Real-Time Precise Targeting of the Subthalamic Nucleus via Transfer Learning in a Rat Model of Parkinson’s Disease Based on Microelectrode Arrays

In neurodegenerative disorders, neuronal firing patterns and oscillatory activity are remarkably altered in specific brain regions, which can serve as valuable biomarkers for the identification of deep brain regions. The subthalamic nucleus (STN) has been the primary target for DBS in patients with Parkinson’s disease (PD). In this study, changes in the spike firing patterns and spectral power of local field potentials (LFPs) in the pre-STN (zona incerta, ZI) and post-STN (cerebral peduncle, cp) regions were investigated in PD rats, providing crucial evidence for the functional localization of the STN. Sixteen-channel microelectrode arrays (MEAs) with sites distributed at different depths and widths were utilized to record neuronal activities. The spikes in the STN exhibited higher firing rates than those in the ZI and cp. Furthermore, the LFP power in the delta band in the STN was the greatest, followed by that in the ZI, and was greater than that in the cp. Additionally, increased LFP power was observed in the beta bands in the STN. To identify the best performing classification model, we applied various convolutional neural networks (CNNs) based on transfer learning to analyze the recorded raw data, which were processed using the Gram matrix of the spikes and the fast Fourier transform of the LFPs. The best transfer learning model achieved an accuracy of 95.16%. After fusing the spike and LFP classification results, the time precision for processing the raw data reached 500 ms. The pretrained model, utilizing raw data, demonstrated the feasibility of employing transfer learning for training models on neural activity. This approach highlights the potential for functional localization within deep brain regions.


I. INTRODUCTION
T HE subthalamic nucleus (STN) has been the most effec- tive therapeutic target for deep brain stimulation (DBS) in patients with Parkinson's disease (PD), especially for addressing motor symptoms and dystonia [1], [2].Misplacement of the DBS lead can potentially aggravate side effects by stimulating adjacent brain regions [3].The subjective judgments made by clinicians increase the risk of electrode misplacement [4], [5].Consequently, the adoption of novel methods is imperative for achieving precise functional localization in targeted deep brain regions.
Clinicians assess the accuracy of stimulating electrode placement by considering the quality of microelectrode recordings in the target region and surgeons' experience.The utilization of electroencephalography (EEG) and electrocorticography (ECoG) electrodes for sensing and stimulation introduces technical challenges, as potential contacts around stimulation sites may be blocked.In contrast, MEAs, when employed for sensing, facilitate higher-resolution signal detection across the temporal and spatial dimensions.[14], [15], [16].MEAs enhance the detection of local field potentials (LFPs) and spikes with millisecond precision thereby yielding a richer neural information profile [17].The incorporation of advanced algorithms enables the automated extraction of these detected neural signals embedded within MEAs.This type of automation plays a crucial role in precisely guiding electrode placement in specific target brain regions.
Microelectrode recordings (MERs) have demonstrated considerable potential, as biomarkers for targeting the STN.Two types of electrophysiological signals, LFPs and spikes, exhibit marked differences in the STN and its adjacent upper and lower regions.LFPs have been proven to be robust guidance tools for identifying the STN.Beta band oscillatory activity (12)(13)(14)(15)(16)(17)(18)(19)(20)(21)(22)(23)(24)(25)(26)(27)(28)(29)(30) is enhanced in patients with PD [13].The delta band (0-4 Hz) has also been positively linked to tremor and voluntary movement [18].On the other hand, spike characteristics vary across different brain regions.Properties such as waveform amplitude and spike firing rate are enhanced in the STN compared to these properties in other brain regions [19], [20], [21].Thus, it may prove valuable to broaden our focus beyond just the specific oscillatory patterns and account for the distinctive features of spike firing characteristics.
In this study, we recorded electrophysiological signals over neighboring brain regions of the STN.The brain regions from top to bottom are zona incerta (ZI), STN and cerebral peduncle (cp).A novel MEA was used to obtain more neural signals, with adequate recording quality for the oscillatory and spike activities in anesthetized PD rats.In addition to analyzing the time-domain and frequency-domain neural signals, a deep residual network based on transfer learning was designed and implemented to classify raw MERs.Such algorithms have the capacity to harness the complex information from raw signals, allowing the differentiation among the ZI, STN, and cp regions.Simultaneously, this approach enhances the efficiency by diminishing the input data size, thereby elevating the frame rate for data acquisition and accelerating the overall data processing speed.This optimization facilitates the advancement of real-time data processing capabilities.

