Alzheimer's disease closed-loop adaptive neuromodulation method and system

Abstract

The application discloses an Alzheimer's disease closed-loop adaptive neuromodulation method and system, which solves the problems of existing system, such as limited sensing modalities, lack of efficacy prediction, simple control strategy, missing safety constraints and insufficient small sample adaptability. The method comprises the following steps: collecting a brain oxygen signal; constructing a functional connectivity matrix and extracting a feature vector; inputting a machine learning model to output a transcranial alternating current stimulation response probability, and dividing subjects into high, medium and low response groups; configuring stimulation parameters according to the prediction results; collecting brain oxygen signals in real time and extracting a functional connectivity remodeling index; and adjusting the stimulation parameters in real time by using a reinforcement learning algorithm. The application uses functional near-infrared spectroscopy to replace electroencephalogram signals as a feedback source, combines individualized efficacy prediction before treatment with reinforcement learning closed-loop control, and realizes individualization and adaptive optimization of Alzheimer's disease neuromodulation.

Description

Technical Field

[0001] This application relates to the field of neuromodulation technology, and more particularly to a closed-loop neuromodulation technology for the treatment of Alzheimer's disease. Background Technology

[0002] Alzheimer's disease (AD) is a neurodegenerative disease characterized by progressive cognitive decline. Current drug treatments (such as cholinesterase inhibitors and NMDA receptor antagonists) can only slow symptom progression and cannot reverse neurodegenerative changes. Transcranial alternating current stimulation (tACS), as a non-invasive neuromodulation technique, induces neural oscillations by applying sinusoidal currents of specific frequencies, showing promising clinical applications in AD treatment. Functional near-infrared spectroscopy (fNIRS), a non-invasive brain imaging technique, indirectly reflects neural activity by measuring blood oxygen dynamics. Its low sensitivity to motor artifacts and good ecological validity give it significant advantages in clinical settings. The resting-state fNIRS functional connectivity pattern in AD patients is closely related to the degree of cognitive impairment.

[0003] However, existing tACS neuromodulation systems have the following technical shortcomings: First, existing closed-loop neuromodulation systems generally use electroencephalogram (EEG) signals as feedback signal sources. EEG has low spatial resolution, making it difficult to accurately characterize the topological characteristics of functional connectivity networks across multiple brain regions, and thus failing to effectively capture characteristic cross-brain region functional connectivity abnormalities in AD patients. Second, existing systems are all passive feedback adjustments after stimulation, without establishing a mechanism for predicting therapeutic responses based on individual brain functional characteristics before stimulation, making it impossible to achieve individualized initial parameter configuration, resulting in approximately 30% to 40% of patients showing no significant response to standard tACS protocols. Third, existing systems mostly use rule-driven or simple feedback control, lacking the ability to sequentially optimize decisions for long-term treatment and struggling to handle nonlinear, high-dimensional brain state spaces. Fourth, existing systems lack a systematic safety constraint framework and have not established a multi-layered safety assurance mechanism for clinical applications. Fifth, medical AI model training is usually limited by sample size, and existing solutions lack effective solutions for small-sample clinical scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a closed-loop adaptive neuromodulation method and system for Alzheimer's disease, addressing the following technical problems in existing technologies: existing closed-loop neuromodulation systems use EEG signals as feedback signal sources, resulting in low spatial resolution and an inability to accurately characterize the topological characteristics of multi-brain inter-region functional connectivity networks; existing systems lack a pre-stimulation efficacy response prediction mechanism based on individual brain functional characteristics, making it impossible to achieve individualized initial parameter configuration; existing systems have simple control strategies, lack sequential decision optimization capabilities for long-term treatment, and struggle to handle nonlinear high-dimensional brain state spaces; existing systems lack a systematic safety constraint framework and a multi-layered safety assurance mechanism for clinical applications; and existing solutions lack effective solutions for small-sample clinical scenarios.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a closed-loop adaptive neuromodulation method for Alzheimer's disease, comprising the following steps: Step S1 is baseline functional near-infrared spectroscopy signal acquisition, acquiring cerebral blood oxygenation signals of subjects under resting conditions for 5 to 15 minutes, covering the prefrontal cortex, temporal cortex, temporoparietal cortex, and occipital cortex, which are the main areas of functional connectivity abnormalities in Alzheimer's disease; Step S2 is multi-brain region functional connectivity matrix extraction, calculating functional connectivity correlation coefficients between brain regions using at least one of Pearson correlation coefficient, partial correlation coefficient, or wavelet coherence, constructing a functional connectivity matrix, and extracting functional connectivity feature vectors, which can comprehensively characterize the collaborative activity patterns of multiple brain regions; Step S3 is transcranial alternating current stimulation efficacy prediction based on functional connectivity features, inputting the functional connectivity feature vectors into a machine learning-based classification model, outputting transcranial alternating current stimulation response probabilities, and classifying subjects according to the response probabilities. For high-response, medium-response, and low-response groups, individualized efficacy prediction is achieved before treatment. Step S4 involves configuring initial parameters for transcranial alternating current stimulation (TAC) based on the efficacy prediction results. These initial parameters include stimulation frequency, stimulation intensity, stimulation target, and stimulation waveform. The stimulation frequency for the high-response group is set to the 40 Hz gamma band, which can effectively promote amyloid protein clearance. Step S5 involves real-time functional near-infrared spectroscopy monitoring. During TAC stimulation, cerebral blood oxygenation signals are collected in real time, and the functional connectivity remodeling index is extracted. The functional connectivity remodeling index characterizes the rate of change in functional connectivity strength during treatment, providing an objective quantitative indicator for closed-loop regulation. Step S6 involves closed-loop parameter adjustment based on reinforcement learning. According to the functional connectivity remodeling index and the current stimulation parameters, a reinforcement learning algorithm is used to adjust the TAC stimulation parameters in real time. When the rate of change of the functional connectivity remodeling index is lower than a preset threshold, parameter adjustment is triggered to improve functional connectivity strength.

