TaVNS closed-loop regulation and control method for high-inflammation ARDS

Abstract

The invention belongs to the technical field of taVNS parameter adjustment, and relates to a taVNS closed-loop regulation and control method for high-inflammation ARDS, and the method comprises the steps: obtaining a neuron activation index through employing a dual-frequency auricular conchae impedance analysis technology, inputting the time series data of the neuron activation index into an inflammation risk prediction model, obtaining an inflammation risk score, and achieving the quantitative evaluation of an inflammation state; in combination with a comparison result based on an inflammation risk score and a set threshold value, a corresponding frequency control parameter is dynamically output in a multi-decision mode, and the current intensity is adjusted in combination with a real-time neuron activation index, so that a stimulation parameter can be adjusted in real time along with the change of an inflammation state; after stimulation is applied, by monitoring the neuron activation index change rate and the heart rate variability high-power index, the stimulation frequency is corrected within the preset time, a closed loop of state perception, parameter adjustment and effect feedback to parameter correction is formed, it is ensured that the stimulation parameters are accurately matched with the inflammation state all the time, and the stimulation effect is improved. Therefore, the target of dynamically adjusting parameters according to the inflammation state is achieved.

Description

Technical Field

[0001] The present application belongs to the field of taVNS parameter regulation technology, and more specifically, relates to a taVNS closed-loop regulation method for highly inflammatory ARDS. Background Art

[0002] Acute respiratory distress syndrome (ARDS) is a severe lung disease characterized by fluid filling of the alveoli, which impairs oxygen exchange and leads to multiple organ failure, with a mortality rate as high as 40%. Hyperinflammation is a key factor in ARDS patients, as excessive inflammatory responses can exacerbate lung damage. Currently, treatment options for ARDS are limited, primarily relying on mechanical ventilation and medication. However, these methods cannot directly regulate the inflammatory response and have limited therapeutic efficacy.

[0003] Transcutaneous vagus nerve stimulation (taVNS) is an emerging treatment approach that stimulates the vagus nerve to regulate the autonomic nervous system, thereby influencing inflammatory responses. However, existing taVNS methods typically use fixed frequencies and current intensities, which cannot be adjusted to the patient's specific inflammatory status. Furthermore, the lack of real-time feedback mechanisms for adjusting stimulation parameters can lead to poor therapeutic effects and may even cause side effects.

[0004] Therefore, how to dynamically adjust the taVNS stimulation parameters according to the patient's inflammatory status and achieve precise regulation has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] The present invention provides a taVNS closed-loop control method for highly inflammatory ARDS, which aims to solve the technical problem that the current taVNS method cannot dynamically adjust parameters according to the inflammatory state.

[0006] The taVNS closed-loop regulation method for highly inflammatory ARDS includes the following steps:

[0007] S1. Use dual-frequency concha impedance analysis to simultaneously measure low-frequency and high-frequency impedance parameters. Calculate the low-frequency impedance change rate and high-frequency phase difference based on the measured low-frequency and high-frequency impedance parameters. Then, use the low-frequency impedance change rate and high-frequency phase difference to derive the neuronal activation index.

[0008] S2. sorting the continuously collected neuron activation index values ​​in time series to form neuron activation index time series data of a predetermined time length;

[0009] S3. Inputting the generated neuronal activation index time series data into a lightweight inflammation risk prediction model, and processing the data based on the inflammation risk prediction model to obtain an inflammation risk score;

[0010] S4. Setting multiple thresholds and corresponding multiple decision-making methods, comparing the inflammation risk score with the set thresholds, determining the decision-making method corresponding to the verification risk score, outputting corresponding frequency control parameters based on the decision-making method; and adjusting the current intensity according to the real-time neuron activation index;

[0011] S5. After applying the stimulation, monitor the change rate index of the neuronal activation index and the high-frequency power index in the heart rate variability, and modify the stimulation frequency based on the changes in the indexes within a predetermined time.

