A Knowledge-Driven Framework Discovers Brain ACtivation-Transition-Spectrum (ACTS) Features for Parkinson’s Disease

Dopaminergic treatment has proved effective to Parkinson’s disease (PD), but the conventional treatment assessment is human-administered and prone to intra- and inter-assessor variability. In this paper, we propose a knowledge-driven framework and discover that brain ACtivation-Transition-Spectrum (ACTS) features achieve effective quantified assessments of dopaminergic therapy in PD. Firstly, brain activities of fifty-one PD patients during clinical walking tests under the OFF and ON states (without and with dopaminergic medication) were measured with functional near-infrared spectroscopy (fNIRS). Then, brain ACTS features were formulated based on the medication-induced variations in temporal features of brain regional activation, transition features of brain hemodynamic states, and graph spectrum of brain functional connectivity. Afterwards, a prior selection algorithm was constructed based on recursive feature elimination and graph spectrum analysis for the selection of principal discriminative features. Further, linear discriminant analysis was conducted to predict the treatment-induced improvements. The results demonstrated that the proposed method decreased the misclassification probability from 42% to 16% in the evaluations of dopaminergic treatment and outperformed existing fNIRS-based methods. Therefore, the proposed method promises a quantified and objective approach for dopaminergic treatment assessment, and our brain ACTS features may serve as promising functional biomarkers for treatment evaluation.


I. INTRODUCTION
P ARKINSON'S disease (PD) results from the progressive deterioration of nigrostriatal dopaminergic neurons [1], [2] and impairs motor functions, commonly leading to gait disorders and significantly affecting the quality of life [3], [4].Dopaminergic therapy is effective for PD and has been the predominant therapeutic approach to manage PD symptoms [5], [6].The assessment of dopaminergic treatment is crucial to ensure timely adjustment of dopaminergic strategy, promote precision medicine, and facilitate a deeper understanding of the underlying functional mechanisms [7], [8].
Previous studies have tried to quantify the impact of dopaminergic treatment in PD patients with clinical rating scales [9].However, clinical ratings are discrete assessments and susceptible to inter-rater and intra-rater variances.From a neuroscience perspective, dopaminergic therapy alleviates the symptoms of patients by replenishing the depleted striatum with dopamine [10], [11].The dopaminergic modulation of the brain leads to behavioral improvements of patients.Therefore, incorporating the underlying neurological responses in dopaminergic treatment assessment is a logical step toward providing reliable and objective metrics, which may contribute to the quantified evaluation of dopaminergic therapy and greatly advance our understanding of dopaminergic modulation in PD patients.
But to date, a limited number of fNIRS studies have been reported for dopaminergic treatment evaluation in PD.Orcioli-Silva et al. assessed dopaminergic treatment with the cortical activation in prefrontal cortex during walking [19].Vitorio et al. analyzed dopaminergic treatment with changes in prefrontal cortical activity during walking [20].Orcioli-Silva et al. evaluated the efficacy of dopaminergic therapy based on cortical activity during walking [21].These exploratory studies have quantified the impact of dopaminergic therapy in PD patients with cortical functional activations.However, the transitions of brain states and graph spectrums of brain networks have not been well investigated.Brain state transition supports complex motor and cognitive activities and may be impaired by PD [22], [23], [24], [25].Graph spectrums characterize the variability of brain networks and are crucial in the analysis of neurological diseases [26], [27].Here, we presented brain activation-transition-spectrum (ACTS) features to achieve quantified evaluations of dopaminergic treatment in brain regional activation, brain hemodynamic states, and brain functional connectivity.The results showed that brain regional activation, brain hemodynamic states, and brain functional connectivity comprehensively signified dopaminergic modulation at three brain functional levels, and our brain ACTS features significantly decreased the misclassification probability for the evaluations of dopaminergic treatment.Therefore, the combination of task fNIRS measurement and brain ACTS features promises a quantitative and objective approach for dopaminergic treatment evaluation.The main contributions of this study are as follows: 1) We established a knowledge-driven framework to achieve effective quantified assessments of dopaminergic therapy in PD.
2) We proposed brain ACtivation-Transition-Spectrum (ACTS) features to signify dopaminergic modulation in brain regional activation, brain hemodynamic states, and brain functional connectivity.
3) A prior selection algorithm was constructed to select principal discriminative features based on recursive feature elimination and graph spectrum analysis.

