Closed-loop deep brain regulation and control system for epilepsy treatment

Abstract

The invention discloses a closed-loop deep brain regulation and control system for epilepsy treatment, and the system comprises an electroencephalogram signal collection module which collects neural electrophysiological signals in real time by implanting a recording electrode in a cerebral cortex or hippocampal brain area; the electroencephalogram signal real-time analysis module is used for outputting the epileptic seizure period prediction probability xi through a seven-layer full-connection neural network based on the 33-dimensional feature vector so as to realize epileptic seizure early warning; when the zeta is greater than or equal to a threshold value, the electrical stimulation regulation and control module is used for triggering electrical stimulation, filtering out a frequency band required to be regulated and controlled according to filtering, calculating and predicting electrical stimulation intensity and then correcting and outputting pulse stimulation, so that personalized electrical stimulation regulation and control are carried out on a specific frequency band, and detection is continuously carried out until the frequency band becomes normal; and the medicine intervention module is used for triggering a micro-injection pump to carry out medicine infusion through a pulse signal. According to the system, epilepsy treatment is improved from empirical parameter adjustment to accurate regulation and control based on real-time nerve physiological signals, and a safer and more effective treatment choice is provided for a drug-refractory epilepsy patient.

Description

Technical Field

[0001] The present invention relates to the field of neuroscience and medical equipment technology, and specifically to a closed-loop deep brain regulation system based on multimodal feature analysis of electroencephalogram (EEG) signals, which is suitable for real-time and precise intervention in drug-refractory epilepsy. Background Art

[0002] Epilepsy, a chronic brain network disorder characterized by abnormal, hypersynchronized discharges of neuronal clusters, develops drug-resistant epilepsy in 35% to 40% of the approximately 51 million patients worldwide (WHO, 2023). Deep brain stimulation (DBS) uses implanted electrodes to electrically stimulate specific brain regions. Since its initial application in the 1990s for movement disorders, it has gradually expanded into the treatment of epilepsy. Although continuous electrical stimulation of key nodes in the epileptogenic network, such as the anterior thalamic nucleus and hippocampus, has been shown to reduce seizure frequency in some patients with refractory epilepsy, traditional open-loop DBS, which uses a fixed-frequency (130-185Hz) continuous stimulation pattern, has two core drawbacks: nonphysiological interference and low energy efficiency, thus limiting its therapeutic efficacy.

[0003] Closed-Loop DBS is a new intervention for intractable epilepsy. It can dynamically optimize stimulation parameters through a real-time feedback mechanism of "detection-decision-stimulation". Theoretical breakthroughs in closed-loop control technology The electrophysiological process of epileptic seizures provides a clear feedback target for closed-loop regulation: low-frequency oscillations of the local field potential (LFP) are often enhanced seconds to minutes before the onset of the attack, and the attack is manifested as high-frequency synchronized discharges (80-200Hz), followed by slow-wave activity suppression after the attack. Clinical studies have shown that the closed-loop mode can improve the efficacy by 28% and reduce the incidence of side effects by 43% compared with open-loop stimulation. However, although closed-loop control has shown significant advantages in epilepsy treatment, its clinical application still faces multiple technical bottlenecks, including insufficient sensitivity in identifying the pro-ictal phase of epilepsy initiation, the inability to track network state transitions in real time, resulting in a high false alarm rate, and the inability of stimulation parameters to adapt to the onset phase, resulting in delayed intervention.

[0004] Therefore, addressing these technical pain points, developing a closed-loop deep brain control system with adaptive learning capabilities has significant clinical value. This system must overcome the limitations of traditional closed-loop control. By analyzing the time-frequency domain characteristics of multi-channel EEG signals in real time, it constructs a personalized epileptogenic activity prediction model and dynamically adjusts stimulation parameters based on the model output, achieving a closed-loop control approach of "precise detection - intelligent decision-making - efficient intervention." This technological innovation is expected to elevate epilepsy treatment from empirical parameter adjustment to precise regulation based on real-time neurophysiological signals, providing safer and more effective treatment options for patients with drug-resistant epilepsy. Summary of the Invention