[0012] The present invention adopts dual-frequency concha impedance analysis technology to obtain the neuron activation index, which is used as the basic indicator reflecting the inflammatory state. The neuron activation index time series data is then used as the input of the inflammatory risk prediction model. The inflammatory risk score is obtained based on the inflammatory risk prediction model to achieve a quantitative assessment of the inflammatory state. Combined with the comparison result based on the inflammatory risk score and the set threshold, the corresponding frequency control parameters are dynamically output through a multi-decision method, and the current intensity is adjusted in combination with the real-time neuron activation index, so that the stimulation parameters can be adjusted in real time with the changes in the inflammatory state. Finally, after the stimulation is applied, the stimulation frequency is corrected within a predetermined time by monitoring the change rate of the neuron activation index and the high-power index of heart rate variability, forming a closed loop of state perception, parameter adjustment, and effect feedback to parameter correction, ensuring that the stimulation parameters are always accurately matched with the inflammatory state, thereby achieving the goal of dynamically adjusting parameters according to the inflammatory state.

[0013] Preferably, in step S1, a dual-frequency impedance sensor is attached to the concha area, and an AC bridge method is used to simultaneously measure the amplitude and phase of the impedance at 1 kHz and 10 kHz by applying an AC signal.

[0014] Preferably, after synchronously measuring the low-frequency and high-frequency impedance parameters in step S1, the following steps are included:

[0015] Data preprocessing: Use filtering algorithm to remove noise from the collected low-frequency and high-frequency impedance parameters, and then normalize the denoised data;

[0016] Calculation of low-frequency impedance change rate: Calculate the difference between the low-frequency impedance values ​​at the current moment and the previous moment, and then divide it by the time interval to obtain the impedance change rate;

[0017] High-frequency phase difference calculation: Calculate the difference between the high-frequency phase value at the current moment and the previous moment to obtain the phase change;

[0018] Calculation of neuronal activation index: The neuronal activation index was obtained by dividing the low-frequency impedance change by the high-frequency phase difference.

[0019] Preferably, the inflammation risk prediction model includes an input layer, a dynamic sparse convolution module, a self-attention mechanism module, a feature fusion module and an output module;

[0020] Inputting the normalized neuron activation index time series data into the model through the input layer;

[0021] The dynamic sparse convolution module calculates the local change rate of the data based on the input neuron activation index time series data, and dynamically adjusts the sparsity factor according to the local change rate, and controls the sparsity factor within a predetermined range. Then, a dynamic sparse convolution kernel is generated based on the adjusted sparsity factor. Then, a convolution operation is performed on the input neuron activation index time series data based on the generated dynamic sparse convolution kernel to obtain a feature map after sparse convolution. The sparsity factor and dynamic sparse convolution kernel parameters are dynamically optimized based on back propagation.

[0022] The self-attention mechanism divides the feature map after sparse convolution into multiple local blocks, calculates the attention weight in each local block, and uses multi-head attention. Each head focuses on a different feature dimension. Finally, the outputs of each head are spliced ​​and integrated through a linear layer to obtain a feature map processed by self-attention.

[0023] The feature fusion module collects feature maps after sparse convolution and self-attention processing, and fuses features through weighted summation to obtain a fused feature map; the fused feature map is input into the output model, and the inflammation risk score is obtained based on the output module processing.

[0024] Preferably, there is a residual connection between the dynamic sparse convolution module and the self-attention module, and between the self-attention module and the feature fusion module, wherein the residual connection adopts an identity mapping to ensure that the gradient can be smoothly transmitted.

[0025] Preferably, the attention weight initialization method of the self-attention mechanism is as follows:

[0026] Perform fast Fourier transform on the input neuron activation index time series data to obtain a spectrum graph;

[0027] The energy of each frequency component is calculated based on the obtained spectrum, and the key frequency components are screened out according to the pre-set typical frequency range of inflammation-related signals;

[0028] Based on the screened key frequency components, the energy sum of the key frequency components in each local block is calculated, and the calculated energy sum is normalized to obtain the initial attention weight.

[0029] Preferably, the data processing step of the output layer includes:

[0030] The feature map output by the feature fusion module is reduced in dimensionality through global average pooling, and then the parameters are compressed through two fully connected layers. A Dropout layer is introduced in the fully connected layer to prevent overfitting. Finally, the output results of the fully connected layer are mapped through the Sigmoid activation function to obtain the inflammation risk score.

