Dynamic parameter intelligent adjusting method and system for mouse pulse pump drug delivery

Abstract

The invention discloses a dynamic parameter intelligent adjustment method and system for mouse pulse pump administration, and the method comprises the steps: collecting a physiological feedback signal of an administration mouse in real time through a multi-channel biosensor, carrying out the denoising and feature extraction of the physiological feedback signal, and generating structured real-time physiological feature data; inputting the real-time physiological feature data into a pre-trained individualized drug response model, and outputting a predicted physiological response trajectory and corresponding confidence evaluation; dynamically determining a current optimal pulse frequency and administration dosage parameter combination through a multi-objective optimization algorithm based on a physiological response trajectory and the confidence evaluation; and according to the optimal pulse frequency and administration dosage parameter combination, a real-time control instruction is generated and issued to a pulse pump execution mechanism. By utilizing the embodiment of the invention, individualized and self-adaptive accurate adjustment of mouse pulse pump administration parameters can be realized, and the scientificity and controllability of the administration process and the reliability of experimental results are improved.

Description

Technical Field

[0001] This invention belongs to the field of bioengineering technology, and in particular to a method and system for intelligent adjustment of dynamic parameters for pulse pump drug delivery in mice. Background Technology

[0002] In biomedical and pharmacological research, pulsed pump drug delivery is a key technology for precisely controlling drug exposure dose and time. Currently, conventional pulsed pump control systems mostly employ preset fixed parameters or simple feedback adjustment modes, which are ill-suited to the dynamic changes in laboratory mice caused by individual differences and fluctuations in physiological state during drug administration. Existing technologies lack the ability to perform real-time fusion analysis and intelligent prediction of multi-channel physiological signals, resulting in lag in drug delivery parameter adjustments and preventing truly individualized and adaptive precision control. This limits the reliability and efficiency of complex pharmacodynamic and toxicological studies and fails to meet the research needs of cutting-edge dynamic drug delivery paradigms. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for intelligent adjustment of dynamic parameters for pulse pump drug delivery in mice, in order to overcome the shortcomings of the prior art, and to achieve individualized, adaptive and precise adjustment of pulse pump drug delivery parameters in mice, thereby improving the scientific nature, controllability and reliability of the drug delivery process and the experimental results.

[0004] One embodiment of this application provides a method for intelligent adjustment of dynamic parameters for pulse pump drug delivery in mice, the method comprising: Physiological feedback signals of drug-treated mice are collected in real time using multi-channel biosensors, and the physiological feedback signals are denoised and feature extracted to generate structured real-time physiological feature data. The real-time physiological feature data is input into a pre-trained individualized drug response model, which outputs the predicted physiological response trajectory and the corresponding confidence assessment. Based on the physiological response trajectory and the confidence assessment, the current optimal combination of pulse frequency and drug dosage parameters is dynamically determined through a multi-objective optimization algorithm. Based on the optimal combination of pulse frequency and dosage parameters, a real-time control command is generated and sent to the pulse pump actuator to achieve dynamic and intelligent adjustment of pulse pump drug delivery.

[0005] Another embodiment of this application provides a dynamic parameter intelligent adjustment system for pulse pump drug delivery in mice, the system comprising: The acquisition module is used to acquire physiological feedback signals of drug-treated mice in real time through multi-channel biosensors, and to denoise and extract features from the physiological feedback signals to generate structured real-time physiological feature data. The input module is used to input the real-time physiological feature data into the pre-trained individualized drug response model and output the predicted physiological response trajectory and the corresponding confidence assessment. The determination module is used to dynamically determine the current optimal combination of pulse frequency and drug dosage parameters based on the physiological response trajectory and the confidence assessment through a multi-objective optimization algorithm. The adjustment module is used to generate real-time control commands based on the optimal combination of pulse frequency and drug dosage parameters and send them to the pulse pump actuator to realize dynamic intelligent adjustment of pulse pump drug delivery.

[0006] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0007] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0008] Compared with existing technologies, the present invention provides a method for intelligent adjustment of dynamic parameters for pulse pump drug delivery in mice, which can realize individualized, adaptive and precise adjustment of pulse pump drug delivery parameters in mice, thereby improving the scientific nature, controllability and reliability of the drug delivery process and the experimental results. Attached Figure Description

[0009] Figure 1 A hardware structure block diagram of a computer terminal for a dynamic parameter intelligent adjustment method for pulse pump drug delivery in mice, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a method for intelligent adjustment of dynamic parameters for pulse pump drug delivery in mice, provided in an embodiment of the present invention; Figure 3 A flowchart illustrating another method for intelligent adjustment of dynamic parameters for pulse pump drug delivery in mice, provided in an embodiment of the present invention; Figure 4 A flowchart illustrating another method for intelligent adjustment of dynamic parameters for pulse pump drug delivery in mice, provided by an embodiment of the present invention; Figure 5 This is a schematic diagram of a dynamic parameter intelligent adjustment system for pulse pump drug delivery in mice, provided as an embodiment of the present invention. Detailed Implementation

[0010] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] The present invention first provides a method for intelligent adjustment of dynamic parameters for pulse pump drug delivery in mice. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0012] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a dynamic parameter intelligent adjustment method for pulse pump drug delivery in mice, provided as an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0013] See Figures 2-4 The present invention provides a method for intelligent adjustment of dynamic parameters for pulse pump drug delivery in mice, which may include the following steps: S201, real-time acquisition of physiological feedback signals of drug-treated mice through multi-channel biosensors, and denoising and feature extraction of the physiological feedback signals to generate structured real-time physiological feature data; Specifically, a multi-channel biosensor array can be used to simultaneously collect multimodal raw physiological signals from drug-treated mice, including at least electrocardiogram signals, electroencephalogram signals, blood oxygen saturation, and body temperature data; The core of this step is to use a multi-channel sensor array to simultaneously capture key physiological signals during drug administration in mice, providing complete and accurate raw data for subsequent processing. The specific implementation method is as follows: The multi-channel biosensor array adopts a miniaturized design to fit the size of mice (weighing 20-30g), closely adhering to the surface and key parts of the mouse. The array contains four types of dedicated sensors, corresponding to ECG, EEG, blood oxygen saturation and body temperature detection, respectively. All sensors work synchronously, with the acquisition frequency uniformly set to 250Hz to ensure that the multimodal signals are synchronized in time and avoid timing deviations.