[0005] The purpose of the present invention is to provide a closed-loop deep brain regulation system based on the multimodal characteristics of EEG signals, which solves the problems of low prediction accuracy, delayed regulation and insufficient individualization in the existing technology by accurately identifying biomarkers of the pro-ictal period, dynamically adapting the individual's neurophysiological state and synergistic intervention of electrical stimulation and drugs.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is: a closed-loop deep brain regulation system for epilepsy treatment, the system comprising:

[0007] EEG signal acquisition module: By implanting recording electrodes in the cerebral cortex or hippocampus, neural electrophysiological signals are collected in real time;

[0008] Real-time EEG signal analysis module: Based on a 33-dimensional feature vector, it outputs the predicted probability of the seizure-promoting period ξ through a 7-layer fully connected neural network to achieve epileptic seizure warning;

[0009] Electrical stimulation control module: When ξ ≥ threshold, electrical stimulation is triggered. The frequency band to be regulated is filtered out, the electrical stimulation intensity is calculated and predicted, and then the output pulse stimulation is corrected, thereby performing personalized electrical stimulation control for specific frequency bands. Continuous monitoring is carried out until the frequency returns to normal and the electrical stimulation ends.

[0010] Drug intervention module: triggers the microinjection pump to infuse drugs through pulse signals.

[0011] Furthermore, in collecting neuroelectrophysiological signals, the sampling rate is ≥ 1 kHz.

[0012] Furthermore, the real-time EEG signal analysis module includes a fully connected neural network established for pre-ictal state detection, wherein the fully connected neural network includes an input layer, a hidden layer, and an output layer. The input layer inputs a 33×1-dimensional real-valued EEG feature vector, and the hidden layer includes 7 fully connected layers, which simplify the feature vector into a single real value, called ξ; the output layer outputs the ξ value, ξ∈[0,1].

[0013] Furthermore, in the neural network training, except for epileptic seizures, the interictal period (quiescent period) and the pro-ictal period (before the seizure) were manually marked every 5 minutes, and the average feature vector (33-dimensional parenchymal EEG feature vector ξ value) was calculated 5 minutes before training; 80% of the data were randomly extracted for 4-fold training and cross-validation, and the remaining 20% ​​were used for model accuracy verification.

[0014] Furthermore, the feature vector includes 26-dimensional phase-amplitude coupling values, 1-dimensional complexity of the EEG signal, and 6-dimensional frequency band power.

[0015] Furthermore, the calculation of the phase-amplitude coupling value requires dividing the high frequency band into intervals and using Hilbert transform to extract the phase-amplitude coupling strength. Features 1 to 26 are the average PAC values ​​of each frequency in the high frequency band.

[0016] Furthermore, the high frequency band is 50–100 Hz, which is divided into 26 frequency bands at intervals of 2 Hz. The Hilbert transform is used to calculate the phase-locking value of the amplitude of each high frequency band and the phase of the low frequency band (an indicator to measure the strength of phase-amplitude coupling), and the average value of each frequency in the high frequency band within 5 minutes is taken as features 1–26.

[0017] Furthermore, the complexity of the electroencephalogram signal is calculated by an algorithm of Lempel-Ziv binary sequence complexity measurement to obtain a characteristic value.

[0018] Furthermore, the power of each frequency band is obtained by applying the Welch method to the EEG and averaging the PSD within each frequency band.

[0019] Furthermore, in electrical stimulation regulation, if electrical stimulation is triggered, the initial pulse intensity is 1 mA, and then increases by 0.5 mA every 20 seconds until ξ decreases ≥ 0.05 or reaches the maximum intensity of 5 mA and stops; in drug intervention, the single dose is calculated based on body weight, the metering range is 0.1-10 mg, and the dosing interval is ≥ 15 minutes.

[0020] Compared with the prior art, the present invention has the following technical advantages:

[0021] Multimodal feature fusion: Integrates time domain (power), frequency domain (PAC), and nonlinear (complexity) features to improve prediction sensitivity (reducing false positive rate by 35%).

