AI-driven transcranial photobiomodulation intelligent control method, device and storage medium

Abstract

The present disclosure relates to an AI-driven transcranial photobiomodulation intelligent control method, device and storage medium. The method comprises obtaining a plurality of sets of resting state data of a transcranial photobiomodulation device user and stimulation state data of a controllable external stimulus; an AI chip extracts time series features using a space-time convolution network, uses a multi-head attention mechanism Transformer structure to model the interaction of time series features of different modalities, dynamically calculates the weight of each fused feature vector through a self-attention mechanism, maps to obtain a physiological feature map and stores it as a personalized baseline model; the optimal working parameter is calculated by a deep reinforcement learning algorithm, and after the optimal working parameter is executed, the optimal working parameter is dynamically adjusted according to the personalized baseline model using the deep reinforcement learning algorithm. The present disclosure realizes individualized and real-time adjustment of the stimulation parameter, does not require population average threshold information, and has high accuracy and good effect in adjustment and control.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent parameter regulation technology, and in particular to AI-driven transcranial photobiological regulation intelligent control methods, devices and storage media. Background Technology

[0002] Photobiomodulation (PBM) is a non-invasive technique that modulates the function of biological tissues through light radiation of specific wavelengths (typically 620-1100 nm). Its mechanism of action is primarily based on the absorption of red and near-infrared light by cytochrome c oxidase in mitochondria, which triggers a series of biochemical cascade reactions, including increased adenosine triphosphate (ATP) production, nitric oxide release, and regulation of reactive oxygen species. Due to its non-invasive nature and high safety profile, PBM has been developed for the treatment of various clinical diseases, such as improving ischemic stroke symptoms, treating traumatic brain injury, and intervening in mental illnesses, neurological diseases (such as Alzheimer's and Parkinson's), and neurodevelopmental disorders. It can also be used to improve cognitive and memory abilities in healthy individuals.

[0003] Brain function is typically enhanced through wearable tPBM (Transcranial Photobiomodulation) devices. Chinese Patent CN119546229A, entitled "Method and Apparatus for Closed-Loop Transcranial Photobiomodulation Stimulation Using Cognitive Testing," discloses a physiological and neurostimulation device based on transcranial photobiomodulation. This device provides continuous wave or pulsed light sources to stimulate specific brain regions. It can also utilize physiological signals, such as electroencephalography (EEG) or heart rate variability (HRV), to assess the effectiveness of tPBM intervention or modify tPBM parameters based on these signals. This technical solution, geared towards cognitive testing tasks, combines transcranial photobiomodulation with event-related potentials (ERPs) recorded during cognitive testing. Through time-frequency analysis of ERPs and monitoring of other physiological signals, it achieves neurofeedback modulation.

[0004] Current transcranial photobiological modulation and control methods rely on the average threshold of the population, which cannot meet personalized needs, cannot respond to individual differences in patients, and have low accuracy and poor effect in modulation and control. Summary of the Invention

[0005] Therefore, it is necessary to provide AI-driven transcranial photobiological modulation intelligent control methods, devices, and storage media to address the above-mentioned problems.

[0006] To solve the above problems, the present disclosure adopts the following technical solution:

[0007] In a first aspect, this disclosure provides an AI-driven transcranial photobiological modulation intelligent control method, comprising the following steps:

[0008] Step 1: Obtain multiple sets of resting state data of the transcranial photobiomodulation device user, and obtain the stimulation state data of the controllable external stimulation of the transcranial photobiomodulation device user.

[0009] Step 2: The AI ​​chip preprocesses the state data obtained in Step 1 to obtain preprocessed modal data. It then uses a spatiotemporal convolutional network to extract the time-series features of the preprocessed modal data. A Transformer structure with a multi-head attention mechanism is used to interactively model the time-series features of different modalities to obtain fused feature vectors. The weights of each fused feature vector are dynamically calculated using a self-attention mechanism. The fused feature vectors are then mapped to a multi-dimensional physiological state space to obtain a physiological feature map. This physiological feature map is stored as a personalized baseline model.

[0010] Step 3: The AI ​​chip calculates the optimal operating parameters of the transcranial photobiological modulation device used by the user based on the physiological feature map and through a deep reinforcement learning algorithm.

[0011] Step 4: The transcranial photobiological modulation device executes optimal operating parameters and monitors the physiological signals of the user in real time. The AI ​​chip analyzes the degree of deviation of the physiological signals from their normal range based on a personalized baseline model and a deep reinforcement learning algorithm, and dynamically adjusts the optimal operating parameters accordingly.

[0012] Secondly, this disclosure provides an AI-driven transcranial photobiological modulation intelligent control device, including:

[0013] The biosensor module is used to acquire multiple sets of resting state data of the transcranial photobiomodulation device user and to acquire stimulation state data of the controllable external stimuli of the transcranial photobiomodulation device user.

[0014] The AI ​​chip is used to acquire resting state data and stimulation state data from the biosensor module, preprocess the state data to obtain preprocessed modal data, extract time-series features of the preprocessed modal data using a spatiotemporal convolutional network, interactively model the time-series features of different modalities using a Transformer structure with multi-head attention mechanism to obtain fused feature vectors, dynamically calculate the weights of each fused feature vector through a self-attention mechanism, and map the fused feature vectors to a multi-dimensional physiological state space to obtain a physiological feature map, which is stored as a personalized baseline model. The AI ​​chip is used to calculate the optimal operating parameters of the transcranial photobiomodulation device used by the user based on the personalized baseline model and through a deep reinforcement learning algorithm.

[0015] The optical modulation module is used to regulate the light waves of the transcranial photobiomodulation device according to the optimal operating parameters;

[0016] The power management module is used to supply power to the working area of ​​the transcranial photobiomodulation device according to the optimal operating parameters;

[0017] The biosensor module is also used to monitor the physiological signals of the user of the transcranial photobiomodulation device in real time when the transcranial photobiomodulation device is operating at optimal parameters.

[0018] The AI ​​chip is also used to obtain the physiological signals monitored in real time by the biosensor module, analyze the degree to which the physiological signals deviate from their normal range using a deep reinforcement learning algorithm, and dynamically adjust the optimal operating parameters accordingly.

[0019] Thirdly, this disclosure provides a storage medium storing a computer program. When executed by a processor, the computer program is capable of: preprocessing state data to obtain preprocessed modal data; extracting time-series features of the preprocessed modal data using a spatiotemporal convolutional network; interactively modeling time-series features of different modalities using a Transformer structure with a multi-head attention mechanism to obtain fused feature vectors; dynamically calculating the weights of each fused feature vector using a self-attention mechanism; and mapping the fused feature vectors to a multi-dimensional physiological state space to obtain a physiological feature map. The steps include: storing physiological feature maps as a personalized baseline model; calculating the optimal operating parameters of the transcranial photobiomodulation device used by the user based on the physiological feature maps using a deep reinforcement learning algorithm; obtaining real-time monitoring of the physiological signals of the transcranial photobiomodulation device user; analyzing the degree of deviation of the physiological signals from their normal range using a deep reinforcement learning algorithm based on the personalized baseline model and dynamically adjusting the optimal operating parameters accordingly; the state data includes multiple sets of resting state data of the transcranial photobiomodulation device user and stimulation state data of the transcranial photobiomodulation device user under controllable external stimulation.