[0006] This invention overcomes the limitation of low spatial resolution in electroencephalogram (EEG) signals by utilizing functional near-infrared spectroscopy, enabling precise measurement of blood oxygenation dynamics signals in multiple brain regions and thus constructing a high-precision functional connectivity network. By extracting functional connectivity features and predicting efficacy before treatment, individualized stimulation parameters can be configured for patients with different response types, avoiding the poor efficacy problems caused by using uniform parameters. During treatment, by monitoring the functional connectivity remodeling index in real time, the reinforcement learning algorithm can dynamically adjust stimulation parameters according to the patient's real-time brain state, achieving sequential decision optimization. The reinforcement learning algorithm uses conservative Q-learning for offline pre-training, enabling effective parameter optimization in small sample scenarios of 30 to 80 cases, solving the problem of scarce clinical data.

[0007] Further, in step S2, the extraction of the functional connectivity matrix uses the Pearson correlation coefficient to calculate the linear correlation between brain regions, and / or uses the partial correlation coefficient to eliminate the influence of indirect connections, and / or uses wavelet coherence analysis to analyze frequency domain functional connectivity; the functional connectivity feature vector includes one or more graph theory metrics such as global efficiency, clustering coefficient, and modularity. The Pearson correlation coefficient can quickly calculate the degree of linear correlation between blood oxygenation signals of two brain regions; the partial correlation coefficient can identify direct functional connections by controlling the influence of other brain regions; wavelet coherence analysis can evaluate the strength of functional connections at different frequency scales; global efficiency reflects the information transmission efficiency of the brain network; the clustering coefficient characterizes the degree of aggregation of local networks; and modularity describes the modular organizational structure of the network. The above graph theory metrics can comprehensively characterize the topological characteristics of the brain network.

[0008] Preferably, in step S3, the machine learning-based classification model includes a support vector machine (SVM) model, a graph neural network (Graph NNN) model, and / or a Transformer model. The SVM model is used for small sample scenarios, the Graph NNN model is used to capture brain network topological features, and the Transformer model is used for sequence feature modeling. The classification model employs an ensemble learning strategy to fuse the prediction results of multiple base models. The ensemble learning strategy uses either a weighted average or a voting mechanism. The SVM model achieves robust prediction under small sample conditions by maximizing the classification margin. The Graph NNN model can directly learn the functional connectivity matrix end-to-end and retain the topological information of the brain network. The Transformer model can capture long-range dependencies in functional connectivity feature vectors through a self-attention mechanism. The weighted average assigns weights based on the performance of each model on the validation set, and the voting mechanism uses a majority vote principle to determine the final classification result. By fusing the advantages of different models, the accuracy and robustness of the prediction can be improved.

[0009] Specifically, in step S6, the state space of the reinforcement learning algorithm includes functional connectivity feature vectors, cerebral blood oxygenation dynamic features, current stimulus parameter vectors, and treatment progress; the action space of the reinforcement learning algorithm includes stimulus frequency adjustment amounts. Stimulus intensity adjustment amount Stimulation target adjustment and stimulation waveform adjustment, among which The value range is -5 to +5 Hz. The value range is -0.3 to +0.3 mA; the reward function of the reinforcement learning algorithm is... ,in , , , For preset weighting coefficients, The value range is from 0.3 to 0.5. The value range is from 0.2 to 0.4. The value range is from 0.1 to 0.3. The value range is from 0.05 to 0.15. Indicates the change in functional connections. This indicates the change in the score of the Mini-Mental State Examination. This indicates a function that indicates a security violation. The parameter adjustment range is represented by the following terms: the first term of the reward function encourages increased functional connectivity, the second term encourages improved cognitive function, the third term penalizes behaviors that violate safety constraints, and the fourth term penalizes excessive parameter adjustments to maintain treatment stability. The reinforcement learning algorithm employs a conservative Q-learning algorithm for offline pre-training and a soft actor-critic algorithm for online fine-tuning. Conservative Q-learning avoids overestimation by penalizing Q-values ​​of out-of-policy actions, while the soft actor-critic algorithm achieves a balance between exploration and exploitation by maximizing the entropy regularization objective function.

[0010] Further, in step S6, the closed-loop parameter adjustment includes a safety constraint mechanism, which comprises hard constraints and soft constraints. The hard constraints limit the stimulation intensity range to 0.5 to 2 mA, the stimulation frequency range to 1 to 100 Hz, and the duration of a single treatment session to 10 to 30 minutes. The soft constraints employ a constrained Markov decision process (CMM) framework, which includes a cost function constraint, limits the maximum magnitude of each parameter adjustment, and sets up a human-machine collaborative supervision layer. The hard constraints set absolute limits based on clinical safety guidelines; stimulation intensity exceeding 2 mA may cause discomfort or tissue damage, stimulation frequency exceeding the 1 to 100 Hz range lacks sufficient safety evidence, and excessively long single treatment sessions increase the risk of fatigue and discomfort. The constrained Markov decision process ensures that the expected value of the cost function does not exceed a preset threshold while optimizing the reward function. The human-machine collaborative supervision layer allows clinicians to manually intervene or terminate treatment when abnormalities are detected.

[0011] Preferably, in step S5, the formula for calculating the functional connectivity reshaping index is as follows: ,in Indicates the current functional connection strength. The baseline functional connectivity strength is represented. In step S6, when the functional connectivity remodeling index is greater than or equal to 10%, the current stimulation parameters are maintained. When the functional connectivity remodeling index is between -5% and 10%, the stimulation parameters are finely adjusted (stimulation intensity adjustment range is ±0.2 mA, stimulation frequency adjustment range is ±2 Hz). When the functional connectivity remodeling index is less than -5%, the stimulation intensity is first reduced to 50% of the initial value. If the functional connectivity remodeling index continues to decrease, stimulation is stopped. The functional connectivity remodeling index eliminates the influence of baseline differences between individuals through normalization. Positive values ​​indicate enhanced functional connectivity, and negative values ​​indicate weakened functional connectivity.