II. MATERIALS AND METHODS
This section first presented the clinical characteristics of PD patients during the experiments.Then, the experimental setup was detailed for the description of collecting brain signals from PD patients before and after dopaminergic treatment.Afterwards, a knowledge-driven framework was introduced to detail the process of achieving effective quantified assessments of dopaminergic therapy in PD.

A. Patients
This study was approved by the Ethics Committee of Tianjin Huanhu Hospital, Tianjin, China, and registered in Chinese Clinical Trial Registry (ChiCTR1900022655).Moreover, this study was conducted in accordance with the Declaration of Helsinki.Each participant provided their written consent prior to the experiment.Since gait performance is the fundamental concern of patients and doctors, the clinical walking test was selected as the evaluated motor task.Patients were recruited if they were clinically diagnosed as PD and capable of performing unaided walking.Patients were excluded if they had a history of brain trauma, severe blood pressure fluctuations between OFF and ON periods of dopaminergic treatment, or were unable to follow the instructions of doctors.Fifty PD patients were recruited (Table I), and one PD patient was excluded due to poor signal quality.The MDS-UPDRS subscore of gait decreased significantly ( p < 0.05) between the OFF state (Mean: 1.28, SD: 0.61) and ON states (Mean: 0.88, SD: 0.68).

B. Experimental Setup
PD patients performed the experiments under the OFF and ON states, as shown in Figure 1.Specifically, the experimental procedure consisted of three phases: (1) Patients performed the walking tests for three times under the OFF state.(2) Patients took dopaminergic medication and waited for about 1 hour to ensure that the medication worked.(3) Patients performed the walking tests again for three times under the ON state.Each walking test involved 30 s standing, 35 s walking, 10 s standing, and 2 minutes rest.During the experiments, the brain signals of patients were recorded by a portable and wireless Nirsmart fNIRS system (Danyang Huichuang Medical Equipment Co., Ltd China) with an 11 Hz sampling rate.Specifically, twenty-six optodes were positioned at the left and right prefrontal cortex (L-PFC and R-PFC), premotor cortex (L-PMC and R-PMC), and primary somatosensory cortex (L-S1 and R-S1), as shown in Figure 2. PFC, PMC, and  S1 play critical roles in motor task planning and controlling during walking [28], [29], [30] and were analyzed in this study.

