Deep learning reveals personalized spatial spectral abnormalities of high delta and low alpha bands in EEG of patients with early Parkinson's disease

Parkinson's disease (PD) is a progressive neurodegenerative disease caused by the loss of dopaminergic neurons in the substantia nigra [1]. Its typical clinical manifestations include motor symptoms such as resting tremor, bradykinesia, muscular stiffness [2, 3]. In addition to these motor symptoms, there are various non-motor impairments, such as cognitive dysfunction, deficits in sensory processing, sleep disorders, autonomic nervous system disorders, and visual processing dysfunction, even in the early stage of PD [4–6]. According to epidemiological investigation and analysis, the number of patients with PD is expected to reach 8.7 million globally by 2030 [7]. The incidence of PD is on the rise around the world, and its irreversible course of disease poses a heavy burden on patients' families, social economic and medical security [8]. Early treatment of PD can effectively slow down the rate of disease deterioration [8]. Therefore, early diagnosis of PD is crucial. At present, due to the lack of definite clinical indicators and unclear typical clinical symptoms of early PD, the diagnosis of early PD has always been a difficult problem in clinical medicine [9]. The diagnostic process is demanding and time-consuming for clinical professionals [9]. Therefore, the study of biomarkers that satisfy individual differences of early PD has very important social and medical value.

Since the 1950s, studies have attempted to discover a correlation between PD and electroencephalography (EEG), which was increasingly recognized as a measurement used to systematically examine cortical activity in human brain [10]. Numerous studies have shown that quantifying characteristics of EEG rhythms could provide important evaluation indexes for various neuropsychiatric disorders, such as Alzheimer's disease, schizophrenia, epilepsy, major depressive disorder and so on [11–14]. The most widely reported EEG rhythm deviations in PD are an increase in slow wave power (theta (1–4 Hz) or/and low alpha (8–10 Hz) power), or a decrease in fast wave power (beta (13–30 Hz) or/and gamma (30–45 Hz) power), or both, but these studies [15–17] have not always reported consistent results. So far, there is still no reliable indicator to evaluate the spectral characteristics of early PD. There may be three reasons for this limitation. First, the division of quantitative EEG frequency bands depends on conventional rules, that is, delta (1–4 Hz), theta (4–8 Hz), low alpha (8–10 Hz), high alpha (10–13 Hz), beta (13–30 Hz) and gamma (30–45 Hz) bands [17, 18], while the frequency bands with characteristic rhythm in early PD is still not clear. Second, the methods to explore the rhythm characteristics of the whole brain average or single channel signals ignore the spatial distribution information of EEG. Third, there is still a lack of personalized analysis methods to explore the rhythm characteristics including individual differences in early PD patients. Therefore, to determine the personalized rhythm characteristics of early PD, it is necessary to integrate all the characteristic information expressed by EEG, including temporal, spectral and spatial and individual information, so as to improve the accuracy of the auxiliary diagnostic indicators of early PD.

In recent years, it has been discovered that deep neural networks have strong computing capabilities, which can mine the inherent information in data and can be directly applied to classification [19–21]. Because of this, a growing number of researchers attempt to use deep learning algorithms to automatically analyze EEG data for classification tasks or feature extraction. Schirrmeister et al [22] proposed a comparison between convolutional neural network (CNN) with different architectures and filter bank common spatial patterns (FBCSP) which is a widely used baseline method, and the results revealed that CNN could achieve a better performance than FBCSP (mean decoding accuracies FBCSP 82.1%, CNN 84.0%). The CNN algorithm architecture was also applied to distinguish pathological and healthy EEG signals, which achieved much better accuracy than the best reported results of this data set (85% vs. 79%) [23]. It is worth noting that both of these studies calculate the variation of power spectrum in different frequency bands, and visualize the learned features to explain the rationality of decisions. The visualization technology mentioned here, such as activation maps [24, 25], saliency maps [26], image occlusion [25, 27] and noise iteration [28, 29], is an effective solution to the lack of interpretability of deep learning models. In addition, CNN is a powerful computing tool that can extract the implicit features of data objects containing spatial information. This suggests that CNN is very promising to automatically find characteristic frequency bands and extract rhythm features including temporal and spatial attributes of EEG for early PD. Moreover, CNN combined with visualization technology can express personalized feature information and the feature can be interpretable, which provides technical support for us to study personalized auxiliary diagnostic indicators.