[0012] In some embodiments, in step S3, the machine learning-based classification model integrates EEG signal features and blood biomarker features. The EEG signal features include the power spectral density of the 40 Hz gamma band, and the blood biomarker features include at least β-amyloid protein-related indicators or phosphorylated tau protein-related indicators. The classification model calculates the weight coefficients of each modality feature through an attention mechanism and performs weighted fusion. Gamma-band EEG activity is closely related to synchronous neuronal firing and cognitive function. Alzheimer's disease patients typically exhibit reduced gamma-wave power. β-amyloid protein and phosphorylated tau protein are pathological markers of Alzheimer's disease. The attention mechanism can automatically learn the contribution of different modality features to efficacy prediction. Through multimodal fusion, complementary information from functional near-infrared spectroscopy, EEG, and blood biomarkers can be comprehensively utilized to improve the accuracy of efficacy prediction.

[0013] As a preferred embodiment, the offline pre-training method of the reinforcement learning algorithm includes a behavior cloning warm-up phase and a conservative Q-learning phase. The behavior cloning warm-up phase is executed before the conservative Q-learning phase. The behavior cloning warm-up phase initializes the algorithm parameters by imitating expert policies, while the conservative Q-learning phase achieves conservative estimation by penalizing the Q-values ​​of actions outside the policy. The offline pre-training method supports small-sample learning scenarios with 30 to 80 cases. The behavior cloning warm-up phase utilizes treatment decision data from clinical experts to train the policy network through supervised learning, enabling the algorithm to quickly learn reasonable parameter adjustment patterns and avoid inefficient exploration and potential risks caused by starting with random initialization. The conservative Q-learning phase further optimizes the policy based on the behavior cloning, preventing the algorithm from overconfidently selecting unverified and dangerous actions by assigning lower Q-values ​​to actions outside the offline dataset.

[0014] Further, in step S6, the closed-loop parameter adjustment includes a safety assessment system. This system includes a triple estimator based on ordinary differential equations to quantify model uncertainty. The triple estimator is used to estimate the confidence interval of the model prediction, detect distribution shift using the maximum mean difference with a shift threshold ranging from 0.05 to 0.15, and perform robustness testing using the Monte Carlo method. The triple estimator based on ordinary differential equations quantifies the model's prediction uncertainty for the current state-action pair by constructing multiple independent Q-value estimators and calculating their variance. When the uncertainty is too high, the system adopts a conservative strategy or requests manual intervention. The maximum mean difference method detects whether the patient's brain state deviates from the training data coverage by comparing the kernel mean embedding distance between the training data distribution and the current data distribution. The Monte Carlo method tests the algorithm's robustness under uncertainty conditions by adding random perturbations to the model parameters and input data.

[0015] The invention also provides a closed-loop adaptive neuromodulation system for Alzheimer's disease, comprising: a functional near-infrared spectroscopy acquisition module for acquiring cerebral blood oxygenation signals of subjects in a resting state and during treatment; the functional near-infrared spectroscopy acquisition module includes a multi-channel light source and a detector array covering key brain regions such as the prefrontal lobe, temporal lobe, temporoparietal lobe, and occipital lobe, with a sampling frequency of not less than 10 Hz; a functional connectivity feature extraction module, data-connected to the functional near-infrared spectroscopy acquisition module, for calculating the functional connectivity strength between brain regions and extracting functional connectivity feature vectors based on the cerebral blood oxygenation signals; the functional connectivity feature extraction module includes a signal preprocessing unit and a feature calculation unit; a therapeutic effect prediction module, data-connected to the functional connectivity feature extraction module, for predicting the transcranial alternating current stimulation response probability based on the functional connectivity feature vectors; and a transcranial alternating current stimulation parameter configuration module, data-connected to the therapeutic effect prediction module, for configuring stimulation frequency, stimulation intensity, stimulation target, and stimulation waveform based on the therapeutic effect prediction results. The system comprises: a real-time monitoring module connected to the functional near-infrared spectroscopy acquisition module for real-time acquisition of cerebral blood oxygenation signals and calculation of the functional connectivity remodeling index during treatment; a reinforcement learning closed-loop control module connected to both the real-time monitoring module and the transcranial AC stimulation parameter configuration module for real-time adjustment of transcranial AC stimulation parameters based on the functional connectivity remodeling index and current stimulation parameters; a state encoder, a policy network, a value network, and a parameter adjustment actuator; and a safety monitoring module connected to both the reinforcement learning closed-loop control module and the transcranial AC stimulation parameter configuration module for monitoring whether the stimulation parameters meet safety constraints and triggering safety interventions when necessary; the safety monitoring module includes a hard constraint check unit, a soft constraint assessment unit, and a safety intervention trigger. All modules communicate with each other via a data bus, and the entire system can be integrated into a portable device for convenient clinical application and home rehabilitation scenarios.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. By using functional near-infrared spectroscopy to replace traditional EEG signal monitoring, the spatial resolution of brain functional connectivity measurement is significantly improved. It can accurately characterize the topological characteristics of functional connectivity networks in multiple brain regions such as the prefrontal, temporal, temporoparietal, and occipital lobes of Alzheimer's patients. This overcomes the technical shortcomings of low spatial resolution of EEG signals and inability to accurately locate abnormal functional connectivity regions. As a result, the constructed functional connectivity matrix is ​​more accurate, providing high-quality input features for subsequent efficacy prediction and parameter adjustment. This solves the problem of sensor modality limitations from a technical principle perspective.