A. Animals and MEA Design
In this study, three male Sprague-Dawley (SD) rats (250-300 g; sourced from Vital River Beijing) were used.These animals were group-housed, with two rats per individually ventilated cage (IVC) measuring 42 cm in length, 28 cm in width, and 27 cm in height.The rats were kept on a 12-hour light--dark cycle and provided full access to both food and tap water.The Institutional Animal Care and Use Committee at the Aerospace Information Research Institute, Chinese Academy of Science, approbed all of the experiments performed in this study, which were all carried out in accordance with the regulations established by the Beijing Association on Laboratory Animal Care (AIRCAS-006-1, 9/14/2022).
The MEA probe comprises a total of 16 microelectrode sites labeled "a" to "p," each possessing a diameter of 20 µm (as shown in Fig 1 .a).The distribution of these sites is detailed as follows:11 sites (a-k) are arranged at intervals of 39 • based on the shape of the STN to maximize the detection area, and 5 sites (l-p) are arranged to ensure that the depth of the MEA is sufficient (Fig 1 .a).The A and B sites are reference electrodes.The spacing between each contact site on the probe is 50 to 70 µm, and the total depth of the recording sites along the probe shaft extends to 800 µm.This configuration facilitates the acquisition of recordings from different layers within the ZI, STN, and cp.The MEA probe is 9.2 mm long and 430 µm wide.On a silicon-on-insulator (SOI) substrate, all the fabrication processes were carried out utilizing the microelectromechanical system (MEMS) technique.(Fig. S1).The conductive layer is composed of platinum and is 250 µm thick, and the insulating layer is made of SiO2 and has a thickness of 700 nm.After the MEAs were fabricated, the recording sites were electrodeposited with platinum nanoparticles (PtNPs) to improve the signal-to-noise ratio [30], [31] (Fig. S2).The MEA was used to detect neuroelectrophysiological activities, encompassing spikes and LFPs, in the three targeted brain regions.

B. Drug Regimen and MEA Implantation Procedure
All procedures were conducted under anesthesia using the RWD520 isoflurane anesthesia apparatus (RWD, USA), which maintained an anesthetic concentration of 1.5% and an airflow rate of 600 mL/min.The depth of anesthesia was assessed by toe pinching throughout the surgical procedures.All animals received subcutaneous injections of marbofloxacin for a period of 7 days following the surgery.
The heads of the rats were securely fixed in a stereotaxic frame (51600, Stenting, USA) using ear bars.The scalp was shaved, and disinfected with a swab, and a central incision was carefully made to expose the skull.A PD model was established according to previous study without a desipramine pretreatment to induce Parkinsonian symptoms in rats [32],which was confirmed by APO (apomorphine, Aladdin, China) and behavior observations.First, 6-OHDA (ACMEC, China) (8 µg of 6-OHDA dissolved in 4 mL of a 0.2% ascorbate solution) was injected into SNc (anteriorposterior (AP): −4.9 mm, medial-lateral (ML): 2.0 mm, For the control group, an equivalent dose of saline was injected following the same procedure.After surgery, the rat's scalp was sutured, and 100,000 units of penicillin were injected subcutaneously every day to prevent postoperative infection (Fig 1 .b).After 7 days, an intraperitoneal injection of APO (0.1 mg/kg) was administered to verify whether the rats developed the PD model.A successful PD model was identified if the rat rotated toward the nondrug side at a speed exceeding 7 revolutions per minute (rpm) (Supplemental Video) (Fig. S3).
One month after surgery, the PD rats were placed in a signal shielding box (55 cm × 45 cm × 50 cm).The box was designed to decrease noise from the surroundings.The rats were anesthetized using isoflurane.To ensure accurate insertion of the MEA into these three brain regions, a combined anatomical and physiological approach was employed.A craniotomy (1 × 1 mm 2 ) was performed by drilling on each rat's head, following the guidance of the rat brain atlas [33].The coordinates for the craniotomy, which encompassed the STN, ZI and cp areas, were as follows: AP: −3.6 mm, ML: −2.5 mm (Fig 1 .c,Fig. S4).Three craniotomies were performed on the skull to serve as grounding sites for the animals.MEA recordings were obtained under isoflurane anesthesia with this experimental setup, which suppressed interference from the surroundings and movements.The MEAs were inserted into the target brain regions at the above coordinates.Recording sessions were performed in all animals within the time of 14:00 to 16:00 pm.MEA signals from the ZI, STN and cp were recorded using the Cerebus™Data Acquisition System.To ensured that the MEA was inserted accurately in these three brain regions, a combination of anatomical and physiological approach was applied.Each recording session, which consisted of two consecutive sweeps lasting 600 s, started 10 mins after implantation in the specific brain region to minimize instrumental errors.Raw signals (30 k samples/s) were saved as TXT data format before being transformed to LFP and Spike.The LFPs underwent band-pass filtering within the range of 1-150 Hz and were sampled at a rate of 30 kHz.The spikes were extracted based on a threshold-dependent approach.