[0031] Preferably, three thresholds are set in step S4, including a first threshold and a second threshold, wherein the first threshold is smaller than the second threshold, and three levels of decision-making methods are set, including basic maintenance, intermediate transition and emergency suppression;

[0032] When the inflammation risk score is less than a first threshold, a basic maintenance strategy is adopted to output a preset basic maintenance frequency;

[0033] When the inflammation risk score is greater than or equal to a first threshold and less than a second threshold, an intermediate transition strategy is adopted to linearly adjust the frequency according to the real-time change of the inflammation risk score, wherein the linear adjustment frequency is linearly adjusted based on the starting frequency and the linear growth term;

[0034] When the inflammation risk score is greater than or equal to the second threshold, the emergency suppression frequency is adopted and the preset emergency suppression frequency is called, wherein the emergency suppression frequency is the preset highest stimulation frequency.

[0035] Preferably, the current intensity is adjusted as follows:

[0036] A real-time current intensity is generated based on a linear mapping relationship between baseline current intensity and neuron activation index; and an upper limit constraint of current intensity is constructed based on the baseline current intensity; wherein the baseline current intensity is obtained based on clinical trials.

[0037] The beneficial effects of the present invention include:

[0038] The present invention adopts dual-frequency concha impedance analysis technology to obtain the neuron activation index, which is used as the basic indicator reflecting the inflammatory state. The neuron activation index time series data is then used as the input of the inflammatory risk prediction model. The inflammatory risk score is obtained based on the inflammatory risk prediction model to achieve a quantitative assessment of the inflammatory state. Combined with the comparison result based on the inflammatory risk score and the set threshold, the corresponding frequency control parameters are dynamically output through a multi-decision method, and the current intensity is adjusted in combination with the real-time neuron activation index, so that the stimulation parameters can be adjusted in real time with the changes in the inflammatory state. Finally, after the stimulation is applied, the stimulation frequency is corrected within a predetermined time by monitoring the change rate of the neuron activation index and the high-power index of heart rate variability, forming a closed loop of state perception, parameter adjustment, and effect feedback to parameter correction, ensuring that the stimulation parameters are always accurately matched with the inflammatory state, thereby achieving the goal of dynamically adjusting parameters according to the inflammatory state. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments or descriptions of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0040] Figure 1 This is a flowchart of the overall steps provided by an embodiment of the present invention.

[0041] Figure 2 This is a simplified structural diagram of the inflammation risk prediction model provided in an embodiment of the present invention.

[0042] Figure 3 This is a brief block diagram of step S4 provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the technical problems, technical solutions and beneficial effects to be solved by this application more clearly understood, this application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0044] See also Figure 1 As shown in FIG, the taVNS closed-loop regulation method for highly inflammatory ARDS includes the following steps:

[0045] S1. Use dual-frequency concha impedance analysis to simultaneously measure low-frequency and high-frequency impedance parameters. Calculate the low-frequency impedance change rate and high-frequency phase difference based on the measured low-frequency and high-frequency impedance parameters. Then, use the low-frequency impedance change rate and high-frequency phase difference to derive the neuronal activation index.

[0046] Because the skin in the concha area is thin and has dense nerve distribution, it is suitable for non-invasive impedance measurement. Therefore, a dual-frequency impedance sensor is attached to the concha area. Using the dual-frequency impedance sensor and the AC bridge method, the amplitude and phase of the impedance are measured by applying an AC signal. During the measurement, the amplitude and phase of the impedance at 1 kHz and 10 kHz are measured simultaneously.

[0047] In this embodiment, two frequencies of 1kHz and 10kHz are selected, corresponding to different physiological tissue characteristics. The low frequency mainly reflects the activity of the vagus nerve, while the high frequency reflects more the state changes of the surrounding tissues. By synchronously measuring the impedance parameters of the two frequencies, the physiological changes in the concha area can be more comprehensively captured, thereby more accurately evaluating the activation state of neurons. Therefore, based on this, high-precision, non-invasive physiological indicator measurement is provided, which supports real-time evaluation of the activation state of neurons.

[0048] After obtaining the corresponding measured values, the measured values ​​are preprocessed, and the collected low-frequency and high-frequency impedance parameters are denoised using a filtering algorithm, and the denoised data are then normalized;

[0049] Then, the low-frequency impedance change rate is calculated based on the normalized data. That is, the difference between the low-frequency impedance values ​​at the current moment and the previous moment is calculated, and then divided by the time interval to obtain the impedance change rate. The specific expression is as follows:

[0050]

[0051] Where: Z 1k (t) represents the low-frequency impedance value at the current moment; Z 1k (t-Δt) represents the low-frequency impedance value at the previous moment; Δt represents the time interval; ΔZ 1k Indicates a time interval.