[0017] 2. By extracting functional connectivity features before transcranial alternating current stimulation (TCC) therapy and using machine learning models to predict efficacy, individualized initial parameter configuration was achieved. Based on the predicted response probability, patients were divided into high-response, medium-response, and low-response groups, and targeted stimulation frequencies, intensities, targets, and waveforms were configured for different groups. This is expected to reduce the non-response rate of patients, significantly improve the effectiveness of treatment and the efficiency of resource utilization, and solve the problem of lacking a pre-stimulation efficacy prediction mechanism from a technical perspective.

[0018] 3. By employing reinforcement learning algorithms to achieve closed-loop adaptive parameter adjustment, the technical shortcomings of existing technologies—such as simple control strategies and inability to effectively handle nonlinear high-dimensional brain state spaces—are overcome. The reinforcement learning algorithm constructs a high-dimensional state space that includes functional connectivity feature vectors, dynamic brain blood oxygenation features, current stimulus parameter vectors, and treatment progress. The reward function comprehensively considers changes in functional connectivity, improvement in cognitive function, safety constraints, and parameter adjustment stability. It can find the optimal parameter adjustment strategy under multi-objective constraints and is expected to maximize long-term efficacy through sequential decision optimization, thereby improving the cognitive function of Alzheimer's patients.

[0019] 4. By constructing a multi-layered safety constraint mechanism, including hard constraints, soft constraints, and a human-machine collaborative supervision layer, the clinical safety of transcranial alternating current stimulation therapy is comprehensively guaranteed. The hard constraints limit the stimulation intensity to no more than 2 mA, the stimulation frequency to the range of 1 to 100 Hz, and the duration of a single treatment session to no more than 30 minutes. The soft constraints adopt a constrained Markov decision process framework to limit the maximum range of parameter adjustment for each session through a cost function. The human-machine collaborative supervision layer allows clinicians to monitor in real time and intervene manually when necessary. This technical architecture solves the problem of the lack of a safety constraint mechanism and provides a reliable guarantee for clinical promotion and application.

[0020] 5. By employing an offline pre-training and online fine-tuning strategy that combines conservative Q-learning with the soft actor-critic algorithm, the model training problem in small-sample learning scenarios is effectively solved, supporting clinical application scenarios with sample sizes of 30 to 80 cases. In the behavioral cloning warm-up stage, the model parameters are quickly initialized using treatment decision data from clinical experts. Conservative Q-learning achieves conservative estimation by penalizing the Q-value of out-of-policy actions, avoiding overfitting and risky exploration under limited data conditions. This solves the problem of insufficient adaptability for small samples from a technical implementation perspective, significantly reducing the cost of clinical data collection and the technical application threshold. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall experimental design for the closed-loop adaptive neural modulation method.

[0022] Figure 2 This is a schematic diagram illustrating the calculation process of the dynamic function connection matrix.

[0023] Figure 3 This is a schematic diagram showing the changes in cognitive function over time for the real stimulus group and the sham stimulus group. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings. The present invention provides a closed-loop adaptive neuromodulation method and system for Alzheimer's disease based on functional near-infrared spectroscopy functional connectivity characteristics. Its core improvement lies in: using functional near-infrared spectroscopy instead of electroencephalogram (EEG) signals as the feedback signal source; achieving individualized parameter configuration through pre-treatment efficacy prediction; and employing a reinforcement learning algorithm to achieve closed-loop adaptive modulation, coupled with a multi-layered safety constraint mechanism. This solves the problems of limited sensing modalities, lack of efficacy prediction mechanisms, simple control strategies, and missing safety constraint mechanisms in existing technologies.

[0025] Example 1 This embodiment provides a closed-loop adaptive neuromodulation method for Alzheimer's disease based on functional near-infrared spectroscopy functional connectivity features, such as... Figure 1 As shown, the overall experimental design of this method covers the entire process of baseline signal acquisition, functional connectivity feature extraction, efficacy prediction, parameter configuration, and closed-loop regulation.

[0026] Step S1 involves baseline functional near-infrared spectroscopy signal acquisition. Under resting conditions, cerebral oxygenation signals are acquired from the subject for 5 to 15 minutes. The signal acquisition covers the prefrontal cortex, temporal cortex, temporoparietal cortex, and occipital cortex—these brain regions are the main areas of functional connectivity abnormalities in Alzheimer's disease.

[0027] Step S2 involves extracting the multi-brain region functional connectivity matrix, such as... Figure 2 As shown, Pearson correlation coefficient was used to calculate linear correlation between brain regions, partial correlation coefficient was used to eliminate the influence of indirect connections, and / or wavelet coherence analysis was used to analyze functional connectivity in the frequency domain, constructing a functional connectivity matrix and extracting functional connectivity feature vectors.

[0028] Functional connectivity feature vectors include one or more graph theory metrics such as global efficiency, clustering coefficient, and modularity. Global efficiency reflects the information transmission efficiency of the brain network, clustering coefficient characterizes the degree of aggregation of local networks, and modularity describes the modular organizational structure of the network.

[0029] Step S3 involves predicting the efficacy of transcranial alternating current stimulation (TAPS) based on functional connectivity features. The functional connectivity feature vector is input into a machine learning-based classification model, which outputs the TPS response probability. Based on this probability, subjects are divided into high-response, medium-response, and low-response groups, achieving individualized efficacy prediction before treatment.

[0030] The classification models include support vector machine (SVM) models, graph neural network (Graph NNN) models, and / or Transformer models. SVMs achieve robust predictions under small sample conditions by maximizing the classification margin; Graph NNNs directly learn the functional connectivity matrix end-to-end, preserving information about the brain network topology; Transformer models capture long-range dependencies through a self-attention mechanism.

[0031] The classification model employs an ensemble learning strategy to fuse the prediction results of multiple base models. This ensemble learning strategy uses either a weighted average or a voting mechanism. The weighted average assigns weights based on the performance of each model on the validation set, while the voting mechanism uses a majority vote to determine the final classification result.