[0014] The electrocardiogram (ECG) signals were acquired using miniature electrode sensors attached to both sides of the mouse's chest. The electrodes were made of medical-grade flexible material to avoid irritating the mouse's skin. The acquisition range was set to -5mV to +5mV, with a resolution of 1μV, which could accurately capture the weak signals of the mouse's ECG activity and eliminate interference from skin contact. In the example, the acquired raw ECG signals of the mouse showed typical P waves, QRS complexes, and T waves, with a waveform period of approximately 0.2 seconds (corresponding to a mouse heart rate of 300 beats / minute), which is consistent with the ECG characteristics of normal mice.

[0015] The electroencephalogram (EEG) signals were acquired using a miniature EEG sensor implanted on the surface of the mouse skull. The acquisition range was -100μV to +100μV, with a resolution of 0.1μV. The sensor focused on capturing the electrical activity signals of the mouse cerebral cortex. During the acquisition process, a shielding layer was designed to reduce external electromagnetic interference. In the example, the acquired raw EEG signals included alpha waves (8-13Hz) and beta waves (14-30Hz), with an alpha wave amplitude of approximately 20μV and a beta wave amplitude of approximately 10μV, reflecting the mouse's awake state before drug administration.

[0016] Blood oxygen saturation signals were acquired using a miniature photoelectric sensor clipped to the mouse's ear. The sensor employed dual-wavelength detection with red light (660nm) and infrared light (940nm), achieving a resolution of 1% and a range of 70%-100%. This allowed the sensor to reflect the proportion of oxyhemoglobin in the mouse's blood in real time. In the example, the raw blood oxygen saturation data of the mice before drug administration remained stable between 95% and 97%, with fluctuations ≤1%, which is within the normal physiological range.

[0017] Body temperature data was collected using a miniature temperature sensor implanted under the skin of the mouse abdomen. The measurement range was 36℃-39℃, with a resolution of 0.01℃. The sampling interval was consistent with other signals (4ms / time), which can accurately capture subtle changes in mouse body temperature. In the example, the original body temperature data of the mouse before drug administration was stable between 37.5℃ and 37.6℃, which is in line with the normal body temperature standard of mice.

[0018] All multimodal physiological raw signals collected by the sensors are transmitted in real time to the signal processing unit through a miniature data transmission module. The transmission process adopts a lossless transmission method to avoid signal distortion. At the same time, each signal is given a unique identifier (ECG, EEG, SpO2, T) and a timestamp (accuracy of 1ms) to ensure signal traceability. Finally, the signals are integrated to form a set of multimodal physiological raw signals, providing complete input for subsequent preprocessing.

[0019] The multimodal physiological raw signals were preprocessed, and wavelet transform algorithm was used to remove power frequency interference and electromyographic noise. Kalman filtering was used to eliminate motion artifacts and generate denoised physiological signals. The core of this step is to eliminate various interferences in the original signal, improve signal purity, and ensure the accuracy of subsequent feature extraction. The specific implementation method is as follows: The raw multimodal physiological signals contain three main types of interference: power frequency interference (50Hz, from external power equipment), electromyographic noise (10-300Hz, from mouse muscle contraction), and motion artifacts (random interference, from sensor contact offset caused by mouse activity). Preprocessing is performed in the order of "wavelet transform denoising → Kalman filtering artifact removal" to ensure that interference is completely removed without losing effective signals.

[0020] Wavelet transform algorithm is used to remove power frequency interference and electromyographic noise. The db4 wavelet is selected as the wavelet basis function, which has good time-domain and frequency-domain localization characteristics, suitable for denoising physiological signals. The decomposition level is set to 5 levels, where levels 1-2 mainly correspond to high-frequency electromyographic noise, levels 3-4 correspond to power frequency interference, and level 5 corresponds to the effective signal components. The specific processing is as follows: Each original signal is decomposed into 5 levels of wavelets to obtain the wavelet coefficients of each level. Soft thresholding is applied to the wavelet coefficients of levels 1-4. The threshold calculation formula is λ=σ×√(2lnN), where σ is the noise standard deviation (estimated through the high-frequency layer coefficients of the signal), and N is the number of signal sampling points. In the example, the number of ECG signal sampling points N=10000, σ=0.5μV, and λ≈0.023μV is calculated. Coefficients with absolute values ​​less than λ in levels 1-4 are set to 0, while coefficients greater than λ are retained. The signal is then reconstructed through inverse wavelet transform, thus removing power frequency interference and electromyographic noise.

[0021] Kalman filtering is used to eliminate motion artifacts, which are random interferences manifested as sudden shifts or fluctuations in the signal. A linear Kalman filtering algorithm is used to construct the signal state equation and observation equation to achieve artifact elimination. The state equation is X_k = A × X_{k-1} + W_k, where X_k is the signal state value at time k, A is the state transition matrix (set to 1 to ensure signal continuity), and W_k is the process noise (variance Q = 0.01). The observation equation is Z_k = H × X_k + V_k, where Z_k is the observation value at time k (the denoised signal), H is the observation matrix (set to 1), and V_k is the observation noise (variance R = 0.1).