[0022] Dual closed-loop control: Electrical stimulation and medication work synergistically, applicable to different seizure types (electrical stimulation is preferred for generalized seizures, and medication is preferred for focal seizures). This system elevates epilepsy treatment from empirical parameter adjustment to precise control based on real-time neurophysiological signals, providing a safer and more effective treatment option for patients with drug-refractory epilepsy.

[0023] Personalized thresholds: Based on patient-specific data training model, auROC ≥ 0.85, detection latency ≤ 100ms, sensitivity and specificity ≥ 85%. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is the hardware architecture diagram of the closed-loop control system (showing the connection between electrode implantation, signal processing and control modules);

[0025] Figure 2Schematic diagram of closed-loop electrical stimulation equipment (including EEG recording, electrical stimulation and drug delivery devices);

[0026] Figure 3 It is the closed-loop control logic diagram (trigger-termination conditions of electrical stimulation and drug intervention);

[0027] Figure 4 Receiver operating characteristic curve (auROC=0.8554, showing model classification performance). DETAILED DESCRIPTION

[0028] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] like Figures 1 to 3 , a closed-loop deep brain control system for epilepsy treatment, the system comprising:

[0030] 1. EEG signal acquisition module: By implanting recording electrodes in the cerebral cortex or hippocampus, neural electrophysiological signals are collected in real time.

[0031] 1.1 Data acquisition: EEG signals were continuously recorded using implanted electrodes, segmented into 5-minute windows, and synchronously labeled as pre-ictal or interictal periods.

[0032] 1.2 Feature extraction (33-dimensional vector): including phase amplitude coupling value (26 elements), complexity of EEG signal (1 element) and power of each frequency band (6 elements).

[0033] Phase-Amplitude Coupling (PAC, 26 dimensions):

[0034] Low frequency band (phase): 2–30 Hz (covering δ, θ, α, and β bands), high frequency band (amplitude): 50–100 Hz, divided into 26 frequency bands at 2 Hz intervals (e.g., 50–52 Hz,…, 98–100 Hz).

[0035] The Hilbert transform was used to calculate the phase-locking value between the amplitude of each high-frequency band and the phase of the low-frequency band, and the average value within 5 minutes was taken as features 1–26.

[0036] Signal complexity (1 dimension, features 27):

[0037] After binarizing the EEG signals, the Lempel-Ziv algorithm (LZ76) was applied to calculate the sequence complexity, which reflects the disorder of neuronal activity.

[0038] Band power (6 dimensions, features 28–33):

[0039] The power spectral density (PSD) of the δ (0–4 Hz), θ (4–8 Hz), α (8–13 Hz), β (13–30 Hz), low γ (30–50 Hz), and high γ (50–100 Hz) frequency bands was calculated using the Welch periodogram method (window length 2 s, overlap 50%), and the average value of each frequency band was taken.

[0040] 2. Real-time EEG signal analysis module: Based on a 33-dimensional feature vector, a 7-layer fully connected neural network outputs the predicted probability of the seizure phase, ξ, to provide early warning of epileptic seizures. This module includes the establishment of a fully connected neural network for pre-ictal state detection and the training of this neural network model.

[0041] 2.1 Neural Network Architecture:

[0042] Input layer: 33-dimensional feature vector;

[0043] Hidden layer: 7 fully connected layers, the number of neurons is 128 → 64 → 32 → 16 → 8 → 4 → 2, and each layer uses the Sigmoid activation function;

[0044] Output layer: single neuron, output value ξ∈[0,1], representing the probability of seizure precipitating period (ξ≥threshold is considered abnormal).

[0045] 2.2 Training process:

[0046] Data partitioning: 80% of the data is used for training (4-fold cross validation), and 20% is used for testing;

[0047] Optimization goal: mean square error (MSE) ≤ 1e-5, maximum iteration 200 times, learning rate 0.05;

[0048] Evaluation metrics: area under the receiver operating characteristic curve (auROC), individual thresholds were optimized by Youden index, ranging from 0.24 to 0.5.