In this study, in order to incorporate the spatial distribution of EEG signals, structured power spectral density with spatio-temporal frequency properties were used as the inputs to train the CNN model for distinguishing early PD from healthy controls. We expect that, based on the convincing performance of the model, we can obtain the optimal characteristic range of frequency bands for identifying early PD. Moreover, on the basis of the characteristic range of frequency bands has been confirmed, in order to explore the personalized biomarkers of early PD, we used a visualization technique called gradient-weighted class activation mapping (Grad-CAM) [30] to obtain the contribution parameters of each subject, and then calculated personalized spatial relative power characteristics that could effectively distinguish early PD from healthy control (HC). We expect that this personalized detection index can significantly characterize the nature of early PD, and become a reliable biomarker for clinically assisted diagnosis.

2.1. General information about participants

2.2. EEG recording and processing

All subjects (including early PD patients and HCs) were seated comfortably in a quiet semi-dark room, in a relaxed state, with eyes closed but kept awake. EEG data was collected between 9 a.m. and 11 a.m. Electrical activities of the scalp at 19 Ag/Agcl scalp electrodes (active electrodes, SYMTOP, Beijing, China), and two reference electrodes (A1 and A2) were recorded, and additional channels were also used to monitor electrocardiogram (ECG), electromyogram (EMG) with the electrodes placed on the skin between thumb and index finger on the back of both hands, as well as four electrooculograms (EOGs) for recording the vertical and horizontal movements of the eyes and blinking, for further preprocessing. Nineteen Ag/Agcl scalp electrodes (active electrodes, SYMTOP, Beijing, China) Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T7, T8, P7, P8, Fz, Cz and Pz were placed on the scalp in accordance with international standard 10–20. All electrodes used for collection are kept below 5 k$\Omega$ in impedance. EEG signals were sampled at 500 Hz.

Fast independent components analysis (fast-ICA) algorithm [31] was applied to remove the noise artifacts, including ocular artifacts (from EOG), cardiac artifacts (from ECG) and movement artifacts (from EMG) in the scalp EEG for each subject respectively. Fast-ICA algorithm decomposes EEG signals into several independent components that are statistically independent of each other. Pearson correlation coefficients are calculated between each independent component and the noise artifacts separately. The independent component whose absolute value of correlation coefficient is greater than 0.5 is considered as the component that has strong correlation with a certain artifact signal, and then we zeroed out these components to obtain the de-noised EEG signals with the noise artifacts removed. All of the de-noised EEG data were carefully checked for other artifacts (i.e. body movements, technical artifacts) by an experienced researcher and the epochs containing amplitude > 80 μV were labeled and rejected. Consequently, more than 5 min signals were kept for further analysis for each subject. Then the EEG signals was processed by a 0.5–45 Hz band-pass finite impulse response (FIR) filter, and each channel of EEG recordings was decomposed into two sub-bands: high delta (3.5–4.5 Hz), low alpha (7.5–11 Hz) via the FIR filter (these two sub-bands are obtained according to the Results section).

Since spectral deviations of early PD are mostly reported, the power spectral density (PSD) of each channel was used as the input characteristic. We selected the same number of epochs from each subject for individualized analysis to avoid deviations in the data set. The continuous time series of each subject were cut into 2 s epochs (i.e. window length is 2 s). The first 150 epochs (without overlap) of each participant were selected, and the P-welch function was used to calculate the PSD of each channel within each epoch at 0.5–45 Hz with step size of 0.5 Hz (it forms 90 points of frequency). We defined that the input PSD was represented by the combination of electrode channels and frequency points to form a channel-frequency PSD dataset, which was defined to contain 1000 samples. The EEG signal processing was performed by MATLAB software (MathWorks Inc., Natick MA, United States).