[0022] During the filtering process, the signal state at time k is first predicted using the state equation, and then the predicted value is corrected using the observation equation to obtain the optimal estimate. In the example, the mouse's activity caused a sudden shift in the blood oxygen saturation signal (from 96% to 90%, lasting 100ms). After Kalman filtering, the shift was corrected, and the signal was restored to a stable range of 95%-96%, while preserving the normal subtle fluctuations of the blood oxygen signal without causing signal distortion.

[0023] After preprocessing, the validity of each signal is verified and the signal-to-noise ratio (SNR) is calculated (SNR≥30dB is considered valid). If the SNR does not meet the standard, wavelet transform and Kalman filtering are performed again until the requirements are met. Finally, four denoised physiological signals are generated with smooth waveforms and clear effective components, which can be directly used for subsequent feature extraction.

[0024] Feature extraction is performed on the denoised physiological signals, and the RR interval variability of the electrocardiogram signal, the power spectrum characteristics of the electroencephalogram signal, the trend slope of blood oxygen saturation and the rate of change of body temperature are calculated to generate a multidimensional physiological feature vector. The core of this step is to extract key features that reflect the physiological state of the mouse from the denoised signal, and to convert the continuous signal into quantifiable feature parameters. The specific implementation method is as follows: Feature extraction is performed separately for the four denoised physiological signals, with 1-2 core features extracted from each signal to ensure the comprehensiveness and relevance of the features. All features are calculated using quantifiable methods to ensure accurate and repeatable extraction results.

[0025] The focus of the ECG signal analysis is to extract the RR interval variability. The RR interval refers to the time interval between two adjacent R wave peaks in the ECG signal, reflecting the heart rate variability in mice and indirectly reflecting the state of the cardiovascular system. The specific calculation process is as follows: First, a peak detection algorithm (with a threshold set to 50% of the signal amplitude) is used to identify all R wave peaks in the ECG signal. The timestamp of each R wave is recorded, and the difference between two adjacent R wave timestamps is calculated to obtain the RR interval sequence. Then, the statistical characteristics of this sequence are calculated: standard deviation (SDNN) and mean (MRR). SDNN reflects the overall fluctuation of the RR interval, and MRR reflects the average heart rate. In the example, the RR interval sequence of the denoised ECG signal is 0.20s, 0.21s, 0.19s, and 0.20s, with SDNN=0.008s and MRR=0.20s calculated, which are used as the core features of the ECG signal.

[0026] The power spectrum features of the EEG signal were extracted. The EEG signal was converted from the time domain to the frequency domain using Fast Fourier Transform (FFT) to obtain the power spectral density (PSD). The power proportions of alpha waves (8-13Hz) and beta waves (14-30Hz) were analyzed, as these proportions reflect the alertness and neural activity state of the mouse brain. The specific calculation process was as follows: An FFT was performed on the denoised EEG signal, with 1024 sampling points and a frequency resolution of 0.244Hz. The sum of the power in the alpha and beta wave bands was calculated, and then divided by the sum of the power across the entire frequency band (0.5-300Hz) to obtain the alpha wave power proportion (Pα) and beta wave power proportion (Pβ). In the example, Pα = 35% and Pβ = 25%, indicating that the mouse was in a state of alertness and calmness.

[0027] The trend slope of the blood oxygen saturation signal is extracted to reflect the changing trend of blood oxygen saturation, indirectly reflecting the respiratory and circulatory system status of mice. The specific calculation process is as follows: Blood oxygen saturation data from 50 consecutive sampling points (corresponding to 200 ms) are selected, and a trend line is fitted using a linear regression algorithm. The slope of the trend line is the blood oxygen saturation trend slope k. A positive k indicates an increase in blood oxygen saturation, and a negative k indicates a decrease. The larger the absolute value of k, the more obvious the changing trend. In the example, the blood oxygen saturation data from 50 consecutive sampling points decreased from 96.2% to 95.8%, and the fitted trend slope k = -0.002% / ms reflects a slight decrease in blood oxygen saturation, consistent with normal physiological fluctuations.

[0028] The body temperature data was used to extract the rate of change of body temperature, reflecting the speed at which the mouse's body temperature changes, and indirectly reflecting the effect of drugs on mouse metabolism. The specific calculation process is as follows: the difference in body temperature between two adjacent sampling points is calculated and divided by the sampling interval (4 ms) to obtain the rate of change of body temperature, r, in °C / ms. A positive r indicates a rise in body temperature, and a negative r indicates a decrease. In the example, the body temperatures of two adjacent sampling points are 37.52 °C and 37.53 °C, with a sampling interval of 4 ms. The calculated r = 0.01 °C / 4 ms = 0.0025 °C / ms, reflecting a slight rise in body temperature.

[0029] After extraction, all feature parameters are integrated to generate a multidimensional physiological feature vector with 6 dimensions: ECG SDNN, ECG MRR, EEG Pα, EEG Pβ, blood oxygenation trend slope k, and body temperature change rate r. In the example, the multidimensional physiological feature vector at a certain moment is [0.008s, 0.20s, 35%, 25%, -0.002% / ms, 0.0025℃ / ms], which provides a basis for subsequent standardization processing.

[0030] The multidimensional physiological feature vectors are standardized and time-aligned to construct a structured data matrix containing timestamps and feature identifiers, thereby generating structured real-time physiological feature data.