C. Data Processing and Analysis
Epochs lasting 600 s were employed for time-domain processing of spike signals and frequency-domain processing of LFPs in different brain regions.First, an eighth-order, zero-phase shift, noncausal, bandstop Butterworth filter was employed.The lower and upper cutoff frequencies of this filter were set at 47 Hz and 53 Hz, respectively, aiming to eliminate noise at approximately 50 Hz.Subsequently, a threshold-dependent approach was employed to remove any artifacts related to the data acquisition method.To determine the threshold, the instrument was run for at least 10 min prior to signal recording to ensure the stability of the background noise.The threshold for detecting neural spikes was three times the maximum amplitude of the background noise.Neural spikes were extracted using a high-pass filter (>250 Hz), while LFPs were acquired using a low-pass filter (0-150 Hz).To alleviate the spectral leakage phenomenon, a Hanning window function was applied.

D. Statistics
Spike units were extracted using K-means scan (Offline Sorter, Plexon, USA).The isolation of units was based on three primary features: waveform peak-valley, peak amplitude and principal component analysis.Clusters with similar valid waveforms were manually defined, ensuring that single units exhibited no refractory period of less than 1 ms.The waveform, firing rate and amplitude of the spikes, as well as the power and percentage values of the oscillatory activity, were considered important factors among the different areas (ZI, STN and cp).The characteristics of the spikes and LFPs were statistically analyzed using Origin 2021 (Origin Lab, USA).Significant testing was conducted using one-way analysis of variance (ANOVA) or one-way repeated measures ANOVA test, followed by Tukey's honest significance test (HSD) to control for multiple comparisons (Python 3.8).* P < 0.05 and * * P < 0.01 were considered to indicate statistically significant differences.All the data in this paper were reported as the mean value ± standard error (SE).

E. Transfer Learning
Convolutional neural networks (CNNs) offer substantial benefits in many learning tasks, and can be helpful in understanding unnoticed details.However, as the learning tasks become increasingly complex, the demands on the network layers increase, leading to the vanishing gradient problem.Additionally, deeper neural networks have many more parameters in the training model.Transfer learning methods alleviate these computational requirements by employing pretrained deep models from large databases and adapting the model to new applications [34], [35], [36].Therefore, MERs are generally considered complex, exhibiting noise-like characteristics with significant temporal and subject variance.MERs are typically processed with deep neural networks, including preprocessing steps such as threshold-dependent approaches and downsampling.To enable real-time classification of these three brain regions in the rodents, the recordings were split into n samples of specific sizes.The training and testing were defined to encompass distinct and nonoverlapping time intervals, while the training sets intentionally overlapped to increase the size of the datasets.The proposed brain region classification algorithm, which consists of four main steps (data splitting, transformation, model training, and result fusion), is shown in Fig 2 .1) Data Splitting: Each channel's recordings were employed individually.Every sample was obtained from one of the 16 channels in the MEA.By reducing the input data adopting this strategy, the model considered the measurement obtained by a single electrode, simplifying the configuration contrasted with the 16-channel MEA.To balance the identification accuracy and speed, the raw neural signals were split into different sizes for time-domain and frequency-domain training.Spikes were identified using 500 ms of data with a 50 ms-step, while LFP features were extracted from 10 s of data with a 500 ms step.
2) Transformation: To distinguish valuable information from Gaussian noise with increased temporal precision, spike series were transformed into Gramian angular fields (GAFs), which preserve time dependance [37] (1).In Formula 1, i and j are the vectors in the spike series, and φ i, j is the angle between vectors i and j.Since the frequency-domain features from the LFPs were distinct, the spectrum conversion of the LFPs using the fast Fourier transform (FFT) was used as a dataset.Then the GAFs of the spike series and the FFT of the LFP series were transformed to 128 * 128 images respectively.
3) Model Training: The emergence of CNNs has led to improved signal processing performance.AlexNet, VGGNet, GoogLeNet and ResNet are pretrained CNNs with different convolutional layers for feature extraction and offer various advantages when trained with large databases.AlexNet consists of 5 convolutional layers for feature extraction and 3 fully connected layers.VGGNet includes additional convolutional layers and pooling layers to reduce the number of parameters.GoogLeNet uses a multipath design to extract both high-and low-level features.ResNet mitigates the issue of vanishing gradients in deeper neural networks.The MERs were input into these pretrained CNNs to determine the model that performed best based on the MERs.These models were implemented using TensorFlow 2.0 in Python 3.8.