[0032] Step S4 involves configuring the initial parameters for transcranial alternating current stimulation based on the efficacy prediction results. These initial parameters include stimulation frequency, stimulation intensity, stimulation target, and stimulation waveform. For the high-response group, the stimulation frequency is set to the 40 Hz gamma band, which effectively promotes amyloid clearance.

[0033] Step S5 involves real-time functional near-infrared spectroscopy monitoring. During transcranial alternating current stimulation, cerebral blood oxygenation signals are acquired in real time, and the functional connectivity remodeling index is extracted. The formula for calculating the functional connectivity remodeling index is as follows: ,in Indicates the current functional connection strength. Indicates the baseline functional connection strength.

[0034] The functional connectivity remodeling index eliminates the influence of baseline differences between individuals through normalization. Positive values ​​indicate enhanced functional connectivity, while negative values ​​indicate weakened functional connectivity, providing an objective quantitative indicator for closed-loop regulation.

[0035] Step S6 involves closed-loop parameter adjustment based on reinforcement learning. Based on the functional connectivity remodeling index and the current stimulation parameters, a reinforcement learning algorithm is used to adjust the transcranial alternating current stimulation parameters in real time. When the rate of change of the functional connectivity remodeling index falls below a preset threshold, parameter adjustment is triggered to enhance functional connectivity strength.

[0036] Specifically, the state space of the reinforcement learning algorithm includes a functional connectivity feature vector (128 dimensions), dynamic brain blood oxygenation features (16 dimensions, including the trend of changes in oxyhemoglobin and deoxyhemoglobin concentrations), a current stimulus parameter vector (8 dimensions), and treatment progress (1 dimension), which together constitute a complete 153-dimensional decision basis.

[0037] The action space of the reinforcement learning algorithm includes stimulus frequency adjustment. Stimulus intensity adjustment amount Stimulation target adjustment and stimulation waveform adjustment, among which The value range is -5 to +5 Hz. The value ranges from -0.3 to +0.3 mA.

[0038] The reward function of the reinforcement learning algorithm is: ,in The value range is from 0.3 to 0.5. The value range is from 0.2 to 0.4. The value range is from 0.1 to 0.3. The value range is from 0.05 to 0.15.

[0039] The reward function consists of four parts: a first term encourages increased functional connectivity, a second encourages improved cognitive function, a third penalizes violations of safety constraints, and a fourth penalizes excessive parameter adjustments to maintain treatment stability. The reinforcement learning algorithm is pre-trained offline using a conservative Q-learning algorithm and fine-tuned online using a soft actor-critic algorithm.

[0040] The parameter adjustment strategy is as follows: when the functional connectivity remodeling index is greater than or equal to 10%, the current stimulation parameters are maintained; when the functional connectivity remodeling index is between -5% and 10%, the stimulation parameters are finely adjusted, with the stimulation intensity adjusted by ±0.2 mA and the stimulation frequency adjusted by ±2 Hz; when the functional connectivity remodeling index is less than -5%, the stimulation intensity is first reduced to 50% of the initial value, and if it continues to decrease, stimulation is stopped.

[0041] Through the above method, this embodiment achieves an organic combination of individualized parameter configuration before treatment and closed-loop adaptive regulation during treatment, overcoming the technical defect of unstable efficacy of fixed parameter open-loop stimulation mode in the prior art.

[0042] Example 2 This embodiment, based on Embodiment 1, further describes the safety constraint mechanism in closed-loop parameter adjustment.

[0043] The closed-loop parameter adjustment includes a three-layer safety constraint mechanism, comprising hard constraints, soft constraints, and a human-machine collaborative supervision layer. The hard constraints limit the stimulation intensity range to 0.5 to 2 mA, the stimulation frequency range to 1 to 100 Hz, and the duration of a single treatment session to 10 to 30 minutes.

[0044] Hard constraints are based on clinical safety guidelines that set absolute limits: stimulation intensity exceeding 2 mA may cause discomfort or tissue damage; stimulation frequency exceeding the range of 1 to 100 Hz lacks sufficient safety evidence; single treatment sessions that are too short cannot produce an effective cumulative stimulation effect, while sessions that are too long increase the risk of fatigue and discomfort.

[0045] The soft constraints employ a Constrained Markov Decision Process (CMM) framework, including a cost function constraint that limits the maximum magnitude of each parameter adjustment. The CMM optimizes the reward function while ensuring that the expected value of the cost function does not exceed a preset threshold; the cost function quantifies the risk level of parameter adjustments.

[0046] The human-machine collaborative supervision layer allows clinicians to manually intervene or terminate treatment when abnormalities are detected, forming a multi-layered safety assurance system and solving the problem of the lack of safety constraint mechanisms in existing systems from a technical architecture perspective.

[0047] Example 3 This embodiment, based on Embodiment 1, further describes a small-sample offline pre-training method for reinforcement learning algorithms.

[0048] The offline pre-training method of the reinforcement learning algorithm includes a behavior cloning warm-up stage and a conservative Q-learning stage, wherein the behavior cloning warm-up stage is executed before the conservative Q-learning stage.

[0049] The behavioral cloning warm-up phase utilizes treatment decision data from clinical experts and trains a policy network through supervised learning, enabling the algorithm to quickly learn reasonable parameter adjustment patterns and avoid inefficient exploration and potential risks caused by starting with random initialization.

[0050] The conservative Q-learning phase further optimizes the strategy based on behavior cloning. By assigning lower Q values ​​to actions outside the offline dataset, it prevents the algorithm from overconfidently selecting unverified dangerous actions and ensures that a safe and effective control strategy is learned under limited data conditions.

[0051] The offline pre-training method supports small-sample learning scenarios with 30 to 80 cases, adapting to the realities of high clinical data collection costs and limited sample sizes. The online fine-tuning phase employs a soft actor-critic algorithm, achieving a balance between exploration and utilization by maximizing the entropy regularization objective function.