[0031] The core of this step is to unify the feature scale and eliminate temporal bias, converting multidimensional feature vectors into structured data to facilitate subsequent input into the model for processing. The specific implementation method is as follows: Standardization is used to eliminate scale differences between different features (e.g., ECG SDNN is in seconds, blood oxygen trend slope is in % / ms) and avoid features with large scales dominating subsequent analysis. The Z-score standardization method is used, and the calculation formula is x_norm=(x-μ) / σ, where x is the original feature value, μ is the historical mean of the feature (obtained by statistical analysis of baseline data of mice before drug administration), and σ is the historical standard deviation of the feature. After standardization, all feature values ​​follow a normal distribution with a mean of 0 and a standard deviation of 1.

[0032] Temporal alignment is used to eliminate time discrepancies between different feature vectors. Since the time windows for feature extraction vary slightly (e.g., power spectrum feature extraction requires 500ms, body temperature change rate extraction requires 4ms), all feature vectors need to be aligned to the same time axis. Specifically, a fixed time interval (100ms) is selected as the feature sampling period, based on the timestamp. A standardized multidimensional physiological feature vector is extracted once per period. If a certain type of feature data is missing in a period, linear interpolation is used to supplement it, ensuring that each time point corresponds to a complete feature vector.

[0033] After time alignment is completed, a structured data matrix is ​​constructed. The rows of the matrix correspond to time points (sorted in ascending order by timestamp), and the columns correspond to feature identifiers (ECG SDNN, ECG MRR, EEG Pα, EEG Pβ, blood oxygenation trend slope k, body temperature change rate r). Each matrix element is the standardized value of the corresponding time point and the corresponding feature. At the same time, a timestamp is added to the first column of the matrix and a feature identifier is added to the first row to form a structured data matrix.

[0034] In the example, the structured data matrix contains 100 time points and 6 features. Each element is labeled with a timestamp and feature identifier to ensure that the data structure is clear and traceable. The final result is structured real-time physiological feature data. This data has a uniform format, consistent scale, and continuous temporal sequence, which can be directly input into the subsequent individualized drug response model to provide accurate physiological feature support for drug response prediction.

[0035] S202, The real-time physiological characteristic data is input into the pre-trained individualized drug response model, and the predicted physiological response trajectory and corresponding confidence assessment are output. Specifically, it may include: S2021, loading a pre-trained individualized drug response model, which is based on a long short-term memory network architecture and trained using historical drug administration data and physiological feedback signals; The core of this step is loading a pre-trained model adapted to a single mouse, providing reliable model support for subsequent physiological response prediction. The key lies in clarifying the model architecture design, training foundation, and loading / validation process. The specific implementation method is as follows: The personalized drug response model adopts a long short-term memory network architecture, which is an improved version of recurrent neural networks. Its core advantage is that it can capture the long-short-term dependencies of time-series data, adapt to the temporal continuity of physiological feature data, and effectively alleviate the gradient vanishing or gradient explosion problems of traditional recurrent neural networks. It is very suitable for handling the dynamic temporal correlation between physiological signals and drug response during mouse drug administration.

[0036] The model's architecture is designed to align with the characteristics of mouse physiological data. It consists of four parts: an input layer, hidden layers, an output layer, and an uncertainty estimation module. The input layer's dimensions are consistent with the structured real-time physiological feature data, being 6-dimensional (corresponding to ECG SDNN, ECG MRR, EEG Pα, EEG Pβ, blood oxygenation trend slope k, and body temperature change rate r). Batch normalization is used in the input layer to further unify feature scale and improve the model's generalization ability. The hidden layer uses two layers of Long Short-Term Memory (LSTM) network units, with 64 hidden units per layer. The ReLU activation function effectively accelerates model training convergence while suppressing gradient vanishing. A dropout layer with a dropout rate of 0.3 is added after each hidden layer to prevent overfitting during training and improve the model's adaptability to new data. The output layer uses a linear activation function with a 6-dimensional output, corresponding to the predicted values ​​of six physiological features at future time points, one-to-one with the physiological feature parameters in the input layer. The uncertainty estimation module is embedded after the output layer for subsequent confidence assessment calculations.

[0037] The model's pre-training process is based on the mouse's historical drug administration data and corresponding physiological feedback signals, ensuring the model's individualized adaptability and avoiding prediction bias caused by individual differences among mice. The training data contains two core components: one is historical drug administration parameter data, including pulse frequency (range 0.5-2Hz), dosage (range 0.1-0.5μL / pulse), pulse interval, and single dose volume, sorted by timestamp to form a time-series sequence of drug administration parameters; the other is physiological feedback signal data for the corresponding time period, synchronized with the time-series sequence of drug administration parameters, including raw signals of ECG, EEG, blood oxygen saturation, and body temperature, as well as extracted structured physiological feature data, with a data volume of no less than 5000 time-series samples, covering the physiological response states under different combinations of drug administration parameters.

[0038] The training process employs an adaptive moment estimation optimization algorithm. The initial learning rate is set to 0.001, which gradually decreases with each training iteration (decay coefficient 0.95, decreasing once every 100 iterations). The loss function is the mean squared error loss function, used to measure the deviation between the model's predicted values ​​and the actual physiological characteristic values. The number of training iterations is set to 500. Training is stopped when the loss function value on the validation set no longer decreases after 50 consecutive iterations, and the decrease is less than 1e-5. The model parameters at this point are then saved as the pre-trained model.

[0039] The model loading process is completed through a dedicated model loading interface. During loading, the pre-stored model parameter file is first read, including the weight matrices, bias vectors, dropout rates, activation function types, and other hyperparameters for the input, hidden, and output layers. These parameters are then assigned to the corresponding network layers, completing the reconstruction of the model structure and parameters. After loading, the model's validity is verified. Verification includes checking the completeness of model parameters (checking for missing weights and biases), the consistency of the model architecture (confirming that the number of network layers and hidden units matches those during pre-training), and the matching of input and output dimensions (verifying that the input layer dimension matches the structured real-time physiological feature data dimension, and that the output layer dimension is 6-dimensional).

