A depression neuromodulation method and system based on TMS-EEG and machine learning driving

Abstract

The application relates to the technical field of electroencephalogram rehabilitation, and particularly discloses a depression neural regulation method and system based on TMS-EEG and machine learning driving, which comprises the following steps: S1: collecting electroencephalogram signals before transcranial magnetic stimulation of a patient; S2: pre-processing the electroencephalogram signals and extracting biomarker features; S3: inputting the extracted biomarker features into a trained machine learning model to obtain decision reference data; S4: using the above parameters as auxiliary reference data to adjust the transcranial magnetic stimulation control parameters; and S5: collecting the electroencephalogram signals again after the transcranial magnetic stimulation and using the electroencephalogram signals for optimizing subsequent personalized transcranial magnetic stimulation parameter suggestions. The application converts multi-time-point electroencephalogram signals into quantitative biomarkers, thereby providing doctors with objective decision reference data throughout the whole diagnosis and treatment process. This enables doctors to predict the patient's reactivity before stimulation regulation, thereby avoiding invalid stimulation; and the doctors can dynamically optimize the parameters according to the neurophysiological feedback during the stimulation, thereby realizing personalized and accurate regulation.

Description

Technical Field

[0001] This invention relates to the field of brainwave rehabilitation technology, and specifically discloses a neuromodulation method and system for depression based on TMS-EEG and machine learning. Background Technology

[0002] Transcranial magnetic stimulation (TMS), as a non-invasive neuromodulation technique, has been proven to be an effective means of alleviating depression. Repetitive transcranial magnetic stimulation (rTMS) can improve depressive symptoms by modulating neural activity in specific brain regions (such as the dorsolateral prefrontal cortex). Currently, clinical practice of rTMS stimulation usually relies on standardized treatment parameters (such as stimulation target, frequency, and intensity) and a efficacy evaluation system based primarily on subjective clinical scales (such as the Human Impact Assessment Model), which has many limitations, including subjective diagnosis, homogenized treatment parameters, delayed efficacy evaluation, and a disconnect between diagnosis, treatment, and evaluation.

[0003] To improve the precision of treatment, existing technologies have been explored in various ways. For example, some approaches use multimodal data (such as EEG and eye movements in VR scenarios) for the auxiliary diagnosis of depression, or utilize MRI images to achieve precise localization of individualized TMS stimulation targets. Other studies focus on state-dependent stimulation, attempting to trigger TMS pulses through real-time EEG phase prediction to improve the spatiotemporal accuracy of single stimulation. However, these technological solutions mainly focus on diagnostic assistance or local optimization of stimulation execution, and have not yet systematically addressed the core clinical decision-making challenge of how to prospectively predict efficacy, personalize parameters, and adaptively adjust treatment based on the patient's individual dynamic physiological response throughout the entire treatment cycle.

[0004] Therefore, it is necessary to propose a method and system that can deeply integrate neurophysiological monitoring and intelligent data analysis, and transform the multi-time-point EEG signals generated during stimulation into quantitative indicators that can be used for clinical decision-making. This would empower doctors to shift from experience-based judgment to data-assisted decision-making, and improve the overall accuracy and efficiency of rTMS stimulation for depression. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a neuromodulation system and method for depression based on TMS-EEG and machine learning. By converting multi-timepoint EEG signals acquired through TMS-EEG into quantitative biomarkers such as the prefrontal alpha wave asymmetry index, theta wave power, and N100 amplitude change rate, it provides physicians with objective decision-making reference data throughout the entire diagnostic and treatment process. This allows physicians to predict patient responsiveness before transcranial magnetic stimulation (rTMS), avoiding ineffective stimulation; dynamically optimize parameters based on neurophysiological feedback during stimulation, achieving personalized and precise control; and ultimately form an intelligent auxiliary closed loop driven by objective data for clinical decision-making, systematically improving the accuracy, foresight, and individualization of rTMS stimulation for depression.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: A neuromodulation method for depression based on TMS-EEG and machine learning, comprising the following steps:

[0007] S1: Collect resting-state EEG and / or EEG signals induced by transcranial magnetic stimulation (TMS) before TMS.