[0020] The aforementioned AI-driven intelligent control method, device, and storage medium for transcranial photobiomodulation involves extracting time-series features from the resting state data of a specific transcranial photobiomodulation device user and the stimulation state data of controllable external stimuli using a spatiotemporal convolutional network. It then employs a Transformer structure with a multi-head attention mechanism to interactively model the time-series features of different modalities, obtaining a fused feature vector. The weights of each fused feature vector are dynamically calculated using a self-attention mechanism to obtain a physiological feature map and a personalized baseline model. Based on the physiological feature map, a deep reinforcement learning algorithm is used to calculate the optimal operating parameters of the transcranial photobiomodulation device used by the user. After the transcranial photobiomodulation device executes the optimal operating parameters, it dynamically adjusts the optimal operating parameters based on real-time monitored physiological signals using a deep reinforcement learning algorithm. This achieves individualized, real-time adjustment of the stimulation parameters for the user, fully considering individual differences among users, eliminating the need for group average threshold information, and resulting in high accuracy and good regulatory effect. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an AI-driven transcranial photobiological modulation intelligent control method in one embodiment of the present disclosure.

[0022] Figure 2 This is a view of the surface of a head-mounted transcranial photobiomodulation device before wearing;

[0023] Figure 3 A three-dimensional view of a head-mounted transcranial photobiological modulation device;

[0024] Figure 4 This is a view of the wearing side surface of a head-mounted transcranial photobiomodulation device;

[0025] Figure 5 A schematic diagram of the inner surface of a head-mounted transcranial photobiological modulation device;

[0026] Figure 6 This is a schematic diagram of the system structure in one embodiment of the present disclosure;

[0027] Figure 7 This is a partial structural cross-sectional view of a head-mounted transcranial photobiomodulation device. Detailed Implementation

[0028] The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0029] See Figure 1 This disclosure provides an AI-driven transcranial photobiological modulation intelligent control method, including:

[0030] Step 1, Obtain calibration data: Obtain multiple sets of resting state data of the transcranial photobiomodulation device user, and obtain the stimulation state data of the controllable external stimulation of the transcranial photobiomodulation device user.

[0031] Step 2: The AI ​​chip preprocesses the state data obtained in Step 1 to obtain preprocessed modal data. The AI ​​chip uses a spatiotemporal convolutional network to extract the time-series features of the preprocessed modal data. The AI ​​chip uses a Transformer structure with a multi-head attention mechanism to interactively model the time-series features of different modalities to obtain fused feature vectors. The AI ​​chip dynamically calculates the weights of each fused feature vector through a self-attention mechanism. The AI ​​chip maps the fused feature vectors to a multi-dimensional physiological state space to obtain a physiological feature map. The physiological feature map is stored as a personalized baseline model.

[0032] Step 3: The AI ​​chip calculates the optimal operating parameters of the transcranial photobiological modulation device used by the user based on the physiological feature map and through a deep reinforcement learning algorithm.

[0033] Step 4: The transcranial photobiological modulation device executes optimal operating parameters and monitors the physiological signals of the user in real time. The AI ​​chip analyzes the degree of deviation of the physiological signals from their normal range based on a personalized baseline model and a deep reinforcement learning algorithm, and dynamically adjusts the optimal operating parameters accordingly.

[0034] First, we introduce existing transcranial photobiomodulation devices, taking a head-mounted transcranial photobiomodulation device as an example. (See...) Figures 2 to 5 , Figure 5 This diagram shows the inner surface of the device's hemispherical structure, illustrating the light source partitioning and arrangement. The light source employs portable whole-brain control, providing a large coverage area. The adjustable range includes eight partitions: left frontal lobe, right frontal lobe, left parietal lobe, right parietal lobe, left occipital lobe, right occipital lobe, left temporal lobe, and right temporal lobe—eight illumination areas, each independently controllable. Adjustable parameters for light source control include frequency (adjustable continuous light and pulsed light, such as arbitrary frequencies like 10 Hz, 40 Hz, and 100 Hz), duty cycle, and power density. Simultaneously, the computer screen displays the square wave matrix, real-time temperature, start time, and the multimodal system startup process.

[0035] The circuit boards of the light sources in different brain regions can all be temperature-measured in real time. If the temperature exceeds the set temperature (e.g., room temperature) during light supply, a closed-loop temperature system will automatically activate to cool the area, preventing overheating and thermal interference to the head. Simultaneously, maintaining a constant temperature for the light source circuit boards ensures stable power output. A cooling channel exists between the light source circuit boards and the outer casing, using externally pumped low-temperature air for noiseless and rapid cooling. A multimodal signal acquisition device is located near the scalp on the inner side of the light source, providing real-time feedback of brain region signals from different individuals to drive photobiological regulation. This multimodal signal acquisition device includes an electroencephalogram (EEG) sensor, a heart rate variability (HRV) / photoplethysmography (PPG) sensor, and a skin conductance sensor. The EEG sensor collects electrical activity from the cerebral cortex, covering the delta, theta, alpha, beta, and gamma bands, with a sampling rate ≥500 Hz to capture high temporal resolution neural activity characteristics. The heart rate variability (HRV) sensor extracts the RR interval based on photoplethysmography (PPG) or electrocardiogram (ECG) to assess the balance of the autonomic nervous system. Electrodermal activity (EDA) sensors are used to record changes in skin conductance caused by sympathetic nerve activity, reflecting stress and emotional states. These sensors are time-aligned via a Precision Time Protocol (PTP) and transmitted to an edge AI chip.

[0036] In a specific embodiment, the AI-driven transcranial photobiological modulation intelligent control method specifically includes the following steps:

[0037] Step 1: In a quiet environment, the user of the transcranial photobiomodulation device wears the device, and multiple sets (5-10 groups) of resting state data are collected, recording signals such as baseline electrical activity (EEG) and autonomic indicators (HRV, EDA) as the resting baseline. Subsequently, various controllable external stimuli are applied sequentially, including: cognitive task-only stimulation (such as the n-back working memory task) to activate the prefrontal and parietal lobes; emotion-inducing stimulation-only (such as positive / negative images from the International Emotion Picture System (IAPS)) to activate the limbic system and autonomic responses; and sensory stimulation-only (such as visual flashes and auditory rhythmic sounds) to induce responses in specific sensory pathways. All stimulus events are synchronized with the physiological data at the millisecond level using event markers for subsequent accurate analysis.

[0038] Therefore, the resting state data includes EEG signals, heart rate variability signals, and skin surface potential change signals in the resting state. The stimulation state data of the controllable external stimulus includes: EEG signals under cognitive task stimulation only, heart rate variability signals under cognitive task stimulation only, skin surface potential change signals under cognitive task stimulation only, EEG signals under emotion-induced stimulation only, heart rate variability signals under emotion-induced stimulation only, skin surface potential change signals under emotion-induced stimulation only, EEG signals under sensory stimulation only, heart rate variability signals under sensory stimulation only, and skin surface potential change signals under sensory stimulation only.

[0039] It is understood that the user is a user who needs to determine / adjust the corresponding optimal working parameters.

[0040] Step 2: The AI ​​chip acquires all state data (resting state data and stimulus state data) from Step 1. It preprocesses the different modal state data obtained in Step 1 to obtain preprocessed modal data. The AI ​​chip uses a spatiotemporal convolutional network to extract the time-series features of the preprocessed modal data. The AI ​​chip employs a Transformer structure with a multi-head attention mechanism to interactively model the time-series features of different modal data, obtaining a fused feature vector. The AI ​​chip dynamically calculates the weights of each fused feature vector through a self-attention mechanism, achieving adaptive interaction of cross-modal features. The AI ​​chip maps the fused feature vectors to a multi-dimensional physiological state space to obtain a physiological feature map. This physiological feature map is stored as a personalized baseline model.

[0041] Step 2 is the process of predicting the brain state of the user / subject / patient. The final physiological feature map obtained in Step 2 is the predicted brain state and its corresponding probability. Because each user has their own physiological feature map, it is also called a specific physiological feature map.

[0042] Step 2.1, Preprocessing: For all signals obtained in Step 1, i.e. physiological signals of different modalities, perform bandpass filtering, denoising (ICA independent principal component analysis or wavelet denoising), artifact removal and standardization.