In order to make CNN model realize its optimal feature extraction function, the input in receptive field should be spatially correlated. The scalp location of each channel was mapped to a specific location in a two-dimensional (2D) matrix using a data organization based on azimuthal equidistant projection [32]. In order to improve the maps' resolution, 2D interpolation was employed on these 2D maps and then the spatial-temporal 2D inputs were obtained. Since there were 90 frequency points, the input sample, spatial PSD, was organized into a three-dimensional (3D) matrix as structured PSD dataset (as shown in figures 1(A) and (B)).

Zoom In Zoom Out Reset image size

2.3. CNN architecture and training strategy

2.4. Model visualization and spatial-mapping relative power (SRP)

In order to reveal the basis of CNN model to discriminate early PD, we adopted the Grad-CAM method [30] for the visual interpretation of the model. The Grad-CAM can discriminate the position of the input data without attention, and visualize CNN of any structure without modifying the network structure or retraining [30]. The activation feature graph and the weight of each feature in the last convolutional layer of CNN models can be obtained by using the Grad-CAM. The judgment basis of CNN models can be visualized through the activation feature graph and the weight of each feature, and then 'the key for model to distinguish the input data category' can be found.

In this work, the activation feature maps of each group were composed of 90 frequency points. The activation characteristic maps calculated by Grad-CAM which can distinguish the early PD group and HC group, can help us find the frequency band range and the location of the characteristic brain region within this frequency band with the most significant difference between early PD and HC. The weight of each feature ranged from 0 to 1. Through coordinate correspondence, we calculated the corresponding feature weight values of 19 electrode channels. And then, we obtained the single-channel feature weight value of each subject within the range of each window length in EEG, thus realizing the expression of personalized spatial-mapping feature weight.

Based on the personalized spatial-mapping feature weight, we proposed a personalized characteristic index of frequency domain named SRP. The SRP was obtained by multiplying the relative power of each individual with its corresponding spatial-mapping characteristic weight, as shown in the formula below:

$$\begin{equation*}{\text{SRP}} = \frac{1}{m}\left( {\sum\limits_{i = 1}^n \sum\limits_{j = 1}^m {C_{ij}} \times {\text{R}}{{\text{P}}_{ij}}} \right)\end{equation*}$$

where, n represents the number of electrical channels, and m indicates the number of windows in each subject's EEG signal calculated with a window length of 1000. C_ij is the spatial-mapping feature weight and RP_ij is the relative power of each characteristic frequency band which was obtained by the activation characteristic maps. For each characteristic frequency band, the relative power is defined as the ratio between the sum of PSD within the frequency band and the sum of the PSD in the 0.5–45 Hz frequency band. In order to compare whether dopamine drugs have a regulatory effect on SPR, the values of SRP were calculated for early PD patients with dopamine drugs and those who stopped dopamine drugs for more than 12 h.

2.5. Statistical analysis

We divided the input data for the training models into channel-frequency PSD and structured PSD sets, and the performance of three kinds of classified models (SVM, MLP and CNN) using the test dataset are shown in table 2. On both kinds of input datasets, the MLP and CNN models yielded the higher ACC and AUC than the SVM models, demonstrating a better learning ability of the deep learning networks for this study. The CNN model showed better performance on the structured PSD sets than on the channel-frequency PSD set, while the ACC and AUC of MLP model on both datasets were similar and lower than that of CNN model, and the performance of SVM model is the lowest. It shows that using spatial frequency features as input data to train CNN model can achieve better discrimination effect, which is attributed to that frequency features are endowed with spatial scale to deepen the dimension of this feature and CNN is very suitable for processing spatial data. The validation result of CNN model on the test sets are that during the eight-fold cross-validation period, the proposed model yields ACC of 99.87% ± 0.03% and AUC of 99.87% ± 0.05% (as shown in figure 2), revealing the optimal discrimination ability so far.

Zoom In Zoom Out Reset image size

In order to explore the characteristic basis for CNN model to achieve such excellent classification performance, we used the Grad-CAM algorithm to visualize the last convolutional layer of CNN. The Grad-CAMs were normalized from 0 to 1 and averaged as the personalized characteristics. After projecting the Grad-CAMs back to the 2D maps using the inverse azimuthal equidistant projection, the personalized characteristics can be shown in topographic maps at each frequency point. The topographic maps with personalized characteristics are shown in figure 3. As shown, in the characteristic topographic map of 90 frequency points, there are 11 frequency points with significant distinguishing contribution while the other frequency points hardly play a role in differentiating early PD and HC. It is obvious that there are significant indicators of abnormal brain rhythms in such two different frequency bands at the early PD. Therefore, these two frequency bands are defined as the optimal frequency bands indicating the characteristics of abnormal brain rhythms in early PD. According to the conventional frequency band division rules, these two bands are distributed in delta and alpha frequency bands respectively.