[0040] In the example, after loading a personalized drug response model for a mouse, the validation results show that the model parameters are complete and without missing values; the architecture is: input layer (6-dimensional) → batch normalization layer → hidden layer 1 (64 long short-term memory units, dropout 0.3) → hidden layer 2 (64 long short-term memory units, dropout 0.3) → output layer (6-dimensional), consistent with the pre-trained architecture; the input and output dimensions match the real-time physiological feature data, and the validation passes. If the validation fails, the backup pre-trained model for the mouse (the model with the second-best training iterations) is automatically loaded. If the backup model also fails to load, a model anomaly warning is triggered, the prediction process is paused, and the loaded model is ensured to accurately complete the subsequent physiological response prediction task, thus completing the loading and validation of the pre-trained personalized drug response model.

[0041] S2022, structured real-time physiological characteristic data are input into the loaded individualized drug response model, and the physiological state prediction values ​​at multiple future time points are obtained through forward propagation calculation, generating a preliminary physiological response prediction sequence; The core of this step is to use the loaded model to perform time-series prediction on real-time physiological feature data, obtain predicted values ​​of physiological state at multiple future time points, and form a preliminary prediction sequence. The key is to ensure that the input data is adapted and the forward propagation calculation is accurate. The specific implementation method is as follows: First, the structured real-time physiological feature data undergoes input adaptation processing. The structured data is a structured data matrix containing timestamps and feature identifiers. Rows correspond to time points (sorted in ascending order of timestamps), and columns correspond to 6 physiological features. Before inputting it into the model, it needs to be converted into a time-series input format that the model can recognize. That is, feature data is extracted in chronological order to construct a time-series input matrix. The dimension of the input matrix is ​​(T×C), where T is the length of the input time series (set to 50 time points, corresponding to 5000ms, to ensure that sufficient physiological time-series information can be captured), and C is the feature dimension (6 dimensions). At the same time, the timestamp information of the input data is retained for the time-series alignment of subsequent prediction sequences.

[0042] In the example, the timestamp interval of the structured real-time physiological feature data is 100ms. The standardized feature data of the most recent 50 time points are extracted to construct an input matrix (50×6). Each row of the matrix corresponds to 6 standardized physiological feature values ​​at a time point. The timestamps range from 20240701100000000ms to 20240701100004900ms to ensure the continuity and integrity of the input time sequence. After the input is adapted, the input matrix is ​​input into the loaded individualized drug response model.

[0043] Forward propagation computation is the core process of model prediction. Input data is progressively passed from the model's input layer to the output layer, and prediction results are obtained through calculations at each layer. The specific process is as follows: The input matrix first enters the input layer, undergoes batch normalization to further eliminate feature scale differences, and ensures stable data distribution. It is then passed to the Long Short-Term Memory (LSTM) network units in the first hidden layer. These LSM units, through the synergistic action of the forget gate, input gate, and output gate, filter and retain key information (such as trends in physiological characteristics) from the input time-series data. The forget gate discards irrelevant historical physiological data, and the input gate updates the current physiological data. The system extracts effective information about the physiological features. The output gate is used to output the current hidden state. After processing through the first hidden layer, a 64-dimensional hidden state vector is obtained. Then, a dropout layer randomly discards 30% of the hidden units to avoid overfitting and affecting prediction accuracy. The processed hidden state vector is fed into the second hidden layer, and the long short-term memory network operation and dropout processing of the first layer are repeated to obtain the final 64-dimensional hidden state vector. Finally, the hidden state vector is fed into the output layer, and through the linear activation function operation, the 64-dimensional hidden state is transformed into a 6-dimensional output vector, which is the predicted value of 6 physiological features at a single future time point.

[0044] To achieve multi-timepoint prediction, a rolling prediction method is adopted. Each time a physiological characteristic prediction value for a future time point is obtained, this prediction value is used as a new input feature. This value is then combined with the feature data from the previous 49 input time points to form a new input matrix, which is then input into the model again to predict the next time point. This process is repeated until a preset number of future time points have been predicted. The preset number of future prediction time points is 10, with each time point spaced 100ms apart, corresponding to the predicted physiological response state of mice within the next 1000ms. This ensures that the prediction results cover the critical time period for subsequent drug efficacy, providing sufficient predictive data for subsequent parameter optimization.

[0045] During the forward propagation calculation, the calculation precision is set to 1e-6 to ensure the accuracy of the predicted values. At the same time, the timestamp of each prediction is recorded. The prediction timestamp is based on the timestamp of the last input time point and is incremented sequentially (e.g., if the last input timestamp is 20240701100004900ms, the first prediction timestamp is 20240701100005000ms, the second is 20240701100005100ms, and so on). In the example, after 10 rolling predictions, 10 predicted values ​​of physiological characteristics at future time points are obtained, forming a preliminary physiological response prediction sequence. Each prediction item in the sequence includes a timestamp and predicted values ​​of 6 physiological characteristics (unstandardized; the standardized values ​​output by the model are converted into true physiological parameter values ​​through inverse standardization). For example, the predicted values ​​at a certain prediction time point are: timestamp 20240701100005000ms, ECG SDNN=0.0082s, ECG MRR=0.202s, EEG Pα=34.8%, EEG Pβ=25.3%, blood oxygenation trend slope k=-0.0018% / ms, and body temperature change rate r=0.0026℃ / ms.