[0043] Step 2.2, Time Series Modeling: Use a spatiotemporal convolutional network to extract the short-term dependence and local spatial dynamic features of each modality of physiological signal to obtain time series features. The size of the convolutional kernel is adaptively adjusted according to the signal sampling rate.

[0044] Step 2.3, Cross-modal Feature Interaction: A Transformer structure with a multi-head attention mechanism is used to model the interaction of time-series features from different modalities. The weights (fusion weights) of each modality in the current time state are dynamically calculated through a self-attention mechanism, enabling the adaptive allocation of modal weights to cross-modal features via self-attention weights. For example, the weight of EDA signals is increased during emotional fluctuations, and the weight of EEG signals is increased in cognitive tasks.

[0045] Step 2.4, State Vector Generation: The fused feature vectors are mapped onto the subject's multidimensional physiological state space to form a real-time updated individual neurophysiological state representation, providing input for personalized baseline models and reinforcement learning optimization.

[0046] Understandably, since users' circumstances can change, steps 1 and 2 need to be repeated to update the individual neurophysiological state representation and personalized baseline model.

[0047] Specifically, in step 2.2, time-frequency analysis (EEG wavelet packet decomposition, EDA event-related conductance change extraction, etc.) is performed on the collected multimodal data, and feature indices (such as power spectral density, graph theory indices corresponding to the functional connectivity matrix, and frequency domain parameters of heart rate variability) are calculated. In step 2.3, principal component analysis (PCA) or t-SNE dimensionality reduction is used to map high-dimensional features to a low-dimensional visualization space to discover differences in physiological response patterns of subjects under different stimulus conditions. In step 4, density-based spatial clustering of applications with noise (DBSCAN) or hidden Markov model (HMM) is used to cluster the feature space, and the results are mapped into a specific physiological feature atlas (i.e., individual neurophysiological state representation), which contains the feature distribution regions and transition paths corresponding to each stimulus condition.

[0048] Specifically, the feature map is stored as a personalized baseline model, including: the central value and normal fluctuation range of physiological characteristics in the resting state; the direction and magnitude of feature shifts under different stimulation conditions; and the feature labels and weights of specific abnormal states (such as high anxiety and low arousal) (i.e., the results of step 2). During actual treatment, real-time collected multimodal physiological states are matched with this baseline model using similarity analysis (calculated using Mahalanobis cosine distance or cosine similarity) to determine the degree to which the current state deviates from the normal range. Subsequent reinforcement learning optimization dynamically adjusts the tPBM parameters accordingly, achieving precise regulation for individual subjects.

[0049] The physiological feature map is stored as a personalized baseline model by encoding the structured features of the feature map into vectors as an initial reference model. Preferably, the training process of the personalized baseline model is carried out in the cloud. According to the preset, the model parameters are updated periodically and then deployed on the AI ​​chip. That is, the AI ​​chip realizes the training (and updating) of the personalized baseline model through the cloud.

[0050] The personalized baseline model is established by collecting multimodal physiological response data from subjects in a resting state and under a series of standardized external stimuli. It characterizes "the unique physiological response patterns of the subject in different states." This atlas defines the range, pattern, and transition characteristics of the subject's "normal" or "baseline" physiological signals in various states, providing a crucial benchmark reference for subsequent personalized optimization.

[0051] This embodiment sets up a calibration phase at the beginning of treatment to acquire multimodal physiological response data of subjects in a resting state and under different external stimuli, in order to establish a specific physiological feature atlas and provide a benchmark reference for subsequent reinforcement learning. This method can accurately capture the neurophysiological response patterns of individual subjects, rather than relying on group average parameters; it can provide a quantitative deviation index, which facilitates reinforcement learning to quickly converge to the personalized optimal strategy; and it can enhance the system's ability to adapt to changes in the subject's state during long-term use.

[0052] Step 3: Based on the obtained specific physiological feature map, calculate the optimal operating parameters of the transcranial photobiological modulation device using a deep reinforcement learning algorithm;

[0053] Step 4: The transcranial photobiomodulation device executes optimal operating parameters and monitors the physiological signals of the user in real time (such as local field potential, electroencephalogram, or neurotransmitter concentration). Based on the personalized baseline model, the device analyzes the degree of deviation of the physiological signals from their normal range using a deep reinforcement learning algorithm and dynamically adjusts the optimal operating parameters of the transcranial photobiomodulation device accordingly.

[0054] The AI ​​chip performs similarity matching (using Mahalanobis cosine distance or cosine similarity calculation) between real-time monitored physiological signals and a personalized baseline model. By determining the degree to which the current state deviates from the normal range through similarity matching, the AI ​​chip uses a deep reinforcement learning algorithm to adjust the optimal operating parameters based on the degree of deviation, achieving precise adjustment for each individual subject. This allows for the re-execution of step 4 with the updated optimal operating parameters.

[0055] Here, the operating parameters include information on the emitting area (or working area or irradiation site), light intensity, wavelength, and pulse frequency. The optimal operating parameters are the best combination of information such as emitting area, light intensity, wavelength, and pulse frequency, including at least these four types of information.

[0056] The core of the closed-loop regulation mechanism disclosed herein is to utilize a deep reinforcement learning (DRL) algorithm to dynamically adjust the key parameters of the tPBM based on real-time multimodal physiological signals. These parameters include: wavelength λ: 650-1100 nm, covering the red and near-infrared bands, regulating the photon penetration depth and biological effects; light intensity I: 1-250 mW / cm², affecting the stimulation intensity and safety of photobiological regulation; pulse frequency f: 1-100 Hz, regulating the rhythmic resonance effect of neural activity, or continuous light; and irradiation sites P': left frontal lobe, right frontal lobe, left parietal lobe, right parietal lobe, left occipital lobe, right occipital lobe, left temporal lobe, and right temporal lobe, with different regions corresponding to different disease symptoms in the neural network.

[0057] Specifically, deep reinforcement learning is implemented using either Deep Deterministic Policy Gradient (DDPG) or Proximal Policy Optimization (PPO) algorithms. DDPG is suitable for continuous action spaces, enabling fine-tuning of wavelength, intensity, and frequency; PPO exhibits high stability during policy updates, making it suitable for safety-constrained optimization in medical scenarios. Weighted reward optimization is employed in short-term decision-making, while in long-term policy updates, the optimal parameter combinations under different objective weights are retained to construct a Pareto optimal set, addressing changes in treatment stages or subject preferences.

[0058] The state space of the deep reinforcement learning algorithm consists of fused feature vectors, and the action space is a set of working parameters. The reward function of the deep reinforcement learning algorithm is represented in a weighted form, taking into account the symptom relief speed of the transcranial photobiological modulation device user, the user's comfort level, the power consumption of the transcranial photobiological modulation device, and the specific temperature control of the transcranial photobiological modulation device. The reinforcement learning model is as follows:

[0059] State space S: Composed of fused feature vectors (real-time fused multimodal physiological feature vectors), such as EEG power spectrum features, HRV frequency domain parameters, and EDA change rate.

[0060] Action space A: The set of combinations of tPBM parameters (equivalent to working parameters), which can be encoded as a quadruple of [λ, I, f, P'].

[0061] The reward function R comprehensively considers the speed of symptom relief (e.g., improvement in EEG α / β ratio, decrease in self-rated mood), user comfort (temperature monitored by temperature sensors and subjective questionnaire feedback), power consumption (power consumption of the transcranial photobiomodulation device, specifically the power consumption of the light source, AI chip, and temperature control system), and temperature control (ensuring the temperature of the transcranial photobiomodulation device does not exceed a safe threshold, specifically ensuring the temperature of the light source and AI chip does not exceed a safe threshold). The reward function can be in a weighted form.