4) Results Fusion:
To evaluate the model's performance, cross-validation was employed.The complete dataset was divided into a training set, a validation set and a test set at a ratio of 7: 1.5:1.5.The evaluation metrics selected included the accuracy, precision, recall and the F1 score.After training, the spike and LFP classification results were fused.The fusion weight was calculated according to the training accuracy (Formula 2).Specifically, the fused weights W f used were determined by incorporating the training accuracy (Acc train modeli ) of each individual model into the fusion process.The fused weights were calculated using the following equation: where W modeli represents the weight assigned to the i-th individual model, and n is the total number of models in the ensemble.This approach ensures that the contribution of each model to the ensemble is proportional to its training accuracy, thereby optimizing the overall performance of the fused model.Overall, this algorithm was applied to accurately and rapidly identify the abovementioned three brain regions.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

A. Time Domain Analysis of Spike Data
When the MEAs were inserted into the ZI, STN and cp, sites 'a' to 'k' were inserted into these three brain regions, while sites 'l' to 'p' were inserted into the ZI or STN.The spikes recorded in the ZI, STN and cp were sorted via principal component analysis and valley-seeking methods.After processing the spikes, a total of 204, 267, and 155 spike units were recorded from ZI, STN and cp respectively, across all five PD rats.All spikes were classified into three types according to the target brain regions.As depicted in Fig 3 .a,a significant difference in firing rate was observed between the control group and the PD model group across these three brain regions (P<0.05).In PD model rats, the spike firing rate in the STN was substantially greater than that in the other two brain regions, and was at least threefold greater than that in the ZI and cp.The frequency in the STN was approximately 9.79 ± 0.46 Hz after statistical analysis, while the firing rates between ZI and cp exhibited minimal variation (P<0.01).Further details about the mean spike waveform were extracted.The aggregated waveforms shown in Fig 3 .bindicate that the waveform in the STN is the highest and narrowest compared to that of ZI and cp.Compared to those in the control group, the fluctuations in the aggregated waveforms in the PD model rats were greater (Fig. S5).The spike waveform width was defined as the duration between the peak and valley, and the amplitude is defined as the difference in value between peak and valley.The spike waveform width in the STN was the narrowest among the various regions, with a value of 0.40142 ± 0.01963 ms.In contrast, the spikes in the cp had a maximal width of 0.63095 ± 0.01851 ms (P < 0.05).Regarding the waveform amplitude, the spikes in the STN demonstrated large amplitudes of 73.09 ± 0.71 µV, while the spikes in the ZI had an amplitude of 60.95667 ± 0.8233 µV (P < 0.01) (Fig 3 .c).These results demonstrated that spike characteristics, including waveform and firing rate, contribute significantly to determining the functional localization of the STN in PD rats.The ZI and cp can be distinguished based on their spike waveforms, particularly in terms of their spike amplitude.