[0046] After the preliminary physiological response prediction sequence is generated, its rationality is verified. The verification standard is whether the predicted values ​​of each physiological characteristic are within the normal physiological range of mice (e.g., the normal range of ECG MRR is 0.18-0.22s, and the normal range of body temperature change rate is -0.005-0.005℃ / ms). If a predicted value exceeds the normal range, it is marked as a suspected abnormal predicted value. The value is retained but marked with an abnormality label. If more than 3 predicted values ​​exceed the normal range, forward propagation calculation is performed again to ensure the rationality of the preliminary prediction sequence. Finally, a preliminary physiological response prediction sequence with continuous time sequence and reasonable data is output.

[0047] S2023, the uncertainty estimation module based on the individualized drug response model, uses the Monte Carlo dropout method to calculate the confidence interval of the predicted values ​​at each time point and generate the prediction confidence assessment matrix; The core of this step is to estimate the uncertainty of the prediction results using the Monte Carlo dropout method, calculate the confidence interval of the predicted values ​​at each time point, and generate a confidence assessment matrix. This provides a reliable basis for subsequent result integration and parameter optimization. The specific implementation method is as follows: First, we must clarify the core significance of uncertainty estimation. Due to individual differences in the physiological state of mice and the randomness of drug response, the model predictions inevitably contain uncertainty. The greater the uncertainty, the lower the reliability of the prediction results. Therefore, it is necessary to quantify the confidence level of the prediction results through an uncertainty estimation module. The Monte Carlo dropout method is an uncertainty estimation method adapted to long short-term memory networks. Its core principle is to repeatedly activate the dropout layer during the model prediction process, and obtain multiple prediction results by randomly discarding hidden units multiple times during forward propagation. The confidence interval and confidence level are calculated using the statistical characteristics of these prediction results. This eliminates the need to train a complex uncertainty estimation model, thus balancing computational efficiency and estimation accuracy.

[0048] The Monte Carlo dropout method is implemented based on a pre-loaded individualized drug response model. The model parameters remain unchanged, and only the dropout layer of the hidden layer is enabled (enabled during the training phase and disabled during the regular prediction phase). The number of forward propagation iterations is set to 50. This number balances the accuracy of uncertainty estimation and computational efficiency. Too few iterations will lead to unreliable statistical results, while too many iterations will increase computation time. The 50 iterations can be completed within 100ms, which meets the time requirements for real-time adjustment.

[0049] The specific calculation process is as follows: Based on the input matrix constructed in step two, each time the forward propagation is repeated, the dropout layer randomly discards 30% of the hidden units (the discard ratio is consistent with that during model training). Through forward propagation, a set of physiological feature prediction values ​​for the next 10 time points is obtained. This process is repeated 50 times to obtain 50 independent prediction sequences. Each prediction sequence contains prediction values ​​for 10 time points and 6 physiological features. Subsequently, for each physiological characteristic at each future time point, 50 sets of predicted values ​​were statistically analyzed. The mean (as the final predicted value of that characteristic at that time point, consistent with the predicted value in step two) and standard deviation (reflecting the dispersion of the predicted values; the larger the standard deviation, the higher the uncertainty) of the 50 sets of predicted values ​​were calculated. Based on the mean and standard deviation, a 95% confidence interval was calculated. The formula for calculating the confidence interval is [μ-1.96×σ,μ+1.96×σ], where μ is the mean of the 50 sets of predicted values, σ is the standard deviation, and 1.96 is the normal distribution quantile corresponding to the 95% confidence level. This confidence interval indicates that there is a 95% probability that the mouse's true physiological characteristic value will fall within this interval.

[0050] Simultaneously, the confidence level of each predicted value is calculated. The confidence level is calculated based on the standard deviation and the normal fluctuation range of mouse physiological characteristics. The calculation formula is conf=1-(σ / σ_max), where σ is the standard deviation of the predicted value of the feature, and σ_max is the maximum allowable standard deviation of the physiological feature (obtained based on historical physiological data statistics, such as σ_max=0.002s for ECG SDNN and σ_max=5% for EEG Pα). The confidence level ranges from 0 to 1. The closer the confidence level is to 1, the higher the reliability of the prediction result; the closer it is to 0, the lower the reliability.

[0051] In the example, the ECG SDNN prediction value at a future time point (20240701100005000ms), after 50 repeated calculations, has a mean μ = 0.0082s, a standard deviation σ = 0.0003s, and a 95% confidence interval of [0.0082 - 1.96 × 0.0003, 0.0082 + 1.96 × 0.0003] = [0.0076s, 0.0088s]. The ECG SDNN's σ_max = 0.002s. The confidence level conf = 1 - (0.0003 / 0.002) = 0.85, indicating that the reliability of the predicted value is relatively high; the blood oxygen trend slope k at the same time point, the mean μ of the 50 predicted values ​​μ = -0.0018% / ms, the standard deviation σ = 0.0005% / ms, σ_max = 0.001% / ms, and the confidence level conf = 1 - (0.0005 / 0.001) = 0.5, indicate that the uncertainty of the predicted value is relatively high and the reliability is moderate.

[0052] After calculating the confidence intervals and confidence scores for all time points and all physiological characteristics, a prediction confidence assessment matrix is ​​generated. This matrix is ​​a two-dimensional matrix, with rows corresponding to 10 future time points (sorted in ascending order of timestamps) and columns corresponding to 6 physiological characteristics. Each matrix element contains the 95% confidence interval (upper and lower limits) and confidence score for that characteristic at that time point. The first row of the matrix is ​​labeled with the feature identifier, and the first column is labeled with the timestamp, ensuring that the matrix structure is clear and traceable. For example, in the confidence assessment matrix, the elements of a certain row (timestamp 20240701100005000ms) are: ECG SDNN [0.0076s, 0.0088s] (confidence score 0.85), ECG MRR [0.198s, 0.206s] (confidence score 0.88), EEG Pα [33.9%, 35.7%] (confidence score 0.90), etc. Finally, a complete prediction confidence assessment matrix is ​​generated to quantify the reliability of each prediction result.