[0052] Example 4 This embodiment further describes a multimodal fusion efficacy prediction scheme based on embodiment 1.

[0053] The machine learning-based classification model integrates EEG signal features and blood biomarker features. The EEG signal features include the power spectral density of the 40 Hz gamma band. Gamma-band EEG activity is closely related to synchronous neuronal firing and cognitive function, and Alzheimer's disease patients typically exhibit reduced gamma-wave power.

[0054] The blood biomarker characteristics mentioned include at least the β-amyloid Aβ42 / 40 ratio or phosphorylated tau protein p-tau181 related indicators. β-amyloid protein and phosphorylated tau protein are pathological markers of Alzheimer's disease, and their blood concentrations are correlated with disease severity.

[0055] The classification model calculates the weight coefficients of each modality feature through an attention mechanism and performs weighted fusion. The attention mechanism can automatically learn the contribution of different modality features to efficacy prediction, assign higher weights to important features, and improve the accuracy of efficacy prediction by comprehensively utilizing complementary information from functional near-infrared spectroscopy, electroencephalography, and blood biomarkers through multimodal fusion.

[0056] Example 5 This embodiment describes a closed-loop adaptive neuromodulation system for Alzheimer's disease based on functional near-infrared spectroscopy functional connectivity features.

[0057] like Figure 1 As shown, the system includes: a functional near-infrared spectroscopy acquisition module, a functional connectivity feature extraction module, a therapeutic effect prediction module, a transcranial AC stimulation parameter configuration module, a real-time monitoring module, a reinforcement learning closed-loop control module, and a safety monitoring module. The modules are connected to each other via a data bus.

[0058] The functional near-infrared spectroscopy acquisition module is used to acquire cerebral oxygenation signals in subjects during resting state and treatment. This module includes a multi-channel light source and detector array, covering key brain regions such as the prefrontal, temporal, temporoparietal, and occipital lobes, with a sampling frequency of no less than 10 Hz to capture dynamic changes in cerebral oxygenation signals.

[0059] The functional connectivity feature extraction module is connected to the functional near-infrared spectroscopy acquisition module to calculate the functional connectivity strength between brain regions based on brain blood oxygenation signals and extract functional connectivity feature vectors. This module includes a signal preprocessing unit and a feature calculation unit. The signal preprocessing unit filters, denoises, and performs baseline correction on the raw light intensity signal, while the feature calculation unit calculates the functional connectivity matrix between brain regions and extracts graph theory features.

[0060] The efficacy prediction module is data-connected to the functional connectivity feature extraction module, and is used to predict the transcranial alternating current stimulation response probability based on the functional connectivity feature vector. This module includes a feature input interface, a machine learning model, and a response probability output interface. The machine learning model can be an ensemble of one or more of the following: support vector machine, graph neural network, or Transformer model.

[0061] The transcranial alternating current stimulation parameter configuration module is connected to the efficacy prediction module and is used to configure the stimulation frequency, stimulation intensity, stimulation target, and stimulation waveform based on the efficacy prediction results. This module includes a parameter database and parameter selection logic. The parameter database stores recommended parameter ranges corresponding to different response types.

[0062] The real-time monitoring module is connected to the functional near-infrared spectroscopy acquisition module to collect cerebral blood oxygenation signals in real time during treatment and calculate the functional connectivity remodeling index. This module includes a real-time data acquisition interface and a functional connectivity remodeling index calculation unit, collecting cerebral blood oxygenation signals at preset time intervals during treatment.

[0063] The reinforcement learning closed-loop control module is connected to both the real-time monitoring module and the transcranial AC stimulation parameter configuration module. It adjusts the transcranial AC stimulation parameters in real time based on the functional connectivity remodeling index and the current stimulation parameters. This module includes a state encoder, a policy network, a value network, and a parameter adjustment actuator. The state encoder encodes the functional connectivity remodeling index, dynamic characteristics of cerebral blood oxygenation, current stimulation parameters, and treatment progress into a state vector. The parameter adjustment actuator sends adjustment commands to the transcranial AC stimulation device.

[0064] The safety monitoring module is connected to the reinforcement learning closed-loop control module and the transcranial AC stimulation parameter configuration module via control signals. It monitors whether the stimulation parameters meet safety constraints and triggers safety interventions when necessary. This module includes a hard constraint checking unit, a soft constraint evaluation unit, and a safety intervention trigger. The safety intervention trigger immediately issues an alarm and automatically reduces the stimulation intensity or pauses treatment when it detects a violation of hard constraints or that the cost of soft constraints exceeds a threshold. The entire system can be integrated into a portable device for convenient clinical application and home rehabilitation scenarios.

[0065] Comparative Example 1 The closed-loop adaptive neuromodulation method of Example 1 is different in that: in step S2, the electroencephalogram (EEG) signal is used instead of the functional near-infrared spectroscopy signal as the feedback signal source, and the EEG data is collected and the functional connectivity features are extracted. The remaining steps are the same as in Example 1.

[0066] Because of the low spatial resolution of EEG, it is difficult to accurately characterize the topological characteristics of functional connectivity networks across multiple brain regions. The constructed functional connectivity matrix cannot effectively capture the characteristic cross-brain region functional connectivity abnormality patterns in AD patients, leading to a decrease in the accuracy of subsequent efficacy prediction and limited feedback quality of closed-loop regulation.

[0067] Network statistical methods based on functional near-infrared spectroscopy can identify significantly differentially functioning subnetworks, while EEG-based methods, due to limitations in spatial resolution, struggle to accurately locate these differentially functioning subnetworks. A comparison of Example 1 and Comparative Example 1 demonstrates that functional near-infrared spectroscopy has a significant advantage in characterizing multi-brain region functional connectivity, an effect that was unexpected.