Zoom In Zoom Out Reset image size

We separately extract the feature topographic maps at these 11 frequency points to obtain the optimal frequency bands, that are high-delta (3.5–4.5 Hz) and low-alpha (7.5–11 Hz) frequency bands, as shown in figure 4. In high-delta band (3.5–4.5 Hz), the characteristic region for early PD is right frontal lobe and in low-alpha band (7.5–11 Hz), it is occipital lobe (especially the left part of occipital lobe). It is noting that each frequency point of the two characteristic frequency bands is composed of specific Grad-CAMs, that is, the spatial power spectrum composed of PSD distributions at each characteristic frequency point constitutes the combined features for identifying early PD, and the mean values of specific Grad-CAMs of each electrical channels in the characteristic frequency bands are the spatial-mapping feature weight.

Zoom In Zoom Out Reset image size

Based on the above features and our proposed personalized frequency feature indexes algorithm, the SRP of these two redefined feature frequency bands are calculated as shown in figure 5 and table 3. The early PD patients with dopamine drugs are included in the characteristic analysis to explore the influence of dopamine drugs on the SRP. The SRP values of early PD groups (with drug-off and drug-on) and HC group in high-delta band (3.5–4.5 Hz) are shown in figure 5(A). As shown, the SRP values of early PD patients with dopamine drugs are pretty similar to those of HC (with p = 0.42), while the SRP values of early PD patients without taking drugs are significantly higher than those of patients with dopamine drugs and HC subjects (with p < 0.01). Meanwhile, the SRP values of these groups in low-alpha band (7.5–11 Hz) also have the same phenomenon, as shown in figure 5(B). SRP has been shown to be a personalized characteristic indicator that can completely distinguish early PD patients without medication from both PD patients with medication and HC (as shown in figure 6), suggesting that the personalized relative power of delta and alpha waves combined with spatial frequency distribution characteristics is significantly increased in patients with early PD within the specific frequency bands. Notably, this abnormal increase is significantly restored to healthy levels after medication, indicating that this significant increase in the relative power of specific delta and alpha waves reflects the abnormal brain rhythm characteristics of early PD, and it can be modulated by dopamine treatment.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Figure 7 shows the Spearman correlation between the gross scores of the UPDRS-III and the SRP in the high-delta and low-alpha bands in PD patients without medicine. There was no significant correlation between SRP in any characteristic frequency band and the gross scores of UPDRS-III with p > 0.01 corrected by FDR. This indicates that the increase of SPR in both two characteristic frequency bands can not reflect the deterioration of the overall motor function in early PD. In other words, the increase of slow waves (in the frequency band of 3.5–4.5 Hz and 7.5–11 Hz) is not associated with the deterioration of overall motor function in early PD.

Zoom In Zoom Out Reset image size

We further explored the Spearman correlations between the scores of subtypes (including bradykinesia, rigidity and tremor) of symptoms in UPDRS-III and the SRP in the high-delta and low-alpha bands in PD patients without medicine and figure 8 is the sketch map. There was a significant positive correlation between SRP in the high-delta (7.5–11 Hz) frequency band and the scores of tremor in UPDRS-III in the drug-off PD patients (r= 0.547, p = 0.002 with FDR). However, there were no significant correlation between SRPs and other clinical scales in drug-off PD groups with p > 0.01. The higher the UPDRS-III's score is, the worse the motion function is [42]. It can be seen that, within a certain permissible range, the more serious the tremor symptom of early PD is, the larger value of the SRP in the high-delta (3.5–4.5 Hz) frequency band is, that is, the more the high-delta oscillation is. However, the increase of low-alpha oscillation is not associated with the deterioration of motor function in early PD.