[0053] S2024 integrates the preliminary physiological response prediction sequence and the prediction confidence assessment matrix to construct a continuous physiological response trajectory with confidence intervals, generating a complete physiological response trajectory prediction result.

[0054] The core of this step is to organically integrate the preliminary prediction sequence with the confidence assessment matrix to construct a continuous physiological response trajectory with reliability indicators, providing a comprehensive and accurate predictive basis for subsequent multi-objective optimization. The specific implementation method is as follows: The integration process follows the principles of "time alignment, feature matching, and trajectory smoothing." First, time alignment is performed to ensure that the timestamps of the preliminary physiological response prediction sequence and the prediction confidence assessment matrix are completely consistent. The preliminary prediction sequence contains prediction values ​​for 10 future time points, and the rows of the confidence assessment matrix also correspond to these 10 time points. The timestamps of each time point are one-to-one. If there is a slight timestamp deviation (≤1ms), the timestamps of the preliminary prediction sequence are used as the standard, and the timestamps of the confidence assessment matrix are adjusted to ensure complete time synchronization and avoid integration errors caused by time misalignment.

[0055] Subsequently, feature matching is performed. The predicted values ​​of the six physiological features at each time point in the preliminary physiological response prediction sequence are matched one-to-one with the corresponding time point, confidence interval, and confidence level of the corresponding feature in the prediction confidence assessment matrix. That is, the predicted value of each physiological feature corresponds to a unique 95% confidence interval and confidence level. After the matching is completed, complete prediction information for each time point is formed, including the timestamp, the predicted values ​​of the six physiological features, the confidence interval and confidence level of each feature, ensuring that each predicted value has a corresponding reliability identifier, which facilitates the subsequent analysis of the credibility of the prediction results.

[0056] Since the initial physiological response prediction sequence consists of 10 discrete time points, it cannot intuitively reflect the continuous trend of the mouse's physiological response. Therefore, it is necessary to smooth the discrete prediction values ​​to construct a continuous physiological response trajectory. The smoothing process uses a linear interpolation algorithm. For each physiological feature, with the timestamp as the horizontal axis and the predicted value as the vertical axis, the 10 discrete prediction values ​​are used as interpolation nodes. Linear interpolation is performed between two adjacent prediction time points, with the interpolation interval set to 10ms. That is, each 10ms interpolation yields a new prediction value and a corresponding confidence interval (the confidence interval is obtained by linear interpolation of the confidence intervals of two adjacent nodes). Finally, the 10 discrete time point prediction values ​​are smoothed into 100 consecutive time point prediction values ​​(covering the next 1000ms), thus constructing a continuous physiological response trajectory.

[0057] The specific calculation process of linear interpolation is illustrated below: Taking ECG SDNN as an example, the predicted values ​​for two adjacent discrete time points t1 (20240701100005000ms) and t2 (20240701100005100ms) are 0.0082s and 0.0083s, respectively, with confidence intervals of [0.0076s, 0.0088s] and [0.0077s, 0.0089s]. At t3 (20240701100005010ms) between t1 and t2, the interpolated ECG SDNN predicted value is 0.00821s, with a lower limit of 0.00761s and an upper limit of 0.00881s, and a confidence level of 0.851. This method achieves continuous smoothness between the predicted value and the confidence interval, ensuring that the physiological response trajectory can accurately reflect the dynamic changes in the physiological state of the mouse.

[0058] After the continuous physiological response trajectory is constructed, the rationality of the integrated results is verified. The verification includes three aspects: first, temporal continuity, ensuring that the interpolated continuous trajectory has no time discontinuities and that the timestamps are increasing in an orderly manner; second, feature consistency, ensuring that the trend of the continuous trajectory of each physiological feature is consistent with the trend of the discrete predicted value, without abrupt changes (e.g., if adjacent discrete predicted values ​​increase, the interpolated trajectory should also maintain an increasing trend); and third, confidence level rationality, ensuring that the confidence level changes smoothly after interpolation without large fluctuations. If the confidence level of a certain trajectory suddenly drops below 0.3, it is marked as a high uncertainty region, indicating that it needs to be given special attention during subsequent parameter optimization.

[0059] After successful verification, a complete physiological response trajectory prediction result is generated, which includes three core components: First, continuous physiological response trajectory data, covering 100 consecutive time points within the next 1000ms and the predicted values ​​of 6 physiological characteristics, sorted chronologically to form a continuous trajectory curve; second, confidence level labeling data, with each feature at each time point corresponding to a 95% confidence interval and confidence level, and high uncertainty regions separately marked; and third, trajectory analysis summary, briefly describing the changing trends of each physiological characteristic (e.g., ECG SDNN shows a slight upward trend, and blood oxygen saturation shows a stable trend) and the overall prediction reliability (e.g., confidence level ≥ 0.7 for more than 80% of time points, indicating overall reliable prediction results).