[0049] 3. Electrical stimulation control module: When ξ ≥ threshold, electrical stimulation is triggered. The frequency band that needs to be controlled is filtered out, the electrical stimulation intensity is calculated and predicted, and then the output pulse stimulation is corrected, thereby performing personalized electrical stimulation control on the specific frequency band and continuing to detect until it becomes normal and the electrical stimulation ends.

[0050] 3.1 Electrical stimulation triggering conditions:

[0051] When the prediction parameter ξ reaches the set threshold level, the system will issue an early warning for the time period and send out a signal for automatic drug administration, or use electrical stimulation to stimulate the patient.

[0052] 3.2 Dynamic adjustment of stimulation parameters:

[0053] Initial intensity: 1 mA, increasing by 0.5 mA every 20 seconds until threshold response (ξ decrease ≥ 0.05) or reaching maximum intensity (5 mA);

[0054] Frequency selection: When the abnormal frequency band is γ wave, 100Hz high frequency suppression is used; when it is δ / θ wave, 1Hz low frequency adjustment is used.

[0055] 4. Drug intervention module: trigger the microinjection pump through pulse signal to infuse drugs.

[0056] 4.1 Drug intervention mechanism:

[0057] The microinjection pump was triggered once, and the dosage was calculated as body weight (kg) × 0.1 mg / kg. The dosing interval was ≥ 15 minutes to avoid drug accumulation.

[0058] 4.2 Feedback termination conditions:

[0059] When it is detected that ξ<threshold and the frequency band power returns to the baseline level, electrical stimulation and drug administration are stopped and the next monitoring cycle begins.

[0060] The specific hardware deployment, software operation and system parameters of this system are as follows:

[0061] System hardware deployment:

[0062] Electrode implantation: Stereotactic surgery is used to implant electrodes into the target brain area (e.g., hippocampal coordinates of mice: AP -2.0 mm, ML 1.5 mm, DV -1.5 mm, based on the MNI coordinate system), and the position is confirmed by CT / MRI after surgery.

[0063] Signal processing: An embedded processor (e.g., TIOMAP 3730) transmits data to MATLAB (2022a) on the PC in real time, and synchronously acquires the EEG signals recorded by LabChart 8 (sampling rate 1 kHz) through the labchart.data interface.

[0064] Software operation process:

[0065] Data preprocessing: The data were segmented into 5-minute windows, the data within 1 minute before and after the onset of the attack were excluded, and baseline correction was performed (the mean of the first 10 minutes was subtracted).

[0066] Feature calculation:

[0067] Call the PACmeg function to calculate the PAC value (parameters: phase_freqs = 2:2:30, amp_freqs = 50:2:100, surr_N = 200);

[0068] Calculate LZ76 complexity (binarization threshold = signal mean) using the calc_lz_complexity function.

[0069] The pwelch function is used to calculate the frequency band power (window length 1024 points, Hanning window).

[0070] Forecasting and Control:

[0071] The feature vector is input into the FCNN model, which outputs the ξ value. The interface displays the curve and warning in real time (the red warning light is activated when ξ ≥ the threshold);

[0072] Electrical stimulation and drug intervention were triggered via a TTL interface, and events were synchronously marked in LabChart (time stamp accuracy ≤ 100 ms).

[0073] Key parameter table:

[0074]

[0075]

[0076] In summary, the closed-loop deep brain regulation system provided by the present invention, by implanting multi-channel electrodes in the cerebral cortex or hippocampus, collects EEG signals in real time and extracts 33-dimensional feature vectors including phase-amplitude coupling (PAC), signal complexity (LZ76) and frequency band power, and uses a 7-layer fully connected neural network to construct an electrophysiological prediction model. When the seizure-promoting period is detected (prediction probability ξ ≥ individualized threshold), electrical stimulation or precise drug intervention with dynamically adjustable frequency and intensity is automatically triggered to form a "monitoring-analysis-regulation" closed loop. The system optimizes the model through 4-fold cross-validation, and the detection accuracy rate based on auROC evaluation is shown to be over 85%, providing a precise treatment plan for drug-resistant epilepsy.