Zoom In Zoom Out Reset image size

In this study, we adopted deep learning technology to detect the abnormality of personalized spatial relative power in the optimal frequency band of EEG in early PD patients. We used azimuthal equidistant projection technology to convert the PSD calculated from each electrode channel of EEG into 2D spatial representation as the input of CNN model by preserving the physical meaning of convolution operation and achieved the results of feature visualization. And then, the redefined optimal frequency bands for identifying early PD were obtained, and personalized identification indicators with spatial distribution characteristics were further calculated. We proved that CNN model has convincing performance for the task of distinguishing early PD patients from healthy subjects. The visualization results revealed the high-delta (3.5–4.5 Hz) and alpha (7.5–11 Hz) frequency bands that are most effective in identifying early PD. On this basis, relative power index SRP with spatial significance was calculated, which can identify abnormal personalized rhythm characteristics in early PD.

We take the channel-frequency PSD and 2D structured PSD as input data sets of training CNN models respectively, and compare the performance of each CNN model. PSD with 2D structure was more satisfied with the advantages of CNN in processing image data, and this 2D data representation method had physical correspondence, which satisfied the extraction process of CNN model to intuitively convert convolution operation into spatial information. Therefore, among all models mentioned in this work, CNN model trained with structured PSD achieved the highest ACC and AUC, which proved the effectiveness of the data representation and CNN model we designed.

The visualization results of CNN model expressed the spatial frequency characteristic distribution of each frequency point. The anomalies of early PD were mainly distributed in the high-delta (3.5–4.5 Hz) and low-alpha (7.5–11 Hz) bands, which were innovatively defined as the optimal bands for identifying the spatial frequency characteristic of early PD. Reported delta oscillations (0.5–4 Hz) have been shown to be a prominent feature of basal ganglia pathophysiology in patients with PD in relationship to tremor [43–45]. Although several studies have demonstrated the presence of delta oscillations in the single units [45–47] and LFP [44, 48] of patients with PD, this characteristic has not yet been shown in the EEG of early PD. Through visualization of CNN model, we found that spatial PSD of high-delta frequency band (3.5–4.5 Hz) has the ability to distinguish early PD in EEG data. The SRP we proposed can express spatial relative power characteristics in a specific frequency band. Oscillation was present at a frequency band of significant power [49]. Thus, the significantly elevated oscillations were present at high-delta band in EEG of early PD. Interestingly, these high-delta oscillations were prominently increased in the right frontal lobe. It is reported that less delta activities in frontal lobe are associated with better motor function in brain, that is the delta activities in EEG is negatively correlated with physical activity ability [50]. More specifically, the study [50] found that there was a relative increase in delta activities in the right frontal EEG with poorer performance of brain's ability to control physical movement, which is consistent with increased delta activities in the right frontal cortex in patients with PD. This confirms that the increased high-delta oscillations express the deterioration of motor function in early PD patients, especially the degeneration of the brain's ability to autonomic control physical movement. In addition, we found that the more severe the tremor in patients with early PD without medication, the more high-delta oscillations there were. In an animal study, Timothy et al found that in awake dopamine-depleted mice approximately half of neurons exhibit delta oscillations in dopamine depletion and demonstrated that these oscillations are a strong indicator of dopamine loss [49]. We found that the high-delta oscillations of early PD patients with dopamine drugs whose tremor symptoms were improved were significantly restored to the healthy level, which was consistent with the conclusion of animal experiments. This further suggests that high-delta oscillations in EEG are both a reliable biomarker of early PD associated with tremor and a strong indicator of dopamine loss.

