC values were detected between superior frontal gyrus and right hippocampus.
These findings confirmed alteration and disruption of the regional brain activity in mood regulation network in the DPD group (Sheng et al., 2014).
Salience network (SN) includes brain regions whose cortical hubs are the anterior cingulate and ventral anterior insular cortices. This network coactivates in response to various experimental tasks and conditions, suggesting a domain-general function (Seeley, 2019). A rs-fMRI study including 17 DPD patients, 17 ND PD patients, and 17 HC subjects found that damaged insula networks between the SN and ECN in PD might lead to DPD . As one of the critical nodes in the STM, the left supramarginal gyrus involved in keeping an abstract representation from the serial order information, and independently from all the content, which instead is stored separately (Guidali et al., 2019). In our study, when DPD was compared with NDPD and HC, disturbed STM regional brain network was identified and might demonstrate the attention deficit.
Our findings also selected several mALFF and F I G U R E 5 ROC curves displaying the predictive performance of Lasso prediction, SVM, and random forest in the testing sets for three classifications. (a) DPD versus HC; (b) NDPD versus HC; (c) DPD versus NDPD baseline brain activity in the dorsolateral prefrontal cortex, the rostral anterior cingulated cortex, and the ventromedial prefrontal cortex that were positively correlated with the HAMD score. The results of abnormal ALFF values in these brain regions implied that the prefrontal-limbic network might be associated with abnormal activities in PD patients with depression (Wen et al., 2013).
In our study, disturbed VMHC was found in the right temporal fusiform cortex, posterior division when comparing NDPD to HC.
Indeed, the impaired functional connectivity within the homotopic brain regions of PD extended previous studies that the disconnection of corticostriatal circuit provided new evidence of disturbed interhemispheric connections in PD (Luo et al., 2015). A VMHC study using the seed-based method discovered decreased VMHC values in the bilateral paracentral lobule and medial frontal gyrus in DPD compared with NDPD (Liao et al.,2020). A structural brain network study showed the global efficiency and characteristic path length were impaired in DPD, which indicated the topological property can be used as a potential objective neuroimaging index for early diagnosis of DPD (Gou et al.,2018).
However, our approach failed to extract any significant VMHC features when comparing DPD and NDPD. This may due to the smaller sample size of the DPD group. Though we could perform data augmentation to increase the sample size, the model fitting results might be compromised by the correlated structures resulting from data augmentation. Hence, for future studies, we intend to include more subjects to diminish the threats caused by the high dimensionality and to further confirm our neurological findings. In addition, the values of radiomic features before and after antidepressant treatment will also be evaluated in the future.
In conclusion, the machine learning-based radiomic approach proposed in this study showed that high-order radiomic features that quantify the functional connectivity and activity of the brain can be used for the diagnosis of DPD and NDPD with high accuracy.

ACK N OWLED G EM ENTS
This research was supported in part by Nanjing Health Young Talent Project (No: QRX 11115) and the Simons Foundation (No: 635213).

CO N FLI C T O F I NTE R E S T
The authors have no conflict of interest to declare.

AUTH O R CO NTR I B UTI O N S
QH and WL conceived and designed the study. QH, XZ, XC, JZ, SZ, CX, and WL performed the experiments. XZ, XC, and QH wrote the manuscript. XC and QH reviewed and edited the manuscript. All authors read and approved the manuscript.

PEER R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1002/brb3.2103.
[Correction added on March 20, 2021, after first online publication: Peer review history statement has been added.]

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are available from the corresponding author upon reasonable request. The data are not publicly available due to privacy or ethical restrictions.