[0052] And calculate the high-frequency phase difference. By calculating the difference between the high-frequency phase values ​​at the current moment and the previous moment, the phase change is obtained. The specific expression is as follows:

[0053] Δφ 10k =φ 10k (t)-φ 10k (t-Δt);

[0054] Where: φ 10k (t) represents the high-frequency phase value at the current moment; φ 10k (t-Δt) represents the high-frequency phase value at the previous moment; Δφ 10k Represents high-frequency phase difference;

[0055] The activity intensity of the neuron is then quantified based on the low-frequency impedance change rate and the high-frequency phase difference to obtain the neuron activation index, which is expressed as follows:

[0056]

[0057] Where: NAI represents the neuronal activation index. When the NAI value increases, it means that the activity of pain neurons is enhanced, and when the NAI value decreases, it means that the activity is weakened.

[0058] In this embodiment, the NAI value is collected once per second to form a 5-minute time series data window, and a sliding window technology is used to continuously update the data window to ensure real-time performance.

[0059] In this embodiment, since the inflammatory factors released by immune cells during the inflammation process can cross the blood-brain barrier, affect the function of neurons, change the excitability and signal conduction of neurons, and lead to the transitional activation or inhibition of neurons, the present invention adopts non-invasive dual-frequency auricle impedance analysis to realize the monitoring of neuronal activation index, and thus realize the monitoring of the severity of inflammation.

[0060] S2. sorting the continuously collected neuron activation index values ​​in time series to form neuron activation index time series data of a predetermined time length;

[0061] S3. Inputting the generated neuronal activation index time series data into a lightweight inflammation risk prediction model, and processing the data based on the inflammation risk prediction model to obtain an inflammation risk score;

[0062] See also Figure 2 As shown, the inflammation risk prediction model includes an input layer, a dynamic sparse convolution module, a self-attention mechanism module, a residual connection, a feature fusion module and an output module;

[0063] Input layer: The input data is the standardized NAI time series data with a shape of (300, 1), corresponding to 5 minutes of continuous sampling data with a sampling frequency of 1 time / second;

[0064] Dynamic sparse convolution module: By dynamically adjusting the sparsity of the convolution kernel, it can efficiently extract local features of NAI time series data while reducing the amount of computation. The specific processing logic is as follows:

[0065] First, the sparse factor is calculated. The local change rate of the data is calculated based on the standardized NAI time series data. The local change rate is calculated by calculating the mean of the absolute difference of adjacent data points through a sliding window, reflecting the local fluctuation intensity (local change rate). The specific expression is as follows:

[0066]

[0067] Where: x j represents the characteristic value at the jth time point; represents the average value of the data in the sliding window; w represents the size of the sliding window; i represents the time point; σ(i) represents the local change rate;

[0068] The sparse factor is dynamically adjusted according to the local change rate, and the range is [0.2, 0.8]. Assuming that the change rate σ(i) is in the range [0, 1], the expression of the sparse factor is as follows:

[0069] s(i)=0.2+0.6·σ(i);

[0070] Where: s(i) represents the sparse factor of the i-th point, ranging from [0.2, 0.8];

[0071] Introducing coefficient factors into the convolution layer to dynamically generate sparse convolution kernels:

[0072]

[0073] Where y(i) represents the output feature at position i; w(m) represents the weight of the convolution at position m; x(im) represents the feature value of the input data at position im; a(m) represents the coefficient activation function, which determines whether to activate the position according to the sparsity factor, and is defined as follows:

[0074]

[0075] The number of activation positions is determined by the coefficient factor, that is, the proportion of activated neurons is s(i), and the specific method is as follows:

[0076] In the position m of the convolution kernel, the position of the s(i) ratio is randomly selected, a(m) is set to 1, and the rest are set to 0. During the training process, according to the gradient information, the neurons that have a greater impact on the loss function are selected for activation, thereby improving the expressiveness of the model. The gradient method can avoid the problem of model instability caused by randomness, and at the same time improve the training efficiency and performance of the model; that is, the sparse factor is optimized through directional propagation to adapt it to the change rate of different data. During the optimization process, the sparse factor is adjusted as a learnable parameter to minimize the loss function.