In addition to the increase in high-delta (3.5–4.5 Hz) power, low-alpha (7.5–11 Hz) power also increased significantly in early PD compared with the healthy subjects. This became evidence of a (pronounced) slowdown in resting state brain activity in patients with early PD. Most studies on EEG have not been able to confirm the higher low-alpha power in patients with PD, but in a study of magnetoencephalography (MEG) in de novo PD patients, the researchers also found the higher low-alpha power [17]. As a consequence of the resulting superior spatial resolution of MEG compared to EEG, it might very well capture the oscillating brain activity changes that are not evident in the EEG signals [51]. Our proposed SRP incorporated the spatial characteristics of EEG signals to overcome the omission of spatial resolution of EEG by conventional relative power methods. In addition, we found the optimal frequency band to identify the early PD patients' power characteristics through deep learning. Therefore, the slowing rhythm features of early PD could be reflected in scalp EEG signals. Moreover, we also found the most significant distribution of power differences in low-alpha band is in the occipital lobe. The occipital lobe is the center of the visual cortex, which processes language, motor perception, abstract concepts and visual information [52]. Abnormalities in the occipital lobe can lead to impaired vision and impaired motor perception [52]. Although highly speculative, an increase in occipital low-alpha oscillations in our early PD patients may be a marker of visual processing dysfunction and impaired pathologic self-motion perception. This notion is further supported by the fact that visual self-motion perception of PD participants was significantly impaired compared with age-matched healthy controls [53]. Other researches also reported that there is visual processing dysfunction which is characterized by abnormal color vision, contrast sensitivity and visuospatial deficits in early PD [3, 48]. However, the current results are obviously insufficient to make a definitive judgment in this matter, and the phenomenon that visual self-motion perception can be improved by taking dopamine drug needs to be verified by more experiments and studies.

The brains of PD patients are heterogeneous [54], which makes it difficult to find a common biomarker that can be present in the brain of each PD patient. It is worth noting that the spatial spectral characteristic proposed in this work, SRP in the high-delta band (3.5–4.5 Hz) and the low-alpha band (7.5–11 Hz), describes the common feature of early PD population (i.e. on the basis of inter-group differences, the spatial spectral characteristic found in this work not only meet the heterogeneity of early PD but also can completely distinguish early PD from HC), thus fully distinguishing early PD patients from healthy controls at the individual level. At present, the heterogeneity of the brain is often ignored in the studies of the spectral characteristics of early PD, which reflects the advanced nature of this work. Our findings not only provide an effective physiological indicator for the diagnosis of early PD, but also provide a new perspective for the study of the mechanism of brain pathology in early PD.

We demonstrated the feasibility of applying CNN model to identify the patients with early PD with the ACC of 99.87% ± 0.03% and the AUC of 99.87% ± 0.05%. Based on the interpretability of the model and the visualization of the decision process, the model indicated the characteristic frequency bands (high-delta (3.5–4.5 Hz) and low-alpha (7.5–11 Hz) frequency bands) and spatial distribution features that are used to identify the early PD. According to obtained spatial distribution characteristics, we proposed a novel personalized indicator incorporated spatial features, SRP, for identifying frequency abnormalities of early PD. We also found that the more severe the tremor in early PD patients, the higher the high-delta bandpower in the EEG. In addition, increased high-delta and low-alpha bandpower suggested that slowing resting state brain activities occurred in early PD, and these abnormalities were significantly improved after dopamine ingestion. The above results indicate that deep learning technology, with its powerful learning ability, has enabled us to find the optimal frequency bands for early PD and the effective personalized biomarker, spatial spectral characteristic, to identify the abnormal characteristics of early PD. The findings of this study will provide a novel research perspective for the medical field and provide an effective reference for the auxiliary diagnosis of early PD in clinical practice.

The authors would like to thank the editors and the reviewers for their critical and constructive comments and suggestions.

The datasets generated for this study are available on request to the corresponding author.

The data generated and/or analysed during the current study are not publicly available for legal/ethical reasons but are available from the corresponding author on reasonable request.

The studies involving human participants were reviewed and approved by Medical Ethics Committee of Tianjin Medical University General Hospital. All subjects understood the purpose of collecting the data and the significance of the study, and signed the informed consent form.

This work was supported in part by the National Natural Science Foundation of China under Grant No. 62173241, in part by the Natural Science Foundation of Tianjin, China (Grant No. 20JCQNJC01160), in part by Ministry of Science and Technology of China (Grant Nos. 2016YFC1306500, 2016YFC1306504), in part by National Key R&D Program of China (Grant No. 2016YFC1306501) and in part by the Foundation of Tianjin University under Grant 2020XRG-0018. The authors also gratefully acknowledge the financial support provided by Opening Fundation of Key Laboratory of Opto-technology and Intelligent Control (Lanzhou Jiaotong University), Ministry of Education (KFKT2020-01).