[0008] S2: Preprocess the EEG signal and extract at least two biomarker features from the prefrontal alpha wave asymmetry index, prefrontal theta wave power, and N100 component amplitude induced by transcranial magnetic stimulation.

[0009] S3: Input the extracted biomarker features into the trained machine learning model to obtain decision reference data including the probability of depression risk, the probability of stimulus response, and personalized transcranial magnetic stimulation parameter suggestions. The machine learning model includes logistic regression and random forest algorithms.

[0010] S4: Use the personalized transcranial magnetic stimulation parameter suggestions generated in step S3 as an auxiliary reference to adjust the transcranial magnetic stimulation control parameters;

[0011] S5: After transcranial magnetic stimulation, EEG signals are collected again to calculate the acute changes in biomarker characteristics and feed them back to the model to optimize subsequent personalized transcranial magnetic stimulation parameter recommendations.

[0012] The fundamental principle of this technical solution lies in constructing an adaptive closed loop of perception-decision-execution-optimization driven by objective neurophysiological data, centered on machine learning intelligent analysis, and aiming to achieve personalized and precise regulation. Based on the stable correlation between specific EEG characteristics and the pathophysiological state of depression and neural plasticity response in neuroscience, this correlation is transformed into actionable clinical decision support through an engineered data process.

[0013] By combining transcranial magnetic stimulation (TMS) with electroencephalography (EEG), high-dimensional electrophysiological signals reflecting the brain's intrinsic state (resting state) and immediate response to external interventions (evoked state) are simultaneously acquired at key TMS nodes (pre-, mid-, and post-stimulation). This breaks through the limitations of traditional diagnosis and treatment relying on subjective statements and delayed scales, establishing an objective data perception layer throughout the entire process. Furthermore, signal processing extracts core biomarkers with clear neurophysiological significance from the raw data, such as prefrontal alpha wave asymmetry (associated with lateralization of emotional processing), prefrontal theta wave power (reflecting cognitive control and abnormal activity), and TMS-induced N100 amplitude (characterizing cortical inhibitory function and excitation-inhibition balance). These features constitute a quantitative bridge connecting brain function and clinical phenotype.

[0014] Building upon this foundation, a trained machine learning model is introduced as the intelligent decision-making hub. By learning the complex mapping relationships between these biomarker patterns and final clinical outcomes (such as diagnostic classification and treatment response) from a large amount of historical data, the model gains the ability to transform the individual characteristics of new patients into quantitative predictions and recommendations in real time. Its output decision reference data includes risk probabilities, responsiveness predictions, and parameter suggestions. It does not replace physician judgment but rather provides a deep fusion analysis based on population data patterns and individual physiological characteristics, thereby upgrading clinical decision-making from experience-based judgment to data-assisted decision-making.

[0015] Subsequently, by feeding back the EEG response (i.e., acute changes in biomarkers, such as ΔN100) following transcranial magnetic stimulation (TMS) to the decision-making model in real time, the system achieved closed-loop optimization. This feedback mechanism allows the stimulation protocol to be dynamically and directionally adjusted based on the brain's actual physiological response to the previous stimulation. The entire principle thus forms a self-itergencing reinforcement loop: each TMS execution generates new physiological response data, which, after analysis, is used to optimize the next TMS decision, thereby achieving continuous adaptation of TMS parameters to the dynamic changes in the patient's brain and driving TMS towards the expected goal.

[0016] Furthermore, in step S2, the prefrontal α-wave asymmetry index is calculated using the following function:

[0017] ,

[0018] in, and These represent the average power of the electrodes in the right and left prefrontal cortex in the α band, respectively.

[0019] Furthermore, in step S5, the acute change includes the percentage change in the N100 amplitude, denoted as ΔN100, which is calculated using the following function:

[0020] ,

[0021] in, and The N100 average amplitudes were measured before and after transcranial magnetic stimulation, respectively.

[0022] Furthermore, in step S5, the rules for dynamically adjusting parameters based on ΔN100 include: if ΔN100 exceeds the first positive threshold by 10%, it is recommended to maintain the current parameters; if ΔN100 is lower than the first negative threshold by -5%, it is recommended to adjust the stimulation target or scheme; if ΔN100 is between the first negative threshold and the first positive threshold, it is recommended to adjust the stimulation intensity.