[0077] Self-attention mechanism module: Capture the long-range dependencies of time series data through local attention and optimize the attention weights based on the characteristics of biomedical signals. The specific steps are as follows:

[0078] Based on the feature map output by the dynamic sparse convolution module, it is divided into local blocks, where every 5 consecutive data points are a block, for a total of 60 blocks; the features in each local block are linearly transformed to generate query, key, and value matrices; the d dimension of each local block is then divided into n heads, and the dimension of each head is

[0079] For each head, we use zoom-click to calculate attention. First, we calculate the attention score based on the dimension, query, and key matrix of each head. Then, we use the Softmax function to process the calculated attention score to obtain the attention weight. Then, we perform a weighted sum based on the attention weight and the value matrix to obtain the attention output of the i-th head. Then, we concatenate the outputs of the n heads to obtain the multi-head attention output. We perform a linear transformation on the concatenated output to obtain the feature map output by the attention mechanism module.

[0080] The attention weight initialization strategy of the attention mechanism module is as follows:

[0081] Based on the feature graph processed by the dynamic sparse convolution module, which contains the local features of the time series data, the time domain signal is converted to the frequency domain through fast Fourier transform, the energy of different frequency components is analyzed, and the energy of each frequency component is calculated to reflect the strength of the signal at different frequencies. The key frequency components are then screened based on the typical frequency range of predetermined inflammation-related signals, where the typical frequency range of predetermined inflammation-related signals is obtained based on industry experience or experimental data.

[0082] Based on the filtered frequency components, the energy of the key frequency components in each channel is summarized to reflect the importance of the channel. The sum of the energy of each channel is then mapped to a predetermined range, such as 0 to 1, to obtain the initial attention weight. The initial attention weight is applied to the self-attention mechanism to make the model pay more attention to the key frequency components.

[0083] Residual connection: Residual connections are introduced between model levels. Specifically, residual connections are introduced between the dynamic sparse convolution module and the self-attention module, and between the self-attention module and the fusion layer. Identity mapping is used to transfer low-level features to elevation, ensuring that gradients flow sequentially during reflection propagation. The residual connection introduced between the dynamic sparse convolution module and the self-attention module is part of the dynamic sparse convolution module; the residual connection introduced between the self-attention module and the fusion layer is part of the self-attention module.

[0084] Feature fusion module: collects low-level local features output by the dynamic sparse convolution module and high-level dependency features output by the self-attention module, and fuses the features through weighted summation. The weights are learnable parameters, which are initially set to equal and adjusted through training to highlight important hierarchical features. After weighted summation and fusion of features, batch normalization is used to stabilize the feature distribution and improve the robustness of the model. Based on this, the fused feature map is obtained.

[0085] Output module: Using a bottleneck structure, the feature map output by the feature fusion module is first reduced in dimensionality through global average pooling. The parameters are then compressed through two fully connected layers, with the first layer using 16 neurons and the second layer using 1 neuron. Dropout (with a probability of 0.3) is introduced in the first and second fully connected layers to prevent overfitting. The output of the second fully connected layer is mapped to the range [0, 1] using a sigmoid activation function to obtain an inflammation risk score.

[0086] The setting of the above-mentioned output module is only an exemplary structure and is not a limitation of the present invention. The output module can also adopt three fully connected layers, and Dropout can be introduced in each fully connected layer to prevent overfitting. The above-mentioned exemplary structure is only an exemplary structure given for the reason of lightweight model.

[0087] In this embodiment, dynamic sparse convolution, self-attention mechanism and feature fusion module are used to extract local changes and global features in time series data, improve the expressive ability of the model, and capture complex physiological signal patterns. The lightweight design of the model ensures the high efficiency of the model, is suitable for real-time processing needs, and provides timely inflammation risk scores for closed-loop regulation. In this embodiment, the initial attention weight is calculated based on the spectrum graph, so that the model can focus on the key frequency components of inflammation-related signals more quickly, improve the initial performance of the model, and by optimizing the initial process of the attention mechanism, the model can more effectively capture important physiological signal features and improve prediction accuracy.