B. Frequency Domain Analysis of the LFP
The LFP is a complex neural signal that includes the combined electrophysiological activity of a group of neurons surrounding the MEA [38], [39].To improve the identification of the ZI, STN and cp in PD model rats, spectrograms of the LFPs are shown in Fig 4 . a, c, e.As shown in the figures, the LFP power spectral density significantly increased in the STN, mainly in the 0-30 Hz range.Then, 0-30 Hz LFPs were extracted as depicted in Fig 4 . b, d, andf.The total power of the LFP in the STN was the largest among the considered brain areas.Furthermore, the LFP power peaked between 0-4 Hz in both the STN and cp, while the peak appeared at approximately 4-10 Hz in the ZI.Notably, there were large oscillation bands between 10-20 Hz LFPs in the STN.Therefore, the powers of the delta (0−4 Hz) and beta (12−30 Hz) bands of the LFPs were statistically analyzed in each of the three brain regions.As shown in Fig 4 .g,h, the delta and beta band LFP powers of the control group and PD model rats were compared.In the control group, the delta band power predominantly prevailed over the beta band power in all three brain regions (P<0.05)(Fig 4 .g).In PD model rats, as depicted in Fig 4 .h,the LFP power in the delta and beta bands were greatest in the STN (delta band: 0.10475 ± 0.008108 mW, beta band: 0.20505 ± 0.05926 mW), and the beta band exhibited a significant increase in LFP power, surpassing the delta band.Additionally, there was an evidence that the LFP power in the beta band was greater in the ZI (0.092661 ± 0.0029232 mW) than in the cp (0.048437 ± 0.002681 mW).These results indicate that the LFP power is significantly different among various brain regions, both in the overall analysis and the analysis of the different frequency bands, particularly the delta and beta bands.The difference between the ZI and cp was mainly reflected in the LFP beta band.Beta band activity may serve as a potential biomarker to identify brain regions around the STN, which is consistent with the results of previous work [40], [41].

C. Transfer Learning
The analysis of the neural activity data using frequency and time domain signal processing methods revealed substantial  I.The number of datasets is also included in Table I.
The model performances based on the spike and LFP data was validated by indices including the accuracy, precision, recall and F1 score (Table II).An accuracy of 94.85%, a precision of 92.42%, a recall of 92.74%, and an F1 score of 92.56% were achieved based on the 500 ms-spike data, and an accuracy of 97.57%, a precision of 98.5%, a recall of 96.60%, and an F1 score of 98.3% were achieved based on the 10 s-LFP data.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.Fusing these two results could help balance the data processing speed and the brain region classification accuracy (Table III).These findings demonstrate that a simpler hardware configuration for the MEA is adequate for identifying different brain regions.
IV. DISCUSSION In studies on electrophysiological information, including multiple-frequency spikes and neural oscillatory activity, researchers have focused on analyzing differences in spike firing patterns and commonly defined frequency bands such as delta and beta oscillatory activity in the brain.
Microelectrode recording signals, including spikes and LFPs, play crucial roles in intraoperative navigation.Analyzing various neuron firing patterns and specific frequency bands can elucidate the roles of distinct brain regions.The precise functional location of the target brain region is important not only in PD, but also in other neurodegenerative diseases such as epilepsy [42].The application of advanced algorithms for analyzing MERs is projected to lead to more precise lead placement during surgical procedures in the future.
In this study, we examined the spike and LFP data obtained from the ZI, STN and cp.The spike firing rate in the STN was greater than that in the ZI and cp.The spike amplitude was highest in the STN, and the spike width was smaller than that in the other two brain regions.The differences in the spikes between the ZI and cp were mainly reflected in the width and amplitude.When analyzing the LFP characteristics of these three brain regions, the total LFP power in the STN was the highest, especially in the delta and beta bands, and the LFP power in the beta band in the ZI was greater than that in the cp.Therefore, the interaction between spikes and oscillatory activities, including beta and delta activity, demonstrates the significance of using oscillatory activity and spikes as biomarkers for function localization of STN [13], [43], and elucidating the size of the training data needed for deep learning algorithms.
To further improve the accuracy of such biomarkers, the concealment of patterns within MEA recordings may yield additional insights, through training the ResNet model with pretrained weights [44].During the experiment, the feasibility of identifying MEA signals among the three brain regions was successfully demonstrated by employing the proposed pretrained neural networks with fusion results.Moreover, transfer learning, which can capture underlying features using versatile algorithms, offers a logical foundation for an alternative approach that eliminates the necessity of preprocessing and achieves reliable and leading classification performance [45].
The algorithm utilizing transfer learning could be applied to analyze both time and frequency domain signals acquired from the recording device linked to the MEA [42].The input data were restricted to a minimum of 500 ms, allowing this algorithm to synergize effectively with hardware processing to increase refresh rates and accelerate the overall data processing.In addition, the proposed algorithm included automated detection of different brain regions through the combination of LFPs and spikes in PD rats, thus offering better biological interpretability.Although this approach cannot be applied to aDBS treatments, it provides feasible real-time evidence for functional localization.Future investigations could focus on the application of aDBS treatments using biocompatible MEAs in conjunction with advanced algorithms.
Importantly, MEA recordings have promising advantages in medical treatments for various neurodegenerative disorders such as Alzheimer's disease [46].MEAs were mass-fabricated using microelectromechanical system techniques, with the detection sites featuring a diameter of 20 µm diameter, allowing for the detection of neurons at higher spatial resolution.The MEAs were modified with PtNPs, which markedly improved their electrical performance by reducing the impedance and minimizing the phase delay [47].In the present study, novel materials such as SWCNTs and PEDOT: PSS were used to improve biocompatibility [48], [49].The proposed MEAs provided higher spatial and temporal resolution for detecting single-unit neuronal activity and multiunit neuronal activities than did existing devices.
Importantly, to validate the 6-OHDA-induced rat model of Parkinson's, we established a proof of concept for employing pretrained ResNet models in identifying functional brain regions.These promising outcomes should be put into practice and rigorously examined in long-term studies involving