C. Proposed Knowledge-Driven Framework
Figure 3 presents the proposed knowledge-driven framework to discover brain ACtivation-Transition-Spectrum (AC-TS) features for PD patients.The proposed framework was applied to classify the patients into two groups: y 0 and y 1 : where y O F F and y O N represented the MDS-UPDRS sub-score of gait under the OFF and ON states.Group y 1 consisted of 20 PD patients who showed significant gait improvements under dopaminergic treatment, while group y 0 comprised 30 PD patients without significant gait improvements.The proposed framework includes the following steps: (1) Preprocessing was performed to denoise the collected fNIRS brain signals; (2) Brain ACTS features were constructed to signify dopaminergic modulation in brain regional activation, brain hemodynamic states, and brain functional connectivity; (3) A prior selection algorithm was proposed to select principal discriminative features based on recursive feature elimination and graph spectrum analysis; (4) The classification performances were evaluated; (5) The evaluation of regional brain features was performed.
1) Preprocessing of fNIRS Signals: The recorded fNIRS signals were preprocessed with the following steps: (1) A sliding window and cubic spline interpolation were used to eliminate motion artifacts [31], [32].(2) A 0.01-0.2Hz bandpass filter was used for the removal of physiological noises [33], [34].(3) The modified Beer-Lambert law was utilized to convert the signals into the relative concentration changes of oxyhemoglobin (△H bO) and deoxyhemoglobin (△H b R) [35].(4) Principal component analysis was employed to filter out scalp blood flow interference [36], [37].The preprocessed signals in the walking period were referred to the 5-second baseline prior to walking and extracted for the following analysis.
2) Construction of Brain ACTS Features: a) Temporal features of brain regional activation: Five temporal features including mean, variance, peak, kurtosis, and skewness were extracted from △H bO to quantify the time-varying brain activation properties.
b) Transition features of brain hemodynamic states: Brain state transition features including transition acceleration (TAcc), transition angle (TAng), and transition strength (TStr) were extracted to measure the capacity of information transmission.TAcc, TAng, and TStr are defined as: Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
Fig. 3.The proposed knowledge-driven framework to discover brain ACtivation-Transition-Spectrum (ACTS) features to assess dopaminergic treatment, including (1) preprocessing of fNIRS data, (2) the construction of brain ACTS features in brain regional activation, brain hemodynamic states, and brain functional connectivity, (3) feature selection with the proposed prior selection algorithm based on recursive feature elimination and graph spectrum analysis, (4) classification performance of the proposed method and ( 5) the evaluation of regional brain features.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
where N is the point number of fNIRS channel signals, △H bO i and △H b R i are the values of △H bO and △H b R at time stamp t i .d i, j is the distance between (△H b R i , △H bO i , t i ) and (△H b R j , △H bO j , t j ) and calculated as: L i is the norm of (△H b R i , △H bO i , t i ) and defined as: f t i ,t j is the transition factor and constructed as: s t i indicates the brain hemodynamic state at time stamp t i .Three brain hemodynamic states were defined as: • Expansion State: △H bO > 0 and △H b R < 0, hyperactivation.
• Intermediate State: (△H bO < 0 and △H b R < 0) or (△H bO > 0 and △H b R > 0), fluctuating around the baseline activation.The hemodynamic states were formulated based on the correlation analysis of the neural activity and hemoglobin responses within the balloon model [38], [39].x and y were set to be 1.5 and 3, indicating concentrating on the impact of inner-state for transition features.
c) Graph spectrums of brain functional connectivity: Graph spectrums were proposed to quantify the spectral variations of brain functional networks.Specifically, brain connectivity network is constructed with phase lag index (PLI).PLI measures the connection strength between channel signals.The Hilbert transformation xi (t) of each fNIRS channel x i at time t is calculated as: where C pr is the Cauchy principle value.Given two fNIRS channel signals x i and x j , the phase difference between two channel signals x i and x j at time t is computed as: where ϕ i (t) and ϕ j (t) indicate the instantaneous phases of x i (t) and x j (t).The PLI score between x i and x j is defined as: where N is the signal length.sign(•) is the signum function.P L I i, j ranges from 0 to 1, where 0 and 1 indicate the weakest and strongest connection strengths, respectively.P L I i, j is used to construct the connectivity matrix A: where M represents the channel number.The connectivity matrix was derived from △H bO in this study since △H bO was more sensitive to locomotion-related activities than △H b R [40], [41].
A brain graph is constructed as the pair G = (V, A), where V = {v 1 , v 2 , . . ., v M } is the set of M vertices.The degree matrix D ∈ R M×M is a diagonal matrix with its i-th diagonal element D ii = M j=1 P L I i, j .The Laplacian matrix L ∈ R M×M of G is defined as L = D − A. Thus, the components of L are explicitly given by L i, j = −P L I i, j and L i,i = M j=1 P L I i, j .L is real, symmetric, diagonal dominant, and positive semidefinite.It can be decomposed as: where U = [u 1 , u 2 , . . ., u M ] is the eigenvector matrix, u i indicates the i-th eigenvector.U H is the Hermitian of U . is the diagonal eigenvalue matrix with = diag(λ 1 , λ 2 , . . ., λ M ) ∈ R M×M , λ i indicates the eigenvalue of the i-th eigenvector u i .The eigenvalues are arranged in ascending order, i.e., Eigenvalues were extracted as graph spectrum features.The channel number M was set as 30, resulting in 30 graph spectrum features extracted in this study.
3) Dopaminergic State Crossing: Brain activation, brain state transition, and graph spectrum features were extracted under the OFF and ON states.However, the relationship between extracted features under the OFF and ON states has not been well explored.Thus, dopaminergic state crossing was defined to build up the relationship of extracted features under the OFF and ON states: where M F i t , M F i b , M F j e are the treatment-induced variations of extracted features in brain regional activation, brain hemodynamic states, and brain functional connectivity.T f ea i and B f ea i indicate the temporal features and transition features in the i-th channel (i ranges from 1 to 30).E f ea j is the j-th graph spectrum feature ( j ranges from 1 to 30).
To select the most discriminative features and avoid overfitting the classification model, a prior selection algorithm was proposed, as shown in Algorithm 1.For M F t and M F b , n 1 features were selected with recursive feature elimination [42].For Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
M F e , M F 2 e was selected since the second smallest eigenvalue (called algebraic connectivity) was crucial to the analysis of brain functional network [43].Moreover, n 2 features were chosen from the remaining M F e features with recursive feature elimination.The recursive feature elimination was with a linear support vector machine estimator.n 1 and n 2 were set as 11 and 3. Fifteen features were selected in this study.The feature selection procedure was repeated independently for each run of leave-one-out cross-validation.