[0088] S4. Setting multiple thresholds and corresponding multiple decision-making methods, comparing the inflammation risk score with the set thresholds, determining the decision-making method corresponding to the verification risk score, outputting corresponding frequency control parameters based on the decision-making method; and adjusting the current intensity according to the real-time neuron activation index;

[0089] See also Figure 3 As shown, in step S4, two thresholds are set, including a first threshold and a second threshold, wherein the first threshold is smaller than the second threshold, and three levels of decision-making methods are set, including basic maintenance, intermediate transition and emergency suppression;

[0090] When the inflammation risk score is less than a first threshold, a basic maintenance strategy is adopted to output a preset basic maintenance frequency;

[0091] When the inflammation risk score is greater than or equal to a first threshold and less than a second threshold, an intermediate transition strategy is adopted to linearly adjust the frequency according to the real-time change of the inflammation risk score, wherein the linear adjustment frequency is linearly adjusted based on the starting frequency and the linear growth term;

[0092] When the inflammation risk score is greater than or equal to the second threshold, the emergency suppression frequency is adopted and the preset emergency suppression frequency is called, wherein the emergency suppression frequency is the preset highest stimulation frequency.

[0093] An applicable implementation of this embodiment is as follows:

[0094] The first threshold is 0.3, the second threshold is 0.7, and the specific steps for generating stimulation parameters are as follows:

[0095] When the inflammatory risk score is less than 0.3, the basic maintenance frequency is used, where the basic maintenance frequency is 5 Hz;

[0096] When the inflammation risk score is greater than or equal to 0.3 and less than 0.7, the stimulation frequency is calculated using linear interpolation:

[0097] f stim=10+5×(IRS-0.3);

[0098] Where: f stim represents stimulation frequency; IRS represents inflammation risk score; 5×(IRS-0.3) represents linear growth coefficient;

[0099] When the inflammation risk is greater than or equal to 0.7, the emergency inhibition frequency is adopted, where the emergency inhibition frequency is the preset highest stimulation frequency, which is 20 Hz.

[0100] The steps for adjusting the current intensity are as follows:

[0101] A real-time current intensity is generated based on a linear mapping relationship between baseline current intensity and neuronal activation index; and an upper limit constraint of the current intensity is established based on the baseline current intensity; wherein the baseline current intensity is obtained based on clinical trials, and an expression for generating the real-time current intensity is as follows:

[0102] I=I0×(1+NAI);

[0103] Where: I represents the real-time stimulation current intensity; I0 represents the baseline current intensity; NAI represents the neuronal activation index.

[0104] As a further embodiment, in this embodiment, a stimulation frequency upper limit and a current upper limit are set, wherein the stimulation frequency upper limit is 20 Hz and the current upper limit is 5 times the baseline current intensity; and a parameter change rate limit is set, wherein Δf stim_max =2Hz / s,ΔI max =0.1mA / s; ΔI max Indicates the rate of change of current intensity; Δf stim_max Indicates the rate of change of frequency.

[0105] In this embodiment, the multi-threshold design improves the flexibility and safety of regulation, and can adapt to the individual differences and changes in inflammatory responses of different patients. The linear frequency adjustment method avoids sudden changes in stimulation parameters, reduces patient discomfort, and improves the comfort and compliance of treatment. Secondly, the emergency inhibition strategy can timely collect high-intensity stimulation when the verification risk rises sharply, effectively control the inflammatory response, and ensure patient safety.

[0106] S5. After applying the stimulation, monitoring the rate of change of the neuronal activation index and the high-frequency power index in the heart rate variability, and modifying the stimulation frequency based on the changes in the indexes within a predetermined time;

[0107] The step S5 comprises the following steps:

[0108] After the stimulation is applied, the NAI value (neuronal activation index) is continuously detected, and the rate of change of NAI is calculated every fixed time interval (1 minute):

[0109]

[0110] Where: ΔNAI represents the difference between the NAI value at the current time point and the NAI value of the previous monitoring round; Δt represents the time interval;

[0111] Collect ECG signals and perform preprocessing on them, including filtering and denoising. Then, use fast Fourier transform or wavelet transform to analyze the spectrum of the ECG signals, and then calculate the high-frequency power in heart rate variability (HRV). The specific steps are as follows:

[0112] An EGC sensor is used to collect ECG signals. High-frequency noise and low-frequency drift are removed through a bandpass filter, and further noise is removed using a median filter. Based on this, a preprocessed EGC signal is obtained. The RR interval sequence is extracted from the EGC signal, and the difference between adjacent RR intervals is calculated to obtain a difference sequence. The difference sequence is converted into a frequency domain signal, and the power spectral density is calculated. The high-frequency power is then calculated based on the power spectral density:

[0113]

[0114] Where: HF represents high frequency power; f high Indicates the upper limit frequency of the high frequency band; f low It represents the lower limit frequency of the high frequency band; PSD(f) represents the value of the power spectrum density at frequency f; df represents the frequency resolution.