[0062]

[0063] in, This indicates the speed at which symptoms are relieved for users of the transcranial photobiological modulation device. express The weight, Indicates the comfort level of the user of the transcranial photobiological modulation device. express The weight, This indicates the power consumption of the transcranial photobiological modulation device. express The weight, This indicates the temperature control of the transcranial photobiological regulation device. express The weight, This represents the reward value.

[0064] The following detailed description focuses on the processing of a specific EEG signal. The processing of other signals in step 1 is the same as that of this EEG signal.

[0065] Taking EEG signal processing as an example, this paper explains how to perform signal preprocessing and feature extraction to generate personalized state space vectors (that is, mapping the fused feature vectors to a multi-dimensional physiological state space to obtain a specific physiological feature map), which is understandable. This disclosure proposes a cascaded Transformer model for identity embedding (EEG Identity-embedding cascaded Transformer, EEG-IECT). For the example of EEG signals, it is a cascaded Transformer model based on EEG signal identity embedding.

[0066] The cascaded Transformer model for identity embedding comprises three modules in its end-to-end process: a shallow convolutional feature encoding module, a cascaded Transformer module for coarse-to-fine granular feature extraction (CCFT), and an identity embedding module (IE). EEG-IECT employs a shallow convolutional feature encoding module to encode temporal and spatial features. The subsequent CCFT module aims to capture coarse- and fine-grained temporal patterns. Finally, an IE module is integrated into the classifier to retrieve the best-matching template from the features learned by the CCFT module and outputs a predicted probability of the current brain state. In other words, the CCFT module learns general features for distinguishing brain states, while the IE module outputs a set of parameters specific to the individual, which refines the general features to determine the current brain state.

[0067] To extract the time-frequency features of EEG signals, this embodiment designs a shallow feature encoder based on CNN, namely a shallow convolutional feature encoding module. This shallow convolutional feature encoding module utilizes a spatiotemporal convolutional network to extract short-term dependencies and local spatial dynamic features from each preprocessed modality data to obtain time-series features. Two one-dimensional convolutional layers are applied along the time and spatial dimensions to extract features from the EEG signals. Based on the prior knowledge that the duration of EEG microstates is 100ms, the kernel size of the first convolutional layer is set to... Step length This means that the convolution operation is performed along the time dimension. Here, Indicates the number of convolution kernels. This represents the sampling rate. The kernel size of the second convolutional layer is set to... Step length This is to study the interactions between different electrodes. This represents the number of electrode channels in the EEG. Batch-Normalization is then used to enhance the training process and alleviate overfitting. After activation by exponential linear units (ELUs), the learned representations are max-pooled every two data points (without overlap) along the time dimension to reduce computational complexity in case of overfitting. Finally, the feature maps from the time-frequency convolution are arranged into tokens (the basic units of the model input / output) and combined with learnable positional encodings for use by subsequent modules. The shallow convolutional feature encoding module is defined as follows:

[0068]

[0069] in, This represents the output value of the shallow convolutional feature encoding module, which in turn represents the time-series features of the preprocessed modality data. This represents the input to the convolutional feature encoder, i.e., the preprocessed modal data. Represents the convolution function. This represents the batch normalization function. Represents the activation function of the exponential linear unit. Represents the max pooling function. This represents the rearrangement function, or rearrangement operation, which rearranges the dimensions of the EEG features processed by the convolution module to fit the input of the subsequent Transformer module. This represents a learnable positional encoding used to provide the model with positional information of elements in a sample sequence.

[0070] In deep learning, especially in natural language processing, models such as RNNs, LSTMs, and Transformers are often used to capture the contextual information of input sequences. Because neural activity is highly temporally correlated, the temporal context of EEG signals is analogous to contextual understanding in language models. Utilizing the encoded shallow time-frequency representation, we use a cascaded Transformer module for coarse- and fine-grained feature extraction. One Transformer-based branch captures the correlation between input tokens, while a CNN-based branch extracts local contextual information. This CCFT module repeats... Secondary output .

[0071] For each CFT module (Transformer module for coarse and fine granular feature extraction) of the CCFT module, a max pooling layer with a pooling size of 2 and a stride of 2 is used to reduce the input label. The feature dimensions are then used. Then, there are two parallel branches: one based on Transformer and the other based on CNN.

[0072] The Transformer-based branch projects the max-pooled tokens through a linear projection layer into three identical copies: query( ), key( ) and value( ), query( ) is the query vector, used To represent; key( ) is the key vector, used This means that value( ) is a value vector, used This is represented by a scaled dot product for evaluation. and The correlation between different tokens is analyzed. After calculating the attention score using the Softmax layer, a weighted average is obtained by performing a dot product. This self-attention process can be defined as:

[0073]

[0074]

[0075] in, Represents the query vector. Represents the key vector. Represents a value vector. This represents the input to the CFT module, specifically the input to the max-pooling layer within the CFT module. This represents the projection matrix of the query / key / value. Indicates based on , , The results of single-head attention calculation, This represents the feature dimension after the key vector is projected. This represents the Softmax function.

[0076] The CFT module employs a multi-head self-attention mechanism, enabling different attention heads to focus on different parts of the coarse-grained temporal encoding, thereby enhancing the diversity of representations. Input labels are implemented through their own independent self-attention modules, and the results are concatenated as the output of a Transformer-based branch. This process can be defined as:

[0077]

[0078] in, This represents the function for the multi-head attention mechanism. Indicates the index of the attention head. , This represents the number of attention heads in a multi-head attention mechanism. Indicates attention head The output matrix, This represents the output matrix of the attention head at index 0. This represents the output matrix of the attention head with index 1. Indicates that the index is The output matrix of the attention head, Indicates the first A query matrix with attention heads Indicates the first The key matrix of each attention head. Indicates the first The value matrix of each attention head, This represents the attention function. Understandably, in the Transformer architecture, the head is a parallel computational unit in the multi-head attention mechanism, used to learn an attention pattern.

[0079] Following MHA, the previously max-pooled input tokens are summed with the MHA result, and a feedforward neural network is introduced to enhance the fitting ability. This network consists of two fully connected layers. Therefore, the coarse-grained embedding of the CFT module is defined as:

[0080]

[0081] in, This represents the coarse-grained features output by the CFT module. This represents a feedforward neural network. The representation layer normalization function.

[0082] To learn fine-grained temporal embedding features, this embodiment designs a CNN-based branch that operates in parallel with a Transformer-based branch. After a Dropout layer, a one-dimensional CNN operation is performed along the temporal dimension, with a small kernel size of (1, 11), followed by a batch normalization layer, an ELU activation function, and a max-pooling layer. Therefore, the fine-grained embedding of the CFT module... Defined as:

[0083]

[0084] Combining coarse-grained and fine-grained temporal representations, the final output of the CFT module is defined as:

[0085]

[0086] in, This represents the final output of the CCFT module. This represents the fine-grained features output by the CFT module. This represents the random deactivation function.

[0087] The identity embedding module is used to embed... Mapping to the identity space, retrieving the user's personalized embedding vector based on the mapping result, and fusing the personalized embedding vector with... Based on personalized embedding vectors and The fusion results are used to assess the individual's brain state.

[0088] To accurately predict individual brain states, this disclosure designs an identity embedding module as a two-branch classifier. The identity embedding branch introduces individual differences and supplements personalized information that may be lost in the general features of the cascaded Transformer module extracted by coarse and fine granular features. Given the spatial, coarse-grained, and fine-grained temporal representations from the CCFT module, the identity embedding module uses two fully connected layers to map the feature space to the identity space, thereby predicting the probability of the subject's (i.e., user's) identity. Based on the predicted identity, the personalized embedding vector of the corresponding user is retrieved from a predefined embedding matrix (the identity embedding vector acts as a dynamic, personalized feature filter or amplifier, which can be understood as telling the model: "For this person, pay more attention to feature A and less attention to feature B."), where each row of the matrix represents an individual's identity feature. During training, the embedding matrix is ​​updated to represent personalized patterns associated with different individual brain states. The personalized embedding vector is fused with the general features (the output of the first fully connected layer of the identity embedding module) to form a personalized spatial state vector, which is then input into the second fully connected layer and evaluated for individual brain state through an activation function. The classifier is defined as follows:

[0089]

[0090] in, Indicates will The mapping result to the identity space, i.e., the identity prediction value, Represents a fully connected transformation function; Indicates an embedded function; express Activation functions introduce nonlinearity.