Algorithm 1 Prior Selection
Input: n 1 : the number of selected features for M F t and M F b n 2 : the number of selected features for M F e Output: S F: selected features 1. Initialize S F as an empty set.2. Rank M F t and M F b with recursive feature elimination.3. Add n 1 features S F n 1 with the highest rank to S F.

Add M F 2
e to S F based on the analysis of algebraic connectivity.5. Rank M F j e ( j = 1, 3, 4, . . ., 30) with recursive feature elimination.6. Add n 2 features S F n 2 with the highest rank from M F e to S F.

5) Classification and Statistical Analysis:
Linear discriminant analysis (LDA) was an effective classifier for fNIRS data analysis [44] and was applied to differentiate patients with different gait improvements on the basis of brain ACTS features.LDA finds the optimal hyper-planes in which the projected features of two classes (for example, patients without and with significant gait improvements) have the minimum intra-class variance and highest separation between the classes.The predicted LDA scores were objectively compared for each class, and the performance was assessed with misclassification probability (MCP).MCP was defined as the probability that patient group y 0 is misclassified into group y 1 (MCP 01 ) and patient group y 1 is misclassified into group y 1 (MCP 10 ).Moreover, three commonly used metrics: accuracy, sensitivity, and specificity were used to evaluate the classification performances.The five-fold cross validation results were used to demonstrate the effectiveness of our method.Min-max scalers were learned with the training data and applied for the test data.Two-sample t-tests were calculated for features in each brain region between patient groups.The FDR correction with the Benjamini-Hochberg method was used for multiple testing in this study [45].A significance level of 5% was utilized to reject the null hypothesis.

A. Classification Performances
Figure 4 presented the classification performances for conventional fNIRS hemodynamic metrics and our brain ACTS features.The results indicated that the classification based on conventional metrics was relatively poor.On the other hand, the combination of our brain ACTS features in brain regional activation (Activation), brain hemodynamic states (Transition), and brain functional connectivity (Spectrum) resulted in the best performances.The results of our brain ACTS features in PFC, PMC, and S1 were provided in Figure 5.The results indicated that the combination of our brain ACTS features in PFC, PMC, and S1 led to the best performances (the highest accuracy, sensitivity, and specificity as well as the lowest MCP).

B. Comparison Results
The proposed method with brain ACTS features was compared with five effective methods for fNIRS analysis of PD patients.In method 1, Maidan et al. differentiated PD patients and healthy controls with fNIRS-based oxyhemoglobin variability [46].In method 2, Schejter-Margalit et al. discriminated between PD patients and healthy controls with fNIRS-based slope features [47].In method 3, Abtahi et al. distinguish PD patients with fNIRS mean concentration changes of oxyhemoglobin and support vector machine [48].In method 4, Wang et al. proposed a transformer-based deep neural network (named fNIRS-T) for fNIRS classification [49].In method 5, Wang et al. integrated the domain knowledge of delayed hemodynamic responses and developed a concise network model named fNIRSNet for fNIRS classification [50].Our method with brain ACTS features outperformed these methods in terms of accuracy, sensitivity, specificity, and MCP, as shown in Figure 6.

C. Validation Results
An external validation experiment was performed to further demonstrate the effectiveness of the proposed method.Specifically, a newly unseen dataset was constructed by collecting fNIRS brain signals of 11 PD patients according to the same criteria in the "A.Patients" and "B.Experimental Setup" of Section "II.MATERIALS AND METHODS".Four patients were with significant gait improvements while seven patients were without significant gait improvements.The clinical characteristics of the newly collected PD patients were shown in Table II.The previously constructed dataset with 50 PD patients was used as the training dataset, and the newly unseen dataset with 11 PD patients was utilized as the validation dataset.Experimental results showed that our method with brain ACTS features outperformed other effective methods in the newly unseen dataset in terms of accuracy, sensitivity, specificity, and MCP, as shown in Figure 7.

D. Regional Brain Features With and Without Significant Gait Improvements
To further ascertain that our brain ACTS features can differentiate different gait improvements, ACTS features over the PFC, PMC, and S1 cortical regions were quantified.Graph spectrums of brain functional connectivity were global indicators calculated based on the whole brain networks and not analyzed in this experiment.Significant differences were observed in PMC and S1 between patient groups with different gait improvements.Specifically, patients with significant gait improvements have significantly higher variance ( p < 0.05 FDR) in PMC and S1, as well as transition strength in S1, and significantly lower skewness in PMC compared with patients without significant gait improvements (as shown in Figure 8).Table III presented the statistical difference of ACTS features in PFC, PMC, and S1 between patient groups with different Fig. 4. Comparison of conventional fNIRS hemodynamic metrics and our brain ACTS features in brain regional activation (Activation), brain hemodynamic states (Transition), and brain functional connectivity (Spectrum).The combination of features in brain regional activation, brain hemodynamic states, and brain functional connectivity resulted in the best performances (the highest accuracy, sensitivity, and specificity as well as the lowest MCP).gait improvements in brain regional activation and brain hemodynamic states.