CONCLUSION
This research describes a functional localization method within the brains of PD rats, employing spike firing patterns and LFP spectrum signatures.The administration of 6-OHDA to the SNc and MFB leads to Parkinsonian-like akinesia and rigidity.The designed 16-channel MEA enables real-time monitoring of neuronal activity across the upper to lower regions of the STN in PD rats.This setup presents the opportunity to potentially employ the pretrained ResNet algorithm as a comprehensive approach for classifying MEA recordings in real time.The fusion of 500 ms-spike and 10 s-LFP identification data in different brain regions could lead to more accurate results with high temporal precision.The high accuracy and temporal precision of this classification model offer a solid theoretical basis for functional localization research.

Fig. 1 .
Fig. 1.A diagram of the approach for functional localization of brain regions using a pretrained ResNet model based on the MEA recordings for the classification and prediction of the ZI, STN and cp.(a) The 16-channel MEA was designed based on the shape of the target brain regions.(b) The distribution of the 16 sites in the MEA.(c) The MEA recording data were collected in the ZI, STN and cp over 10 minutes after establishing and validating PD symptoms in the rats.(d) The target brain regions include three brain regions at different depths (ZI, STN and cp).(e) LFP signals and spike firing patterns of different channels in the target brain regions of PD rats (n=5).(f) The spike and LFP data were separately input into the pretrained ResNet separately, then the results were fused to identify brain regions.

Fig. 2 .
Fig. 2. A flow chart showing the processing of raw data recorded by the MEAs via transfer learning.

Fig 1 .
d demonstrates that the LFP and spike firing time stamps recorded by electrodes in the ZI (3 representative electrodes), STN (3 representative electrodes) and cp (4 representative electrodes) differ.Note that there were distinct differences among the spike firing patterns, which led to the analysis of the time domain activity.

Fig. 3 .
Fig. 3. Characteristics of the spikes in different target brain regions.(a) Analysis of the average spike firing rates in different target brain regions for both the control group and PD model group is shown as bar graphs (mean ±SEM).(b) Aggregated waveforms of spikes in the ZI, STN and cp determined by K-means clustering in the PD model group.(c) Statistical analysis of the aggregated waveforms in terms of amplitude and width in different target brain regions.The data are shown as the mean ± SE.One-way repeated-measures ANOVA.* = p < 0.05; * * p < 0.01.

Fig. 4 .
Fig. 4. Characteristics of LFPs in different target brain regions.(a-f) Spectrograph analysis and power of LFPs in the ZI, STN and cp.(g-h) Statistical analysis of LFP power in the delta (0-4 Hz) and beta (12-30 Hz) frequency bands in the control group and PD model group in different target brain regions are shown as bar (mean ±SEM) graphs.One-way repeated-measures ANOVA.* = p < 0.05.

TABLE I PARAMETERS
OF PRE-TRAINED NETWORK ARCHITECTURE AND DATASETS Fig. 5. Accuracy of different pretrained CNNs after fusion of spike and LFP results.

TABLE II PERFORMANCE
OF RESNET TO CLASSIFY SPIKE AND LFP IN DIFFERENT BRAIN REGIONS

TABLE III SUMMARY
OF IDENTIFYING BRAIN REGIONS METHODS FOR PARKINSON'S DISEASEthe application of aDBS, encompassing diverse experimental models of neurological disorders.