[0091] The loss function of this framework consists of two parts: identity prediction cross-entropy and brain state assessment cross-entropy, which are defined as follows:

[0092]

[0093] in, Indicates the loss value. This indicates the number of samples in each batch during model training. This indicates the index of the calculated sample. Index representing brain states, Indicates the number of brain states. Indicates the number of participants. This indicates the actual value of the brain's state. Indicates brain state assessment value, Indicates the actual value of the identity. This represents the predicted identity value.

[0094] The EEG signal classification model learns to decompose the input signal into two parts: a task-related general pattern (extracted by CCFT): "This is a feature of fatigue"; and an individual-related offset pattern (provided by the embedding matrix): "This is Zhang San's unique way of expressing fatigue." In fatigue assessment, some people show a significant increase in frontal alpha wave power when fatigued, while others show an increase in occipital alpha waves. General features may contain both frontal and occipital alpha information, causing confusion. Identity embedding adaptively strengthens the feature channels most sensitive to the current individual and suppresses insensitive or interfering channels.

[0095] Through the above processing, an identity embedding matrix is ​​generated (the output values ​​of the identity embedding module are presented in matrix form), thereby optimizing the final evaluation performance.

[0096] See Figure 6 This disclosure provides an AI-driven transcranial photobiological modulation intelligent control device, including:

[0097] The biosensor module is used to acquire multiple sets of resting state data of the transcranial photobiomodulation device user and to acquire stimulation state data of the controllable external stimuli of the transcranial photobiomodulation device user.

[0098] The AI ​​chip is used to acquire resting state data and stimulation state data from the biosensor module, preprocess the state data to obtain preprocessed modal data, extract time-series features of the preprocessed modal data using a spatiotemporal convolutional network, interactively model the time-series features of different modalities using a Transformer structure with multi-head attention mechanism to obtain fused feature vectors, dynamically calculate the weights of each fused feature vector through a self-attention mechanism, and map the fused feature vectors to a multi-dimensional physiological state space to obtain a physiological feature map, which is stored as a personalized baseline model. The AI ​​chip is used to calculate the optimal operating parameters of the transcranial photobiomodulation device used by the user based on the personalized baseline model and through a deep reinforcement learning algorithm.

[0099] The optical modulation module is used to regulate the light waves of the transcranial photobiomodulation device according to the optimal operating parameters;

[0100] The power management module is used to supply power to the working area of ​​the transcranial photobiomodulation device according to the optimal operating parameters;

[0101] The biosensor module is also used to monitor the physiological signals of the user of the transcranial photobiomodulation device in real time when the transcranial photobiomodulation device is operating at optimal parameters.

[0102] The AI ​​chip is also used to obtain the physiological signals monitored in real time by the biosensor module, and to analyze the degree to which the physiological signals deviate from their normal range using a deep reinforcement learning algorithm and dynamically adjust the optimal operating parameters accordingly.

[0103] The operating parameters include information on the irradiated area, light intensity, wavelength, and pulse frequency. The power management module supplies power to the working area of ​​the transcranial photobiomodulation device based on the irradiated area information, and the optical modulation module regulates the light waves of the transcranial photobiomodulation device based on the light intensity, wavelength, and pulse frequency information.

[0104] The AI ​​chip described herein is a miniaturized AI chip. The AI ​​chip disclosed herein can process AI algorithms directly on a local device via edge computing. Preferably, the AI ​​chip disclosed herein adopts a cloud-edge collaborative architecture.

[0105] The device disclosed herein adopts an integrated hardware and software architecture with cloud-edge collaboration. Through modular hardware integration and layered software deployment, specifically, sensors continuously monitor physiological responses and input the data into the next round of optimization, achieving high real-time treatment control with a closed-loop latency of <50 ms. The device consists of two main components: hardware and software.

[0106] The hardware components include an AI chip, an optical modulation module, a biosensor module, and a power management module.

[0107] The miniaturized AI chip utilizes a low-power, high-performance embedded AI chip (such as an ARM Cortex-A series + NPU or a RISC-V AI SoC), possessing ≥2 TOPS (Tera Operations Per Second) of AI computing power. It supports mixed-precision operations of half-precision floating-point numbers (FP16) and 8-bit integers (INT8), balancing computational accuracy with power consumption control. Integrated hardware acceleration modules (CNN accelerator, Transformer inference engine) are used for rapid multimodal feature extraction and reinforcement learning inference. A built-in hardware security module (HSM) provides data encryption and privacy protection, complying with medical information security standards and HIPAA (Health Insurance Portability and Accountability Act) / GDPR (General Data Protection Regulation). The AI ​​chip provides computing power either on its own or through a combination of itself and the cloud.

[0108] The optical modulation module includes a light supply system, a real-time temperature monitoring system, and a temperature control system (specifically a closed-loop constant temperature control system). Further, it also includes a computer control system, a real-time data recording system, and an external power supply. The light supply system features evenly distributed LED light sources, controlled by different brain regions: the working areas corresponding to the left frontal lobe, right frontal lobe, left parietal lobe, left temporal lobe, left occipital lobe, right parietal lobe, right temporal lobe, and right occipital lobe. The real-time temperature monitoring system is located in the central circuit boards of different regions, facilitating real-time monitoring of temperature changes caused by the activation of the light sources. It also includes temperature monitoring of the AI ​​chip. The closed-loop constant temperature control system continuously supplies low-temperature air via an external cooling device. When the detected temperature exceeds a constant temperature, cooling is initiated until the temperature reaches a stable level. The system sets the temperature, which can be achieved through a built-in heat pipe and low-noise fan, and adjusts the temperature in real time to prevent overheating of the light source. Further measures include preventing overheating of the AI ​​chip. This closed-loop system uses real-time temperature monitoring to maintain a constant temperature. The computer control system updates or controls the internal program code of the helmet via Arduino software (code functions can be enabled simultaneously or used to control the light source separately), including controlling the operation of the lighting system, real-time temperature monitoring system, temperature control system, and real-time data recording system. The real-time data recording system can collect and record the lighting intervention process in real time, forming a complete file including data recorded during the operation of the lighting system, temperature monitoring system, and temperature control system. The power management module controls the external power supply, which provides partial / full power to the lighting system, temperature monitoring system, temperature control system, computer control system, and real-time data recording system.

[0109] It is understood that in other embodiments, the light supply system may not include all of the above-mentioned working areas. For example, the LED light source does not cover the entire head, and the arrangement shape of the LED light source is not limited to that shown in this embodiment (the LED beads or lasers can be arranged in a circular arrangement, a rectangular arrangement, or other arbitrary arrangement combinations, etc.). For example, the LED light source may only be provided in the working area corresponding to the left frontal lobe and / or the working area corresponding to the right frontal lobe (the above eight areas can be activated in any combination).

[0110] Understandably, when using it, you can control the operation by dividing the work area into zones so that some work areas work while others do not.

[0111] The biosensor module is the multimodal signal acquisition device mentioned above. In this embodiment, the EEG sensor features a high input impedance amplifier (>1 GΩ), a 24-bit analog-to-digital converter (ADC), and a sampling rate of 500–1000 Hz, supporting multi-channel acquisition. The HRV / PPG sensor uses red light + infrared dual-wavelength PPG acquisition with a sampling rate ≥200 Hz for accurate calculation of the RR interval (a complete cardiac cycle). The EDA sensor measures skin conductance using the constant voltage method, with a resolution of 0.01 microseconds and a sampling rate ≥50 Hz.