[0068] Comparative Example 2 The closed-loop adaptive neuromodulation method of Example 1 is similar to that in Example 1, except that steps S3 (therapeutic efficacy prediction step) and S4 (individualized parameter configuration step) are removed. All subjects are uniformly subjected to transcranial alternating current stimulation with fixed initial parameters. The stimulation frequency is fixed at 40 Hz and the stimulation intensity is fixed at 1 mA. The remaining steps are the same as in Example 1.

[0069] like Figure 3 As shown, the cognitive function change curves over time for the real stimulation group (Example 1) and the sham stimulation group indicate that Example 1, with its individualized parameter configuration achieved through a efficacy prediction mechanism, demonstrated a significantly higher treatment effectiveness than the fixed parameter regimen in Comparative Example 2. In Comparative Example 2, due to the lack of a pre-stimulation efficacy prediction mechanism, approximately 30% to 40% of the subjects showed no significant response to the fixed parameter regimen.

[0070] The bootstrap mediation analysis pathway validated the complete causal link from transcranial alternating current stimulation intervention to dynamic brain functional network state reconstruction and then to cognitive function improvement. Example 1, through individualized parameter configuration, significantly enhanced the mediating effect of this causal link, while in the fixed parameter scheme of Comparative Example 2, the aforementioned mediating effect was significantly weakened. Comparing Example 1 and Comparative Example 2, it can be seen that the individualized parameter configuration brought about by the pre-treatment efficacy prediction mechanism can significantly improve the treatment effectiveness rate, an effect that was unexpected.

[0071] Comparative Example 3 The closed-loop adaptive neuromodulation method of Example 1 differs in that: in step S6, a simple rule-driven feedback control is used instead of the reinforcement learning algorithm, that is, when the functional connectivity remodeling index is lower than the preset threshold, the stimulation intensity is increased by 0.1 mA at a fixed step size, and the remaining steps are the same as in Example 1.

[0072] Simple rule-driven feedback control cannot effectively handle nonlinear, high-dimensional brain state spaces, lacks the ability to optimize sequential decisions for long-term treatment, and cannot find the optimal parameter adjustment strategy under multi-objective constraints (improvement of functional connectivity, enhancement of cognitive function, safety assurance, and stability of parameter adjustment), resulting in unstable long-term efficacy.

[0073] Comparing Example 1 with Comparative Example 3, it can be seen that the closed-loop adaptive regulation scheme using reinforcement learning algorithm has significant advantages in handling complex brain state spaces and achieving multi-objective optimization. It can maximize long-term efficacy through sequential decision optimization, which is an unexpected effect.

[0074] Test Example 1 The efficacy of the neuromodulation protocols used in Example 1 and Comparative Examples 1 to 3 was evaluated. The test indicators included changes in the functional connectivity remodeling index and changes in the Mini-Mental State Examination score. The results are shown in Table 1.

[0075] Table 1. Comparison of therapeutic effects between Example 1 and each comparative example. The test results above show that Example 1, through the synergistic effect of functional near-infrared spectroscopy replacing EEG signals, pre-treatment efficacy prediction mechanism, and reinforcement learning closed-loop control, is superior to the comparative examples in terms of functional connectivity remodeling and cognitive function improvement, fully verifying the technical effects of the different technical features of the present invention.

[0076] It is understood that the functional connectivity calculation method in the above embodiments is not limited to the Pearson correlation coefficient, and can also use partial correlation coefficient, wavelet coherence, or a combination thereof. Obviously, the above classification model is not limited to support vector machines, graph neural networks, or Transformer models, and can also use machine learning models commonly used in this field, such as random forests and convolutional neural networks. It is understood that the above reinforcement learning algorithm is not limited to a combination of conservative Q-learning and soft actor-critic algorithm, and can also use other reinforcement learning algorithms such as proximal policy optimization and deep Q-networks. Obviously, the preset threshold of the above functional connectivity remodeling index is not limited to 10% and -5%, and can be adjusted according to specific clinical application scenarios. It is understood that the above system is not limited to integration into portable devices, but can also be integrated into desktop medical devices or telemedicine platforms.

Claims

1. A closed-loop adaptive neuromodulation method for Alzheimer's disease, characterized in that, Includes the following steps: S1: Collect resting-state cerebral blood oxygenation signals from the subjects for 5 to 15 minutes, covering the prefrontal cortex, temporal cortex, temporoparietal cortex, and occipital cortex. S2: Calculate the functional connectivity correlation coefficient between brain regions using at least one of Pearson correlation coefficient, partial correlation coefficient or wavelet coherence, construct the functional connectivity matrix, and extract the functional connectivity feature vector; S3: Input the functional connection feature vector into a machine learning-based classification model, output the transcranial alternating current stimulation response probability, and divide the subjects into a high response group, a medium response group, and a low response group according to the response probability; S4: Configure the initial parameters of transcranial alternating current stimulation based on the efficacy prediction results obtained in step S3. The initial parameters include stimulation frequency, stimulation intensity, stimulation target point and stimulation waveform. The stimulation frequency of the high response group is set to the 40 Hz gamma band. S5: During transcranial alternating current stimulation, brain blood oxygenation signals are collected in real time, and the functional connectivity remodeling index is extracted. The functional connectivity remodeling index represents the rate of change in functional connectivity strength during the treatment process. S6: Based on the functional connectivity remodeling index and the current stimulation parameters, the transcranial alternating current stimulation parameters are adjusted in real time using a reinforcement learning algorithm. When the rate of change of the functional connectivity remodeling index is lower than a preset threshold, parameter adjustment is triggered to improve the functional connectivity strength.