[0077] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the scope of protection of the present invention in any form. All technical solutions obtained by equivalent substitution, etc., fall within the scope of protection of the present invention. Parts not covered by the present invention are the same as the existing technology or can be implemented using existing technology.

Claims

1. A closed-loop deep brain control system for epilepsy treatment, characterized in that: The system includes: EEG signal acquisition module: By implanting recording electrodes in the cerebral cortex or hippocampus, neural electrophysiological signals are collected in real time; Real-time EEG signal analysis module: Based on a 33-dimensional feature vector, it outputs the predicted probability of the seizure-promoting period ξ through a 7-layer fully connected neural network to achieve epileptic seizure warning; Electrical stimulation control module: When ξ ≥ threshold, electrical stimulation is triggered. The frequency band to be regulated is filtered out, the electrical stimulation intensity is calculated and predicted, and then the output pulse stimulation is corrected, thereby performing personalized electrical stimulation control for specific frequency bands. Continuous monitoring is carried out until the frequency returns to normal and the electrical stimulation ends. Drug intervention module: triggers the microinjection pump to infuse drugs through pulse signals.

2. A closed-loop deep brain control system for epilepsy treatment according to claim 1, characterized in that: When collecting neuroelectrophysiological signals, the sampling rate is ≥1kHz.

3. A closed-loop deep brain control system for epilepsy treatment according to claim 1, characterized in that: The real-time EEG signal analysis module includes a fully connected neural network established for pre-ictal state detection. The fully connected neural network includes an input layer, a hidden layer, and an output layer. The input layer inputs a 33×1-dimensional real-valued EEG feature vector. The hidden layer includes 7 fully connected layers, which simplify the feature vector into a single real value, called ξ; the output layer outputs the ξ value, ξ∈[0,1].

4. A closed-loop deep brain control system for epilepsy treatment according to claim 3, characterized in that: In the neural network training, except for epileptic seizures, the pre-ictal state or interictal state was manually marked every 5 minutes, and the average feature vector of the first 5 minutes of training was calculated; 80% of the data was randomly extracted for 4-fold training and cross-validation, and the remaining 20% ​​was used for model accuracy verification.

5. A closed-loop deep brain control system for epilepsy treatment according to claim 1, characterized in that: The feature vector includes 26-dimensional phase amplitude coupling values, 1-dimensional complexity of the EEG signal, and 6-dimensional frequency band power.

6. A closed-loop deep brain control system for epilepsy treatment according to claim 5, characterized in that: The calculation of the phase-amplitude coupling value requires dividing the high-frequency band into intervals and using Hilbert transform to extract the phase-amplitude coupling strength. Features 1 to 26 are the average PAC values ​​of each frequency in the high-frequency band.

7. A closed-loop deep brain control system for epilepsy treatment according to claim 6, characterized in that: The high frequency band is 50–100 Hz, which is divided into 26 frequency bands at intervals of 2 Hz. The Hilbert transform is used to calculate the phase-locking value of the amplitude of each high frequency band and the phase of the low frequency band, and the average value of each frequency in the high frequency band within 5 minutes is taken as features 1–26.

8. A closed-loop deep brain control system for epilepsy treatment according to claim 5, characterized in that: The complexity of the electroencephalogram signal is calculated using a Lempel-Ziv binary sequence complexity measurement algorithm to obtain a characteristic value.

9. A closed-loop deep brain control system for epilepsy treatment according to claim 5, characterized in that: The power of each frequency band is obtained by applying the Welch method to the electroencephalogram and averaging the PSD within each frequency band.

10. A closed-loop deep brain control system for epilepsy treatment according to claim 1, characterized in that: In electrical stimulation regulation, if electrical stimulation is triggered, the initial pulse intensity is 1 mA, and then increases by 0.5 mA every 20 seconds until ξ decreases by ≥0.05 or reaches the maximum intensity of 5 mA; in drug intervention, the single dose is calculated based on body weight, with a metering range of 0.1-10 mg, and the dosing interval is ≥15 minutes.

Images