[0112] The power management module includes a battery module and a control module, and can use a high-energy-density lithium polymer battery (3.7 V, 2000 mAh), supporting continuous operation for >6 hours. An integrated DC-DC buck-boost converter module stabilizes the power supply to the optical module and AI chip, with ripple <10 mV, reducing power supply noise interference to the sensor. Specifically, it uses a 3.7 V lithium polymer battery + PMU, supports USB-C charging, and has overcurrent / overvoltage / overtemperature protection functions.

[0113] Communication between modules, and between modules and AI chips, is achieved through a built-in wireless communication unit: it supports Wi-Fi 6 and BLE 5.3 to achieve data synchronization and can also be used for remote control.

[0114] The transcranial photobiological modulation device has a built-in thermal management circuit and temperature sensor to monitor the temperature of the chip and light source in real time to prevent overheating.

[0115] The software adopts a cloud-edge collaborative architecture, with the edge performing real-time control and main AI inference, and the cloud used for offline model training and long-term data analysis.

[0116] (a) Edge: Run a Real-Time Operating System (RTOS) or lightweight Linux (such as Yocto) to ensure task scheduling latency <1ms. Deploy a multimodal data acquisition and fusion algorithm module (temporal convolution + Transformer) for real-time physiological state estimation. Run a reinforcement learning inference engine to adjust tPBM parameters based on real-time state and directly drive the optical modulation module to perform stimulation. Include a safety policy daemon process to immediately cut off optical stimulation when an anomaly is detected (such as excessively high temperature or abnormal signal).

[0117] (b) Cloud-based: The cloud server is used for offline model training and long-term trend analysis. Anonymized multimodal data of subjects is uploaded periodically for deep model retraining and personalized baseline model updates. A data visualization platform is provided for physicians to view the long-term neurophysiological trends of patients. The long-term trend analysis refers to the fact that physiological data is recorded after each tPBM treatment, and the cloud can analyze the changes in certain physiological data after multiple treatments.

[0118] (c) Cloud-Edge Collaboration Mechanism: All real-time inference and decision-making during the operation of the transcranial photobiological modulation device are completed at the edge, independent of the cloud, to ensure latency and privacy. The cloud provides update packages for personalized baseline models, which are securely synchronized with the AI ​​chip each time a network connection is established, ensuring continuous parameter optimization. Therefore, even in a network-free environment, it can still independently complete closed-loop parameter control.

[0119] (d) Closed-loop latency optimization strategy: To control the closed-loop latency to within 50 ms, this disclosure adopts the following measures in the hardware-software co-design: Multiple sensors use Direct Memory Access (DMA) to acquire data in parallel, reducing CPU interrupt burden. The edge algorithm uses a sliding window and incremental calculation method, calculating only the newly entered data segment in each sampling cycle to avoid full recalculation. The hardware NPU of the AI ​​chip is used to perform reinforcement learning inference, reducing latency related to CPU / GPU involvement. The optical modulation module is directly connected to the AI ​​chip via a high-speed SPI / I2C bus, with instruction transmission latency < 1 ms.

[0120] In this embodiment, the transcranial photobiomodulation device is taken as a head-mounted transcranial photobiomodulation device. See [link to example]. Figures 2 to 5 The headband features a lightweight ring-shaped support structure, weighing ≤ 250 g, suitable for various head shapes. It incorporates an adjustable telescopic arm and flexible silicone padding to ensure comfort during extended wear. The light supply system (i.e., the light source) is located at the innermost layer of the light supply system, allowing for precise location of the target brain region through accurate measurement, strict zoning, and modular control. Simultaneously, a real-time temperature monitoring system is located outside the light supply system. Outside this system are a closed-loop constant temperature control system and a multimodal signal acquisition device. Specifically, the electrodes of the EEG sensor, the photodetector of the PPG sensor, and the electrodes of the EDA sensor are distributed in different locations and connected to the AI ​​chip via FPC (Flexible Printed Circuit) cables to ensure signal transmission stability. In one specific embodiment, the integrated AI chip serves as a central processing unit (CPU / NPU integrated). The AI ​​chip also drives the optical modulation module and the power management module. Figure 7This is a cross-sectional view of a portion of the structure of a head-mounted transcranial photobiomodulation device. The figure illustrates the outer shell 11, lamp mounting points 12, air duct 13, inner shell 14, lamp assembly 15, air vents 16, and transparent cover 17. The transparent cover 17 is heat-resistant up to 80 degrees Celsius, radiation-resistant, and has several 0.4mm micro-ventilation holes evenly distributed on its surface. In this embodiment, a fully integrated design (sensing-computation-stimulation integration) is achieved based on the transcranial photobiomodulation device, eliminating reliance on remote servers, supporting home and mobile use, and offering high real-time performance.

[0121] The aforementioned AI-driven transcranial photobiological modulation intelligent control device achieves "personalized customization," solving the problem of treatment precision caused by individual differences; it improves feedback accuracy through multi-dimensional collaborative assessment; and it resolves the triple contradiction of latency, privacy, and portability through edge-cloud collaboration; it fully considers brain region specificity, temperature closed-loop regulation, and real-time monitoring, resulting in higher safety.

[0122] This disclosure provides a storage medium storing a computer program. When executed by a processor, the computer program can perform the following steps: preprocessing state data to obtain preprocessed modal data; extracting time-series features of the preprocessed modal data using a spatiotemporal convolutional network; interactively modeling the time-series features of different modalities using a Transformer structure with a multi-head attention mechanism to obtain fused feature vectors; dynamically calculating the weights of each fused feature vector using a self-attention mechanism; mapping the fused feature vectors to a multi-dimensional physiological state space to obtain a physiological feature map; storing the physiological feature map as a personalized baseline model; calculating the optimal operating parameters of a transcranial photobiomodulation device used by a user based on the physiological feature map using a deep reinforcement learning algorithm; obtaining real-time monitored physiological signals of the user of the transcranial photobiomodulation device; and analyzing the degree of deviation of the physiological signals from their normal range using a deep reinforcement learning algorithm based on the personalized baseline model and dynamically adjusting the optimal operating parameters accordingly. The status data includes multiple sets of resting status data of the transcranial photobiomodulation device user and stimulation status data of the transcranial photobiomodulation device user under controllable external stimulation.

[0123] In one specific embodiment, this disclosure is applied to assessing mental fatigue. The method of this disclosure is compared with state-of-the-art methods on two datasets (Dataset I and Dataset II). Dataset I consists of self-collected experimental data: in a controlled laboratory environment, 40 subjects (23 ± 1.75 years old) performed a 60-minute 2-back working memory task to induce mental fatigue, including subjective fatigue scores, objective performance, and EEG signals throughout the task. Dataset II is a publicly available dataset in the prior art; in a virtual reality driving environment, 27 subjects performed a 90-minute simulated driving task to assess cognitive fatigue during driving, including task performance and EEG signals throughout the task. The mental fatigue classification results on Dataset I show that this disclosure (referred to as the EEG-IECT model) outperforms all baseline methods. It achieves the highest accuracy (97.39%), macro-F1 score (97.37%), and Kappa score (96.08%). Compared to the second-best performing model, EEG-IECT (EEG Conformer: Convolutional Transformer for EEG Decoding and Visualization, a brain signal analysis model combining CNN and Transformer architectures for EEG decoding and visualization), this disclosure shows a 6.06% improvement in accuracy, a 6.14% improvement in macro-F1 score, and a 9.12% improvement in Kappa score. In contrast, EEGNet (EEGNet: A CompactConvolutional Network for Brain-Computer Interfaces, a lightweight convolutional neural network designed for EEG signal classification), LGGNet (Local-Global Graph Neural Network, an EEG analysis model based on spatiotemporal graph convolution), and EEG-ViT (EEG Vision Transformer, an EEG signal processing model based on a visual Transformer architecture), which are based on basic CNN, GNN, and Transformer architectures, achieved relatively lower scores, with an accuracy and macro-F1 score of approximately 80% and a Kappa score of approximately 70%. These methods focus on information at a single scale, granularity, or time period, limiting their ability to capture complex data patterns. In contrast, TSception (Temporal-Spectral Convolutional Network, an emotion recognition model based on EEG signals) and EEG-Conformer overcome the limitations of single structures through multi-level pattern fusion and structural combination, thus demonstrating superior performance.However, their ability to fully capture complex patterns remains limited when tackling more challenging classification problems. This disclosure achieves superior performance by integrating features at different scales, including spatiotemporal features, coarse-to-fine features, local-to-global features, and subject-specific features. These results highlight the importance of multi-level feature fusion for capturing complex EEG patterns and demonstrate the effectiveness of combining CNN and Transformer architectures to extract discriminative features.