[0115] Then evaluate the NAI change rate, specifically calculating the NAI decrease rate within 10 minutes after stimulation:

[0116]

[0117] Where: NAI 初始 Indicates the NAI value at the beginning of the stimulus; NAI 当前 Indicates the current NAI value;

[0118] HRV high power assessment, calculate the rate of increase of HRV high frequency power within 10 minutes after stimulation:

[0119]

[0120] Where: HF 初始 Indicates the HF power value at the beginning of stimulation; HF 当前 Indicates the current HF power value;

[0121] If the NAI decrease rate does not reach 10%, it means that the inhibitory effect is insufficient and the stimulation frequency needs to be increased:

[0122] f stim新 =f stim当前 +2Hz;

[0123] Where: f stim当前 Indicates the current stimulation frequency; f stim新 represents the corrected stimulation frequency;

[0124] If the HRV high-frequency power increases by 15%, it indicates that the parasympathetic nerves are fully activated and the current parameters should be maintained.

[0125] The corrected stimulation parameters are transmitted to the stimulation generator for the next round of stimulation.

[0126] The present invention adopts dual-frequency concha impedance analysis technology to obtain the neuron activation index, which is used as the basic indicator reflecting the inflammatory state. The neuron activation index time series data is then used as the input of the inflammatory risk prediction model. The inflammatory risk score is obtained based on the inflammatory risk prediction model to achieve a quantitative assessment of the inflammatory state. Combined with the comparison result based on the inflammatory risk score and the set threshold, the corresponding frequency control parameters are dynamically output through a multi-decision method, and the current intensity is adjusted in combination with the real-time neuron activation index, so that the stimulation parameters can be adjusted in real time with the changes in the inflammatory state. Finally, after the stimulation is applied, the stimulation frequency is corrected within a predetermined time by monitoring the change rate of the neuron activation index and the high-power index of heart rate variability, forming a closed loop of state perception, parameter adjustment, and effect feedback to parameter correction, ensuring that the stimulation parameters are always accurately matched with the inflammatory state, thereby achieving the goal of dynamically adjusting parameters according to the inflammatory state.

[0127] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. The taVNS closed-loop control method for highly inflammatory ARDS is characterized by: The following steps are involved: S1. Use dual-frequency concha impedance analysis to simultaneously measure low-frequency and high-frequency impedance parameters. Calculate the low-frequency impedance change rate and high-frequency phase difference based on the measured low-frequency and high-frequency impedance parameters. Then, use the low-frequency impedance change rate and high-frequency phase difference to derive the neuronal activation index. S2. sorting the continuously collected neuron activation index values ​​in time series to form neuron activation index time series data of a predetermined time length; S3. Inputting the generated neuronal activation index time series data into a lightweight inflammation risk prediction model, and processing the data based on the inflammation risk prediction model to obtain an inflammation risk score; S4. Setting multiple thresholds and corresponding multiple decision-making methods, comparing the inflammation risk score with the set thresholds, determining the decision-making method corresponding to the verification risk score, outputting corresponding frequency control parameters based on the decision-making method; and adjusting the current intensity according to the real-time neuron activation index; S5. After applying the stimulation, monitor the change rate index of the neuronal activation index and the high-frequency power index in the heart rate variability, and modify the stimulation frequency based on the changes in the indexes within a predetermined time.

2. The taVNS closed-loop control method for highly inflammatory ARDS according to claim 1, characterized in that: In step S1, a dual-frequency impedance sensor is attached to the concha area, and an AC bridge method is used to simultaneously measure the amplitude and phase of the impedance at 1 kHz and 10 kHz by applying an AC signal.