IV. DISCUSSION
Starting from the fact that PD is a neurological disease affecting motor performances, an objective method was proposed for dopaminergic treatment assessment with task fNIRS measurements and brain ACTS features.To the best of our knowledge, this is the first fNIRS-based study for the assessment of dopaminergic treatment on the basis of brain regional activation, brain hemodynamic states, and brain functional connectivity.Our method was developed to evaluate the Fig. 6.Comparison results of our method with brain ACTS features and five effective methods: method 1 [46], method 2 [47], method 3 [48], method 4 [49], and method 5 [50].Our method has the best performances (the highest accuracy, sensitivity, and specificity as well as the lowest MCP).Fig. 7. Validation results in the newly unseen dataset of our method with brain ACTS features and five effective methods: method 1 [46], method 2 [47], method 3 [48], method 4 [49], and method 5 [50].Our method has the best performances (the highest accuracy, sensitivity, and specificity as well as the lowest MCP).
efficacy of dopaminergic treatment, but also enabled objective analysis of rehabilitation and deep brain stimulation therapy by providing definitive evidence at the neural functional level.
This paper proposed brain ACTS features to quantify the impact of dopaminergic treatment in brain regional activation, brain hemodynamic states, and brain functional connectivity.In brain regional activation, previous studies approached the analysis of dopaminergic therapy with fNIRS mean hemodynamic changes [19], [20], [21].Besides mean hemodynamic activation, Maidan et al. demonstrated that the variability of  oxyhemoglobin measures was a promising indicator in PD diagnosis [46].Moreover, the peak, kurtosis, and skewness features were effective in fNIRS-based brain-computer interfaces and may promise valuable analysis in PD [44].Thus, brain regional activation was investigated with fNIRS-based mean, variance, peak, kurtosis, and skewness indicators.The human brain is a dynamic system with complex transitions supporting motor and cognitive activities [22], [23].PD affects the functional dynamics of the brain and may deteriorate the changes of brain states [24], [51].Thus, brain state transition features were constructed according to brain hemodynamic states and used for the quantification of dopaminergic modulation.Graph spectrums were the fundamental property of brain functional networks in the spectral representation.The spectral brain networks permitted the extraction of graph pieces associated with different modes of variations, which were crucial in the investigation of neurological disorders [26], [27].Thus, graph spectrums were quantified and used in this study.Our brain ACTS features in brain regional activation, brain hemodynamic states, and brain functional connectivity were instructively complementary and had better performances when compared to other features.The proposed method had better performances than five effective fNIRS-based methods [46], [47], [48], [49], [50].
The methods [46], [47], [48] focused on the analysis of brain temporal characteristics.However, previous studies showed that brain hemodynamic states and brain spectral features were also crucial in the exploration of neurological disorders [24], [26], [27], [51].Neglecting these properties may result in performance degradation and incomplete understanding of the brain's response to neurological conditions.Thus, the proposed method constructed brain ACTS features for the evaluation of dopaminergic treatment by considering brain temporal characteristics, brain hemodynamic states, and brain spectral features.In [49] and [50], deep neural networks were developed for fNIRS classification.These methods are solely data-driven and may not fully capture the PD-related features.The proposed method was knowledge-driven and discovered effective brain ACTS features to signify neural degeneration of PD.Compared with [49], [50], the proposed method had better classification performances.
This paper proposed a prior selection algorithm to choose the most discriminative features.The proposed algorithm integrated domain-specific knowledge by selecting M F 2 e based on the second smallest eigenvalue, which is critical in brain network analysis [43].Moreover, instead of applying recursive feature elimination (RFE) uniformly, RFE was sequentially applied for feature sets in brain temporal and spectral domains, allowing for more precise feature selection specific to each feature set's characteristics.By combining a domain-specific selection criterion with RFE, a hybrid method was proposed to leverage the strengths of both techniques, enhancing the overall discriminative power of the selected features.
Gait performance is a fundamental concern for patients and doctors.Thus, the clinical walking test was used as the motor task for dopaminergic treatment assessment.During walking, three gait-related brain regions, PFC, PMC, and S1, were analyzed.PFC is associated with the execution and adjustment of complex behavior [28], [52].PMC plays a pivotal role in motor functions and movement planning [29], [53].S1 is crucial in receiving sensory information and controlling rhythmic movements [30], [54].The results showed that the combination of brain ACTS features in PFC, PMC, and S1 resulted in better assessment performances compared with them only in PFC, PMC, or S1.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
Besides fNIRS, EEG can also measure brain activities for the investigation of PD, but generally for non-motor functions [55], [56].The fundamental challenges of EEG in motor tasks consist of the need to quickly set up, remove motion artifacts, and locate the brain sources.fMRI and PET could not permit task-state measurements when patients were trying to perform a motor task.fNIRS was used in this study for its portability in task-related measurement, good tolerance to motion noises, flexibility for experiments, etc [17].
The reasons for using linear support vector machine (SVM) during recursive feature elimination (RFE) were that linear SVM was effective in high-dimensional spaces, less prone to overfitting during the feature selection process, and could provide a clear margin of separation between classes, which was beneficial when identifying the most discriminative features.The reasons for using linear discriminant analysis (LDA) as the final classifier were that LDA could provide a robust decision boundary that maximizes the separation between classes and was effective in fNIRS data classification [44].Using linear SVM for RFE, we leverage its strengths in feature selection to identify the most relevant features.Subsequently, applying LDA as the final classifier allows us to take advantage of its power in fNIRS data classification.This combination could optimize the overall model performance by ensuring robust feature selection followed by effective classification.
By solving the task of distinguishing patients with different medication-induced gait improvements at the brain functional level, we could achieve finer assessment of medication effects and explore the compensation neural mechanism of dopaminergic medication, which contributes to achieving effective adjustment of medication treatment plans [19], [20], [21].Moreover, crucial cortical regions could be identified by the proposed method and targeted with further interventions such as transcranial magnetic stimulation and direct current stimulation, which may help to alleviate the cortical burden of walking in PD [57].In addition to dopaminergic treatment assessment, predicting medication effects from pre-medication fNIRS signals holds great promise for enhancing personalized medicine.By quantifying the likely response of a patient to a particular medication with fNIRS signals before taking medication, doctors can make more effective decisions for medication treatment, potentially reducing trial-and-error prescribing and optimizing therapeutic effectiveness from the outset.In the future, we will explore whether the change in gait performance can be predicted from fNIRS signals before taking medication.
We hope this work can encourage more study, and more clinical evidence will promisingly enable quantified optimization of dopaminergic treatment for PD patients.In the future, we will try to employ the proposed method on more datasets.