[0023] Furthermore, in step S3, the logistic regression model will output the probability of depression risk and the probability of stimulus response P, which are between 0 and 1. depress and P response When P depress A P value of 0.65 or higher is considered a high risk of depression; when P < 0.65, the risk of depression is considered high. response A value below 0.5 indicates a potentially poor response.

[0024] Furthermore, in step S3, the random forest algorithm model will directly output the classification labels of high responders or poor responders.

[0025] A neuromodulation system for depression based on TMS-EEG and machine learning-driven methods for depression includes:

[0026] The EEG signal acquisition module is used to acquire the patient's EEG signals before and after transcranial magnetic stimulation;

[0027] The feature extraction module is used to preprocess the EEG signal and extract biomarker features;

[0028] The machine learning decision module internally stores trained machine learning models, which are used to generate decision reference data including the probability of depression risk, the probability of stimulus response, and personalized parameter suggestions based on the biomarker characteristics.

[0029] The stimulation control module, connected to the machine learning decision module, is used to generate control commands for the transcranial magnetic stimulation device based on personalized parameter suggestions.

[0030] Furthermore, the machine learning decision module also includes a parameter adjustment rule setter, which dynamically optimizes the next personalized transcranial magnetic stimulation parameter recommendations based on acute changes and adjustment rules.

[0031] Furthermore, biomarker features used for predicting stimulus response probability include the amplitude and / or latency of the N100 component extracted from the EEG response induced by transcranial magnetic stimulation.

[0032] Furthermore, the system is configured to cyclically execute steps S1 to S5 of the neuromodulation method for depression.

[0033] The beneficial effects of this invention are: (1) By extracting and analyzing quantifiable neurophysiological biomarkers such as prefrontal alpha wave asymmetry and theta wave power, objective biological evidence is provided for the identification of depression. This effectively reduces the reliance on subjective scales and personal experience in diagnosis and improves the consistency and scientific nature of diagnosis.

[0034] (2) Before the stimulation modulation begins, by using a machine learning-based model to analyze the characteristics such as the N100 amplitude induced by TMS, it is possible to predict the patient's possible response to transcranial magnetic stimulation (high responders / potentially poor responders). This provides valuable decision-making reference data for clinicians, enabling them to identify potentially ineffective patients in advance, consider alternative solutions, thereby avoiding waste of medical resources and saving patients' time.

[0035] (3) The system can intelligently generate personalized initial TMS stimulation parameter suggestions (such as stimulation intensity and frequency) based on the baseline EEG biomarker characteristics of individual patients, breaking the traditional one-size-fits-all fixed parameter mode, so that the stimulation regulation is as close as possible to the patient's unique brain functional state from the beginning.

[0036] (4) By calculating the acute changes in characteristics such as N100 amplitude after each stimulation and feeding them back to the decision-making model, the system can automatically assess the brain's physiological response to the stimulation and dynamically optimize the parameter recommendations for the next personalized transcranial magnetic stimulation. This forms an online adaptive closed loop of stimulation-assessment-optimization-restimulation, enabling the stimulation control plan to be continuously and accurately adjusted according to the dynamic changes in the patient's brain state.

[0037] (5) The core output of the entire solution is decision reference data that serves clinicians and integrates multidimensional predictions and suggestions. This is not a replacement for doctors, but rather a significant enhancement of doctors' information base at key decision points (such as triage, initiation of treatment, and parameter adjustment) by providing objective and quantitative neurophysiological insights, reducing the reliance on pure experience in decision-making, and improving the intelligence and standardization of the diagnosis and treatment process.

[0038] (6) The continuous operation of the system will accumulate a large amount of structured data that correlates precise stimulation parameters, multidimensional biomarkers and final efficacy. These data assets can be used to continuously iterate and optimize the built-in machine learning model, forming a virtuous cycle that becomes smarter with use, and providing valuable resources for clinical research. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the TMS-EEG and machine learning-driven neuromodulation system for depression in Embodiment 1 of the present invention.

[0040] Figure 2 This is a flowchart of the operation of the TMS-EEG and machine learning-driven neuromodulation system for depression in Embodiment 1 of the present invention.