3. The taVNS closed-loop control method for highly inflammatory ARDS according to claim 1, characterized in that: After synchronously measuring the low-frequency and high-frequency impedance parameters in step S1, the following steps are included: Data preprocessing: Use filtering algorithm to remove noise from the collected low-frequency and high-frequency impedance parameters, and then normalize the denoised data; Calculation of low-frequency impedance change rate: Calculate the difference between the low-frequency impedance values ​​at the current moment and the previous moment, and then divide it by the time interval to obtain the impedance change rate; High-frequency phase difference calculation: Calculate the difference between the high-frequency phase value at the current moment and the previous moment to obtain the phase change; Calculation of neuronal activation index: The neuronal activation index was obtained by dividing the low-frequency impedance change by the high-frequency phase difference.

4. The taVNS closed-loop control method for highly inflammatory ARDS according to claim 1, characterized in that: The inflammation risk prediction model includes an input layer, a dynamic sparse convolution module, a self-attention mechanism module, a feature fusion module and an output module; Inputting the normalized neuron activation index time series data into the model through the input layer; The dynamic sparse convolution module calculates the local change rate of the data based on the input neuron activation index time series data, and dynamically adjusts the sparsity factor according to the local change rate, and controls the sparsity factor within a predetermined range. Then, a dynamic sparse convolution kernel is generated based on the adjusted sparsity factor. Then, a convolution operation is performed on the input neuron activation index time series data based on the generated dynamic sparse convolution kernel to obtain a feature map after sparse convolution. The sparse factor and dynamic sparse convolution kernel parameters are dynamically optimized based on back propagation; The self-attention mechanism divides the feature map after sparse convolution into multiple local blocks, calculates the attention weight in each local block, and uses multi-head attention. Each head focuses on a different feature dimension. Finally, the outputs of each head are spliced ​​and integrated through a linear layer to obtain a feature map processed by self-attention. The feature fusion module collects the feature maps after sparse convolution and self-attention processing, and fuses the features through weighted summation to obtain the fused feature map; The fused feature map is input into the output model, and the inflammation risk score is obtained based on the output module processing.

5. The taVNS closed-loop control method for highly inflammatory ARDS according to claim 4, characterized in that: There are residual connections between the dynamic sparse convolution module and the self-attention module, and between the self-attention module and the feature fusion module, where the residual connections use identity mapping to ensure that the gradient can be smoothly transmitted.

6. The taVNS closed-loop control method for highly inflammatory ARDS according to claim 4, characterized in that: The attention weight initialization method of the self-attention mechanism is as follows: Perform fast Fourier transform on the input neuron activation index time series data to obtain a spectrum graph; The energy of each frequency component is calculated based on the obtained spectrum, and the key frequency components are screened out according to the pre-set typical frequency range of inflammation-related signals; Based on the screened key frequency components, the energy sum of the key frequency components in each local block is calculated, and the calculated energy sum is normalized to obtain the initial attention weight.

7. The taVNS closed-loop control method for highly inflammatory ARDS according to claim 4, characterized in that: The data processing steps of the output layer include: The feature map output by the feature fusion module is reduced in dimensionality through global average pooling, and then the parameters are compressed through two fully connected layers. A Dropout layer is introduced in the fully connected layer to prevent overfitting. Finally, the output results of the fully connected layer are mapped through the Sigmoid activation function to obtain the inflammation risk score.

8. The taVNS closed-loop control method for highly inflammatory ARDS according to claim 1, characterized in that: In step S4, three thresholds are set, including a first threshold and a second threshold, wherein the first threshold is smaller than the second threshold, and three levels of decision-making methods are set, including basic maintenance, intermediate transition, and emergency suppression; When the inflammation risk score is less than a first threshold, a basic maintenance strategy is adopted to output a preset basic maintenance frequency; When the inflammation risk score is greater than or equal to a first threshold and less than a second threshold, an intermediate transition strategy is adopted to linearly adjust the frequency according to the real-time change of the inflammation risk score, wherein the linear adjustment frequency is linearly adjusted based on the starting frequency and the linear growth term; When the inflammation risk score is greater than or equal to the second threshold, the emergency suppression frequency is adopted and the preset emergency suppression frequency is called, wherein the emergency suppression frequency is the preset highest stimulation frequency.

9. The taVNS closed-loop control method for highly inflammatory ARDS according to claim 1, characterized in that: The steps for adjusting the current intensity are as follows: A real-time current intensity is generated based on a linear mapping relationship between baseline current intensity and neuron activation index; and an upper limit constraint of current intensity is constructed based on the baseline current intensity; wherein the baseline current intensity is obtained based on clinical trials.

Images