[0124] This disclosure also evaluated the classification performance of each participant to compare the predictive power of the models. The method proposed in this disclosure achieved superior performance, with all 40 participants achieving a prediction accuracy of over 90%, and 80% of the participants achieving an accuracy of over 95%. In contrast, other models showed varying degrees of accuracy decline, with some participants falling into the 80%, 70%, 60%, or even below 60% accuracy range, highlighting the superiority of this disclosure in individual-level prediction.

[0125] To further validate the effectiveness of this disclosure, a comprehensive evaluation of driver fatigue detection was conducted on Dataset II. The binary classification results revealed trends consistent with those observed in Dataset I. Notably, EEG-IECT achieved excellent performance with an accuracy of 94.79%, a macro-F1 score of 94.47%, and a Kappa score of 88.93%, demonstrating its robustness and reliability in EEG-based fatigue decoding (i.e., fatigue classification). Compared to the second-best performing model, EEG-IECT significantly improved accuracy by 2.18%, macro-F1 score by 2.3%, and Kappa score by 4.6%. Similar to Dataset I, models such as EEGNet, LGGNet, and EEG-ViT exhibited relatively lower performance, with accuracies between 80% and 90%. In contrast, TSception and EEG-Conformer, utilizing multi-layer information fusion, performed slightly better, exceeding 90%. These findings further highlight the effectiveness of combining CNN and Transformer architectures to design multi-level information fusion mechanisms, which can robustly capture complex patterns and significantly improve classification performance. For EEG-IECT, 91% of participants achieved prediction accuracy of 90% or higher, with only one participant falling within the 85%-90% range. In contrast, other models exhibited a wider performance distribution, with a greater number of participants achieving prediction accuracy below 90%.

[0126] This disclosure discloses an AI-driven intelligent control method, device, and storage medium for transcranial photobiomodulation. It utilizes a spatiotemporal convolutional network to extract time-series features from the resting state data of a specific transcranial photobiomodulation device user and the stimulation state data of controllable external stimuli. A Transformer structure with multi-head attention is used to interactively model the time-series features of different modalities to obtain a fused feature vector. The weights of each fused feature vector are dynamically calculated using a self-attention mechanism. The fused feature vectors are mapped to a multi-dimensional physiological state space to obtain a physiological feature map. This physiological feature map is stored as a personalized baseline model. Based on the physiological feature map, a deep reinforcement learning algorithm is used to calculate the optimal operating parameters of the transcranial photobiomodulation device used by the user. The transcranial photobiomodulation device executes the optimal operating parameters. The physiological signals of the transcranial photobiomodulation device user are monitored in real time. Based on the personalized baseline model, a deep reinforcement learning algorithm is used to analyze the degree to which the physiological signals deviate from their normal range and dynamically adjust the optimal operating parameters accordingly. This analysis and calculation achieves individualized, real-time adjustment of stimulation parameters, fully considering individual user differences, eliminating the need for group average threshold information, and resulting in high accuracy and good adjustment effect. This also improves the effectiveness of transcranial photobiological modulation.

[0127] This disclosure achieves closed-loop regulation of "perception-analysis-stimulation" through real-time physiological feedback to match symptom fluctuations. For example, when EEG detects prefrontal hypoactivation, it automatically increases the intensity of a specific wavelength of light, while switching to a high-frequency pulse mode during the acute phase of anxiety (when HRV is reduced).

[0128] This disclosure overcomes the limitations of population parameters by establishing regulatory rules based on patient-specific baselines. In practical applications, a machine learning phase of 1-2 weeks (i.e., step 4 above) is used to construct control parameters, including optimal response wavelengths and sensitive brain regions.

[0129] The reward function disclosed herein considers the symptom relief speed of the transcranial photobiomodulation device user, the user's comfort, the device's power consumption, and its temperature control. This means it doesn't rely solely on changes in a single physiological indicator to adjust parameters, but also considers subjective feelings, all of which are important bases for adaptive parameter adjustment. This achieves optimization from a single treatment goal to a comprehensive experience, improving patient compliance.

[0130] This disclosure can be used to treat neurodevelopmental disorders such as schizophrenia and ADHD, neurodegenerative diseases such as Alzheimer's disease and Parkinson's disease, and brain dysfunction diseases such as stroke and traumatic brain injury.

[0131] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0132] The embodiments described above are merely illustrative of several implementations of this disclosure, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this disclosure, and these all fall within the protection scope of this disclosure. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. An AI-driven transcranial photobiological modulation intelligent control method, characterized in that, Includes the following steps: Step 1: Obtain multiple sets of resting state data of the transcranial photobiomodulation device user, and obtain the stimulation state data of the controllable external stimulation of the transcranial photobiomodulation device user. Step 2: The AI ​​chip preprocesses the state data obtained in Step 1 to obtain preprocessed modal data. It then uses a spatiotemporal convolutional network to extract the time-series features of the preprocessed modal data. A Transformer structure with a multi-head attention mechanism is used to interactively model the time-series features of different modalities to obtain fused feature vectors. The weights of each fused feature vector are dynamically calculated using a self-attention mechanism. The fused feature vectors are then mapped to a multi-dimensional physiological state space to obtain a physiological feature map. This physiological feature map is stored as a personalized baseline model. Step 3: The AI ​​chip calculates the optimal operating parameters of the transcranial photobiological modulation device used by the user based on the physiological feature map and through a deep reinforcement learning algorithm. Step 4: The transcranial photobiological modulation device executes the optimal operating parameters and monitors the physiological signals of the user in real time. The AI ​​chip analyzes the degree of deviation of the physiological signals from their normal range based on the personalized baseline model and uses deep reinforcement learning algorithms, and dynamically adjusts the optimal operating parameters accordingly. The state space of the deep reinforcement learning algorithm is composed of fused feature vectors, the action space is a set of working parameters, and the reward function is represented in a weighted form. The reward function takes into account the symptom relief speed of the transcranial photobiomodulation device user, the comfort of the transcranial photobiomodulation device user, the power consumption of the transcranial photobiomodulation device, and the temperature control of the transcranial photobiomodulation device. The resting state data includes EEG signals, heart rate variability signals, and skin surface potential change signals in the resting state; the stimulation state data of the controllable external stimulus includes: EEG signals under cognitive task stimulation only, heart rate variability signals under cognitive task stimulation only, skin surface potential change signals under cognitive task stimulation only, EEG signals under emotionally induced stimulation only, heart rate variability signals under emotionally induced stimulation only, skin surface potential change signals under emotionally induced stimulation only, EEG signals under sensory stimulation only, heart rate variability signals under sensory stimulation only, and skin surface potential change signals under sensory stimulation only. The AI ​​chip analyzes the degree to which the physiological signal deviates from its normal range based on a personalized baseline model and a deep reinforcement learning algorithm, and dynamically adjusts the optimal operating parameters accordingly. Specifically, the AI ​​chip performs similarity matching between the physiological signal and the personalized baseline model to determine the degree to which the current state deviates from the normal range. The AI ​​chip then adjusts the optimal operating parameters of the transcranial photobiological modulation device based on the degree to which the current state deviates from the normal range using a deep reinforcement learning algorithm. This allows the updated optimal operating parameters to be used to re-execute step 4.