[0041] Figure 3 This is a flowchart of EEG signal data acquisition in Embodiment 1 of the present invention.

[0042] Figure 4 This is a flowchart of the EEG signal data preprocessing process in Embodiment 1 of the present invention.

[0043] Figure 5 This is a flowchart of the feature extraction process in Embodiment 1 of the present invention.

[0044] Figure 6 This is a schematic diagram of machine learning model training in Embodiment 1 of the present invention.

[0045] Figure 7 This is a flowchart of the TMS stimulation optimization process in Embodiment 1 of the present invention. Detailed Implementation

[0046] The specific implementation method is described below with reference to the accompanying drawings.

[0047] Example 1

[0048] Basic as Figures 1-7 As shown: This embodiment provides the process of implementing neuromodulation stimulation on high-risk patients with depression using a TMS-EEG and machine learning-driven neuromodulation system for depression.

[0049] For a 28-year-old newly diagnosed patient presenting with long-term low mood, this TMS-EEG and machine learning-driven neuromodulation system for depression was used. First, the EEG signal acquisition module was activated. This module was configured with a 19-channel EEG cap according to the standard placement of 10-20 electrodes, with a sampling rate of 256Hz and bilateral ear reference electrodes. The patient's resting-state EEG signal was acquired for 8 minutes before stimulation, and the EEG signal induced by single-pulse TMS stimulation (targeting the left dorsolateral prefrontal cortex, with an intensity of 110% of the resting motor threshold) was also acquired. The resting-state EEG covered key areas of the whole brain, including the frontal lobe, temporal lobe, and parietal lobe. The TMS-induced EEG synchronously recorded the EEG response from 0 to 500 ms after stimulation, providing a basic data source for subsequent biomarker extraction. Its core function is to obtain objective electrophysiological representations of the patient's brain physiological state, breaking the reliance of traditional diagnosis on subjective scales.

[0050] After data acquisition, the feature extraction module establishes a communication connection with the EEG signal acquisition module and initiates a preprocessing procedure on the raw EEG signal. First, baseline drift and power frequency interference are eliminated through a 0.5Hz high-pass filter and a 50Hz notch filter. Then, artifacts related to electrooculography (EOG) and electromyography (EMG) are removed using bad segment removal and the ICA algorithm to ensure signal purity. After preprocessing, this module begins to extract core biomarker features. On one hand, for resting-state EEG, the natural logarithmic difference of the power in the alpha band (8-13Hz) of the F3 and F4 electrodes is calculated to obtain the prefrontal alpha wave asymmetry index, whose functional expression is:

[0051] ,

[0052] Simultaneously, the average power of the theta band (4-7Hz) of the Fz / Cz electrodes is extracted and Z-score normalization is performed. On the other hand, for TMS-induced EEG, the average amplitude of the N100 component is extracted by locking the 80-120ms time window after stimulation. Finally, the above features are integrated into a feature vector. The role of this module is to realize the transformation from raw signals to quantifiable biomarkers. These biomarkers are the core basis for subsequent diagnosis and efficacy prediction. For example, the prefrontal alpha wave asymmetry index can reflect abnormal lateralization of emotion processing, the increased power of the prefrontal theta waves is a typical manifestation of abnormal brain function in depression, and the N100 amplitude is directly related to the intensity of cortical inhibitory neuron activity. The three together constitute the physiological basis for objective diagnosis and treatment.

[0053] The feature vectors are then transmitted to the machine learning decision module, which pre-stores a depression identification model and a treatment prediction model trained with 10-fold cross-validation. First, the prefrontal alpha wave asymmetry index and the prefrontal theta wave power are input into the depression identification model, whose decision function is:

[0054] ,

[0055] In the formula, This represents the patient's probability of having depression. For the prefrontal cortex Wave power.

[0056] The model output a probability of 0.72 for the patient's depressive state, which is higher than the judgment threshold of 0.65, serving as a key decision reference data. Next, the N100 amplitude (-2.9 μV), patient disease duration (8 months), and prefrontal θ / α power ratio were input into the efficacy prediction model, whose decision function is:

[0057] ,

[0058] In the formula, This represents the probability that a patient will respond effectively to TMS stimulation. The mean amplitude of the N100 component in the EEG response induced by transcranial magnetic stimulation. For the prefrontal cortex Wave power and The ratio of wave power, i.e. ratio, This refers to the patient's actual age.