V. CONCLUSION
This paper proposed a knowledge-driven framework to discover brain ACtivation-Transition-Spectrum (ACTS) features for the evaluation of dopaminergic treatment.Brain ACTS features were formulated to evaluate dopaminergic treatment in PD patients from brain regional activation, brain hemodynamic states, and brain functional connectivity.Experimental results demonstrated that brain regional activation, brain hemodynamic states, and brain functional connectivity were informatively complementary and comprehensively signified dopaminergic modulation at three brain functional levels.The proposed method significantly decreased the misclassification probability and outperformed other competitive methods for dopaminergic treatment assessment.Prospectively, the proposed method could be employed to other brain disorders.

Fig. 1 .
Fig. 1.Experimental procedures.PD patients performed clinical walking tests for 3 times under the OFF and ON states.The walking test involved 30 s standing, 35 s walking, 10 s standing, and 2 minutes rest.Patients took dopaminergic medication and waited for about 1 hour to ensure that the medication worked.

Fig. 2 .
Fig. 2. Optode arrangement.The optodes including 14 sources and 12 detectors were positioned at the left and right prefrontal cortex (L-PFC and R-PFC), premotor cortex (L-PMC and R-PMC), and primary somatosensory cortex (L-S1 and R-S1), generating 30 channels.The i-th channel is denoted as Ci.

Fig. 5 .
Fig. 5. Comparison of our brain ACTS features in PFC, PMC, and S1.The combination of ACTS features in PFC, PMC, and S1 led to the best performances (the highest accuracy, sensitivity, and specificity as well as the lowest MCP).

Fig. 8 .
Fig. 8. Significant changes of regional brain ACTS features between patients with and without significant gait improvements.(a) Changes of variance in brain regional activation.(b) Changes of skewness in brain regional activation.(c) Changes of transition strength in brain hemodynamic states.n.s.: not significant.*: p<0.05 FDR.