2. The AI-driven transcranial photobiological modulation intelligent control method according to claim 1, characterized in that, Step 2 is achieved through an EEG signal classification model based on identity embedding and cascaded Transformer architecture. The EEG signal classification model based on identity embedding and cascaded Transformer architecture includes a shallow convolutional feature encoding module, a cascaded Transformer module for coarse and fine-grained feature extraction, and an identity embedding module. The shallow convolutional feature encoding module is used to extract short-term dependencies and local spatial dynamic features of each preprocessed modality data using a spatiotemporal convolutional network to obtain time series features. The shallow convolutional feature encoding module is defined as follows: ， in, This represents the output value of the shallow convolutional feature encoding module. This represents the input to the convolutional feature encoder. Represents the convolution function. This represents the batch normalization function. Represents the activation function of the exponential linear unit. Represents the max pooling function. This represents the rearrangement function. This represents a learnable location code; The cascaded Transformer modules for coarse and fine granular feature extraction are used to interactively model time series features of different modalities using a Transformer structure with a multi-head attention mechanism to obtain a fused feature vector; each of the cascaded Transformer modules for coarse and fine granular feature extraction employs a multi-head self-attention mechanism, satisfying: ， in, This represents the function for the multi-head attention mechanism. Represents the query vector. Represents the key vector. Represents a value vector. This represents the number of attention heads in a multi-head attention mechanism. Indicates the index of the attention head. , Represents attention head The output matrix, This represents the output matrix of the attention head at index 0. This represents the output matrix of the attention head with index 1. Indicates that the index is The output matrix of the attention head, Indicates the first A query matrix with attention heads Indicates the first The key matrix of each attention head. Indicates the first The value matrix of each attention head, Represents the attention function; The coarse-grained embedding of the cascaded Transformer module for coarse and fine-grained feature extraction is defined as follows: ， in, This represents the coarse-grained features output by the Transformer module for coarse-grained feature extraction. This represents a feedforward neural network. Indicates the layer normalization function; The fine-grained embedding of the cascaded Transformer module for coarse and fine-grained feature extraction Defined as: ， The output of the cascaded Transformer module for coarse and fine granular feature extraction is defined as follows: ， in, This represents the final output of the cascaded Transformer modules for coarse-grained and fine-grained feature extraction. This represents the fine-grained features output by the Transformer module, which extracts coarse-grained and fine-grained features. Represents the random deactivation function; The identity embedding module is used to embed... Mapping to the identity space, retrieving the user's personalized embedding vector, and fusing the personalized embedding vector with... Conduct an individual brain state assessment to ensure that the following conditions are met: ， in, Indicates will The mapping result to the identity space, Represents a fully connected transformation function. Indicates an embedded function. express Activation function.

3. The AI-driven transcranial photobiological modulation intelligent control method according to claim 1, characterized in that, The training of the personalized baseline model is achieved through the cloud; the preprocessing includes bandpass filtering, denoising, artifact removal, and normalization; the optimal operating parameters include the optimal combination of multiple operating parameters, including luminous region information, light intensity information, wavelength information, and pulse frequency information.

4. An AI-driven transcranial photobiological modulation intelligent control device, characterized in that, include: The biosensor module is used to acquire multiple sets of resting state data of the transcranial photobiomodulation device user and to acquire stimulation state data of the controllable external stimuli of the transcranial photobiomodulation device user. The AI ​​chip is used to acquire resting state data and stimulation state data from the biosensor module, preprocess the state data to obtain preprocessed modal data, extract time-series features of the preprocessed modal data using a spatiotemporal convolutional network, interactively model the time-series features of different modalities using a Transformer structure with multi-head attention mechanism to obtain fused feature vectors, dynamically calculate the weights of each fused feature vector through a self-attention mechanism, and map the fused feature vectors to a multi-dimensional physiological state space to obtain a physiological feature map, which is stored as a personalized baseline model. The AI ​​chip is used to calculate the optimal operating parameters of the transcranial photobiomodulation device used by the user based on the personalized baseline model and through a deep reinforcement learning algorithm. The optical modulation module is used to regulate the light waves of the transcranial photobiomodulation device according to the optimal operating parameters; The power management module is used to supply power to the working area of ​​the transcranial photobiomodulation device according to the optimal operating parameters; The biosensor module is also used to monitor the physiological signals of the user of the transcranial photobiomodulation device in real time when the transcranial photobiomodulation device is operating at optimal parameters. The AI ​​chip is also used to obtain the physiological signals monitored in real time by the biosensor module, and to analyze the degree of deviation of the physiological signals from their normal range using a deep reinforcement learning algorithm and dynamically adjust the optimal operating parameters accordingly. The state space of the deep reinforcement learning algorithm is composed of fused feature vectors, the action space is a set of working parameters, and the reward function is represented in a weighted form. The reward function takes into account the symptom relief speed of the transcranial photobiomodulation device user, the comfort of the transcranial photobiomodulation device user, the power consumption of the transcranial photobiomodulation device, and the temperature control of the transcranial photobiomodulation device. The resting state data includes EEG signals, heart rate variability signals, and skin surface potential change signals in the resting state; the stimulation state data of the controllable external stimulus includes: EEG signals under cognitive task stimulation only, heart rate variability signals under cognitive task stimulation only, skin surface potential change signals under cognitive task stimulation only, EEG signals under emotionally induced stimulation only, heart rate variability signals under emotionally induced stimulation only, skin surface potential change signals under emotionally induced stimulation only, EEG signals under sensory stimulation only, heart rate variability signals under sensory stimulation only, and skin surface potential change signals under sensory stimulation only. The AI ​​chip is used to analyze the degree to which the physiological signal deviates from its normal range based on a personalized baseline model and a deep reinforcement learning algorithm, and dynamically adjust the optimal operating parameters accordingly. Specifically, the AI ​​chip is used to perform similarity matching between the physiological signal and the personalized baseline model to determine the degree to which the current state deviates from the normal range. The AI ​​chip is used to adjust the optimal operating parameters of the transcranial photobiomodulation device according to the degree to which the current state deviates from the normal range using a deep reinforcement learning algorithm. The updated optimal operating parameters are then used to restart the function of the biosensor module to monitor the physiological signals of the user of the transcranial photobiomodulation device in real time when the transcranial photobiomodulation device is operating at the optimal operating parameters.

5. The AI-driven transcranial photobiological modulation intelligent control device as described in claim 4, characterized in that, The operating parameters include information on the irradiated area, light intensity, wavelength, and pulse frequency. The power management module supplies power to the working area of ​​the transcranial photobiomodulation device based on the irradiated area information, and the optical modulation module regulates the light waves of the transcranial photobiomodulation device based on the light intensity, wavelength, and pulse frequency information.

6. The AI-driven transcranial photobiological modulation intelligent control device as described in claim 4, characterized in that, The training of the personalized baseline model is achieved through the cloud; the AI ​​chip provides computing power either on its own or through a combination of itself and the cloud; the biosensor module includes an electroencephalogram (EEG) sensor, a heart rate variability sensor, and a skin conductance sensor; the optical modulation module includes a light supply system for regional light emission, a real-time temperature monitoring system for real-time temperature monitoring of the light supply system, and a temperature control system for constant temperature control of the light supply system based on the temperature monitoring results of the real-time temperature monitoring system; the transcranial photobiomodulation device is a head-mounted transcranial photobiomodulation device.

Images