[0059] The model outputs a stimulus response probability of 0.85 and generates initial TMS regulation parameter suggestions based on baseline biomarker characteristics: stimulation frequency 10Hz, intensity 100% of resting motion threshold, and daily stimulation duration 20 minutes. The core function of this module is to convert biomarkers into quantified probability values ​​and parameter suggestions, forming a decision reference dataset for clinicians to use, thereby supporting precise and personalized transcranial magnetic stimulation decisions.

[0060] After receiving parameter suggestions generated by the machine learning decision-making module, the stimulation control module translates them into specific device control commands. After the clinician confirms the data (high risk of depression, high response probability, and specific parameter suggestions), the TMS device performs the first rTMS stimulation. Within 30 minutes after the stimulation, the EEG signal acquisition module again acquires single-pulse TMS-evoked EEG data. The feature extraction module calculates the N100 amplitude after stimulation to be -3.2 μV, and then uses a function...

[0061] ,

[0062] The acute change was calculated to be 10.3%. This change was fed back to the parameter adjustment rule generator in the machine learning decision module. Because ΔN100 exceeded the first positive threshold of 10%, the rule generator determined a positive response and generated a subsequent stimulus control suggestion to "maintain the current parameters." The stimulus control module then executed subsequent treatment accordingly. Throughout the treatment cycle, the system repeated the above process of data collection, processing, decision-making, and control. The machine learning decision module also generated long-term efficacy trend prediction data based on the ΔN100 sequence data. Ultimately, after completing 4 weeks of stimulus control, the patient's depression scale score decreased by 52%, achieving precise and effective closed-loop control.

[0063] Example 2

[0064] This embodiment provides dynamic parameter optimization and efficacy early warning for patients with poor initial response to stimulation modulation of a neuromodulation system for depression based on TMS-EEG and machine learning.

[0065] A 35-year-old patient diagnosed with depression, whose mood did not improve significantly after two TMS stimulation sessions, was connected to this neuromodulation system for stimulation modulation program optimization. All units of the system started working according to collaborative logic. First, before the third stimulation, the EEG signal acquisition module simultaneously acquired the patient's resting-state EEG and TMS-evoked EEG, with sampling parameters consistent with Example 1. Its function is to obtain dynamic data of the brain's physiological state during the stimulation modulation phase. Compared to the baseline data before stimulation, changes in EEG characteristics during stimulation directly reflect the stimulation response, providing crucial evidence for parameter optimization.

[0066] After preprocessing the acquired signals, the feature extraction module focuses on extracting the dynamic changes of biomarkers. First, it calculates that the current prefrontal alpha wave asymmetry index has improved by 12% compared to the baseline, and the prefrontal theta wave power has decreased by 8%. Then, it extracts the TMS-induced N100 amplitude as -2.7 μV, compared to -2.8 μV before the first stimulation. The acute change calculated using the ΔN100 function is -3.6%, falling within the range of -5% to 10%. Here, the dynamic changes of biomarkers become the core input parameters for generating optimization suggestions. For example, a negative change in N100 amplitude indicates that the activity of cortical inhibitory neurons has not been regulated as expected, directly suggesting insufficient fit of the current stimulation regulation parameters. Conversely, the slight improvement in the prefrontal alpha and theta wave characteristics indicates a potential trend of stimulation regulation taking effect.

[0067] After receiving feature data, the machine learning decision-making module first inputs baseline features and the previous two ΔN100 data points into a long-term efficacy monitoring model (random forest model). The model outputs a predicted probability of poor efficacy of 0.68, and the system immediately generates efficacy warning information as high-level decision-making reference data. Simultaneously, the current ΔN100 data is input into the parameter adjustment rule generator. Because the change is between the first negative threshold (-5%) and the first positive threshold (10%), the rule generator generates a parameter optimization suggestion of "increasing the stimulation intensity by 5%" according to preset rules. In conjunction with the current change in the theta wave power of the prefrontal cortex, it suggests fine-tuning the stimulation frequency to 12Hz. By generating warnings and specific parameter optimization suggestions, this module provides doctors with clear, quantitative evidence based on changes in objective biomarkers for adjusting stimulation protocols.