2. The Alzheimer's disease closed-loop adaptive neuromodulation method according to claim 1, characterized in that, In step S2, the extraction of the functional connectivity matrix uses Pearson correlation coefficient to calculate linear correlation between brain regions, and / or uses partial correlation coefficient to eliminate the influence of indirect connectivity, and / or uses wavelet coherence analysis of frequency domain functional connectivity; the functional connectivity feature vector includes one or more graph theory indices such as global efficiency, clustering coefficient, and modularity.

3. The Alzheimer's disease closed-loop adaptive neuromodulation method according to claim 1, characterized in that, In step S3, the machine learning-based classification model includes a support vector machine model, a graph neural network model, and / or a Transformer model, wherein the support vector machine model is used for small sample scenarios, the graph neural network model is used to capture brain network topological features, and the Transformer model is used for sequence feature modeling; the classification model adopts an ensemble learning strategy to fuse the prediction results of multiple base models, and the ensemble learning strategy adopts a weighted average or voting mechanism.

4. The Alzheimer's disease closed-loop adaptive neuromodulation method according to claim 1, characterized in that, In step S6, the state space of the reinforcement learning algorithm includes a functional connectivity feature vector, dynamic cerebral blood oxygenation features, a current stimulus parameter vector, and treatment progress; the reward function of the reinforcement learning algorithm is... ,in , , , For preset weighting coefficients, The value range is from 0.3 to 0.

5. The value range is from 0.2 to 0.

4. The value range is from 0.1 to 0.

3. The value range is from 0.05 to 0.

15. Indicates the change in functional connections. This indicates the change in the score of the Mini-Mental State Examination. This indicates a function that indicates a security violation. This indicates the parameter adjustment range; the reinforcement learning algorithm uses the conservative Q-learning algorithm for offline pre-training and the soft actor-critic algorithm for online fine-tuning.

5. The Alzheimer's disease closed-loop adaptive neuromodulation method according to claim 1, characterized in that, In step S6, the closed-loop parameter adjustment includes a safety constraint mechanism, which includes hard constraints and soft constraints. The hard constraints limit the stimulation intensity range to 0.5 to 2 mA, the stimulation frequency range to 1 to 100 Hz, and the duration of a single treatment session to 10 to 30 minutes. The soft constraints adopt a constrained Markov decision process framework, which includes a cost function constraint, limits the maximum magnitude of each parameter adjustment, and sets up a human-machine collaborative supervision layer.

6. The Alzheimer's disease closed-loop adaptive neuromodulation method according to claim 1, characterized in that, In step S5, the formula for calculating the functional connectivity reshaping index is as follows: ,in Indicates the current functional connection strength. The baseline functional connectivity strength is indicated. In step S6, when the functional connectivity remodeling index is greater than or equal to 10%, the current stimulation parameters are maintained. When the functional connectivity remodeling index is between -5% and 10%, the stimulation parameters are finely adjusted. When the functional connectivity remodeling index is less than -5%, the stimulation intensity is reduced to 50% of the initial value. If the functional connectivity remodeling index continues to decrease, stimulation is stopped.

7. The Alzheimer's disease closed-loop adaptive neuromodulation method according to claim 1, characterized in that, In step S3, the machine learning-based classification model integrates EEG signal features and blood biomarker features. The EEG signal features include the power spectral density of the 40 Hz gamma band, and the blood biomarker features include at least β-amyloid protein-related indicators or phosphorylated tau protein-related indicators. The classification model calculates the weight coefficients of each modality feature through an attention mechanism and performs weighted fusion.

8. The Alzheimer's disease closed-loop adaptive neuromodulation method according to claim 4, characterized in that, The offline pre-training method of the reinforcement learning algorithm includes a behavior cloning warm-up stage and a conservative Q-learning stage. The behavior cloning warm-up stage is executed before the conservative Q-learning stage. The behavior cloning warm-up stage initializes the algorithm parameters by imitating the expert policy. The conservative Q-learning stage achieves conservative estimation by penalizing the Q-value of out-of-policy actions. The offline pre-training method supports small sample learning scenarios with a sample size of 30 to 80 cases.

9. The Alzheimer's disease closed-loop adaptive neuromodulation method according to claim 1, characterized in that, In step S6, the closed-loop parameter adjustment includes a security assessment system. The security assessment system includes a triple estimator based on ordinary differential equations to quantify model uncertainty. The triple estimator is used to estimate the confidence interval of the model prediction, detect distribution shift using the maximum mean difference with a shift threshold range of 0.05 to 0.15, and perform robustness testing using the Monte Carlo method.

10. A closed-loop adaptive neuromodulation system for Alzheimer's disease, characterized in that, include: A functional near-infrared spectroscopy acquisition module is used to acquire cerebral blood oxygenation signals of subjects in a resting state and during treatment. The functional connectivity feature extraction module is connected to the functional near-infrared spectroscopy acquisition module and is used to calculate the functional connectivity strength between brain regions and extract functional connectivity feature vectors based on the brain blood oxygenation signal. The efficacy prediction module is data-connected to the functional connectivity feature extraction module and is used to predict the transcranial alternating current stimulation response probability based on the functional connectivity feature vector. The transcranial alternating current stimulation parameter configuration module is connected to the efficacy prediction module and is used to configure the stimulation frequency, stimulation intensity, stimulation target and stimulation waveform according to the efficacy prediction results. The real-time monitoring module is connected to the functional near-infrared spectroscopy acquisition module and is used to acquire cerebral blood oxygenation signals in real time and calculate the functional connectivity remodeling index during treatment. The reinforcement learning closed-loop control module is connected to the real-time monitoring module and the transcranial AC stimulation parameter configuration module respectively, and is used to adjust the transcranial AC stimulation parameters in real time according to the functional connection remodeling index and the current stimulation parameters. The safety monitoring module is connected to the reinforcement learning closed-loop control module and the transcranial AC stimulation parameter configuration module via control signals, and is used to monitor whether the stimulation parameters meet the safety constraints and trigger safety intervention when necessary; the modules are connected to each other via a data bus.

Images