[0068] Based on the aforementioned optimization suggestions, the stimulation control module generated new control instructions. After confirmation by the physician, the TMS device executed the third stimulation modulation with the new parameters (stimulation frequency 12Hz, intensity 105% of the resting motor threshold). Within 30 minutes after stimulation, the EEG signal acquisition module again acquired TMS-evoked EEG data. The feature extraction module calculated ΔN100 to be 15.2%, indicating that cortical inhibitory neuron activity was effectively activated. In subsequent stimulations, the system continuously repeated the closed-loop process. The machine learning decision-making module dynamically generated suggestions for fine-tuning parameters based on changes in biomarkers each time. After the sixth stimulation, the long-term efficacy monitoring model's predicted probability of good efficacy increased to 0.82, and the warning was lifted. Ultimately, after completing 6 weeks of stimulation, the patient's depression scale score decreased by 48%, achieving a reversal from poor efficacy to effective stimulation modulation. This fully demonstrates the core driving role of biomarker-driven decision reference data in dynamic closed-loop modulation, as well as the value of the synergistic effect of various system units in ensuring individualized stimulation rehabilitation.

Claims

1. A neuromodulation system for depression based on TMS-EEG and machine learning methods, characterized in that, include: The EEG signal acquisition module is used to acquire the patient's EEG signals before and after transcranial magnetic stimulation therapy; The feature extraction module is used to preprocess the EEG signal and extract biomarker features. These biomarker features include at least two of the following: the prefrontal alpha wave asymmetry index, the prefrontal theta wave power, and the N100 component amplitude induced by transcranial magnetic stimulation. The prefrontal alpha wave asymmetry index is calculated using the following function: ， in, and These represent the average power of the electrodes in the right and left prefrontal cortex in the α band, respectively. The machine learning decision-making module internally stores trained machine learning models used to generate decision reference data based on the biomarker characteristics. This data includes the probability of depression risk, the predicted probability of treatment response, and personalized parameter suggestions. The machine learning models include logistic regression and random forest algorithms. The logistic regression model outputs the probability of depression risk and the probability of treatment response P, both ranging from 0 to 1. depress and P response When P depress A P value of 0.65 or higher is considered a high risk of depression; when P < 0.65, the risk of depression is considered high. response A score below 0.5 indicates a potentially poor responder; the Random Forest algorithm model will directly output the classification labels for either high or poor responders. The treatment control module, connected to the machine learning decision module, is used to generate control commands for the transcranial magnetic stimulation device based on personalized parameter suggestions. The feature extraction module is also used to calculate the acute change in the biomarker features in the EEG signal collected again after treatment, and to feed the acute change back to the machine learning decision module. The machine learning decision module is also configured to dynamically optimize personalized transcranial magnetic stimulation parameter recommendations for the next treatment based on the feedback of the acute changes. This system is configured to cyclically execute the neuromodulation process for depression.

2. The neuromodulation system for depression according to claim 1, characterized in that, The machine learning decision module also includes a parameter tuning rule generator, which dynamically optimizes personalized transcranial magnetic stimulation parameter recommendations for the next treatment based on acute changes and tuning rules. The acute change includes the percentage change in the N100 amplitude, ΔN100. The adjustment rules include rules for dynamically adjusting personalized transcranial magnetic stimulation parameters based on ΔN100: if ΔN100 exceeds the first positive threshold by 10%, it is recommended to maintain the current parameters; if ΔN100 is lower than the first negative threshold by -5%, it is recommended to adjust the stimulation target or protocol; if ΔN100 is between the first negative threshold and the first positive threshold, it is recommended to adjust the stimulation intensity.

3. The neuromodulation system for depression according to claim 2, characterized in that, Biomarker features used for predicting therapeutic responsiveness include the amplitude and / or latency of the N100 component extracted from the EEG response evoked by transcranial magnetic stimulation, and the percentage change in N100 amplitude, denoted as ΔN100, calculated using the following function: ， in, and The mean amplitudes of N100 measured before and after treatment are shown respectively.

Images