[0060] In the example, the complete summary of the physiological response trajectory prediction results is as follows: Within the next 1000ms, the mouse ECG SDNN steadily increased from 0.0082s to 0.0085s, with a confidence level ≥0.8, indicating reliable prediction; the EEG Pα remained between 34% and 36%, with a confidence level ≥0.85, indicating reliable prediction; the blood oxygen trend slope remained between -0.002% and 0.001% / ms, with confidence levels of 0.5-0.7 at some time points, indicating some uncertainty; the body temperature change rate showed a slight upward trend, with a confidence level ≥0.8, indicating reliable prediction. This prediction result is complete and continuous, encompassing both the dynamic trajectory of the physiological response and quantifying the reliability of the prediction results. It can be directly used as the core input for subsequent multi-objective optimization algorithms to determine the optimal drug administration parameters, ensuring the accuracy and rationality of parameter optimization.

[0061] S203, based on the physiological response trajectory and the confidence assessment, the current optimal combination of pulse frequency and drug dosage parameters is dynamically determined through a multi-objective optimization algorithm; Specifically, it may include: S2031, constructing a multi-objective optimization function, including maximizing efficacy indicators, minimizing side effect risk and minimizing drug accumulation, and generating the corresponding multi-objective optimization problem; The core of this step is to quantify the optimization objective of mouse pulse pump drug delivery into a mathematical function, clarifying the core direction of multi-objective optimization and providing a foundation for subsequent algorithm solutions. The key is to ensure that each objective function fits the physiological response characteristics and that the parameters are quantifiable. The specific implementation method is as follows: The core of the multi-objective optimization function is to simultaneously consider three mutually constraining objectives: maximizing efficacy indicators, minimizing side effect risks, and minimizing drug accumulation. These three objectives are interdependent and inseparable, and need to be integrated into a unified optimization model through a quantification function. Each objective function is constructed based on the physiological response trajectory prediction results and confidence assessment matrix generated in step three to ensure that it is linked to the actual physiological response of mice.

[0062] First, we define a function to maximize the drug efficacy index, f1(x), where x is the vector of parameters to be optimized (x=[f,d], f is the pulse frequency in Hz; d is the dose in μL / pulse). The quantification of the drug efficacy index is based on characteristic parameters related to drug action in the physiological response trajectory. Combining historical data from mouse drug administration experiments, we select two core indicators: the proportion of EEG Pα power and the trend slope k of blood oxygen saturation. A weighted summation method is used to construct the drug efficacy index function. The weights are set according to the drug action mechanism: the weight of EEG Pα is w1=0.6, and the weight of blood oxygen trend slope k is w2=0.4. The sum of the weights is 1 to ensure the rationality of the quantification results.

[0063] The specific calculation formula for the drug efficacy index function is f1(x) = w1 × (Pα(x) / Pα_max) + w2 × (1 - |k(x) / k_max|), where Pα(x) is the average proportion of EEG Pα power in the physiological response trajectory within the next 1000ms under the current parameter combination x, Pα_max is the maximum effective value of EEG Pα when the drug takes effect in the mouse (set to 40% based on historical data statistics); k(x) is the average slope of the blood oxygen saturation trend under the current parameter combination x, and k_max is the maximum allowable absolute value of the blood oxygen trend slope when the drug takes effect (set to 0.005% / ms). The value of f1(x) ranges from 0 to 1. The closer the value is to 1, the better the drug efficacy. If f1(x) > 1, it is calculated as 1; if f1(x) < 0, it is calculated as 0.

[0064] In the example, with a parameter combination x=[1.2Hz, 0.3μL / pulse], the physiological response trajectory prediction yields Pα(x)=36% and k(x)=-0.002% / ms. Substituting these values ​​into the formula, we calculate: f1(x)=0.6×(36% / 40%)+0.4×(1-|-0.002 / 0.005|)=0.6×0.9+0.4×(1-0.4)=0.54+0.24=0.78, indicating that the drug efficacy is good under this parameter combination. Simultaneously, considering the confidence assessment matrix, if the average confidence of EEG Pα is <0.7, f1(x) is corrected using the mean confidence coefficient. In this example, the average confidence is 0.85, so no correction is needed. If the average confidence is 0.6, the corrected f1(x)=0.78×0.6=0.47, reducing the drug efficacy score and reflecting the impact of prediction uncertainty.

[0065] Next, we define a function to minimize the side effect risk, f2(x). The side effect risk is mainly reflected in the effect of the drug on the cardiovascular system and body temperature of mice. The quantification is based on two core indicators: electrocardiogram SDNN and body temperature change rate r, which are constructed using a weighted summation method. The weights are set as follows: electrocardiogram SDNN weight w3=0.7, body temperature change rate r weight w4=0.3, and the sum of the weights is 1. The higher the side effect risk, the larger the value of f2(x). The optimization objective is to minimize the value of this function.

[0066] The specific formula for calculating the side effect risk function is f2(x) = w3 × (|SDNN(x) - SDNN_base| / SDNN_base) + w4 × (|r(x)| / r_max), where SDNN(x) is the mean value of the ECG SDNN under the current parameter combination x, SDNN_base is the baseline value of the ECG SDNN under normal physiological conditions of the mouse (based on the baseline data before drug administration, set to 0.007s); r(x) is the mean value of the rate of change of body temperature under the current parameter combination x, and r_max is the absolute value of the maximum rate of change of body temperature that the mouse can tolerate (set to 0.005℃ / ms). The value of f2(x) ranges from 0 to ∞. The closer the value is to 0, the lower the risk of side effects. If f2(x) > 1, it is calculated as 1 (considered a serious side effect).

[0067] In the example, the above parameter combination x=[1.2Hz, 0.3μL / pulse] predicts SDNN(x)=0.0072s, r(x)=0.0022℃ / ms. Substituting these values ​​into the formula, we get: f2(x)=0.7×(|0.0072-0.007| / 0.007)+0.3×(|0.0022| / 0.005)=0.7×(0.0002 / 0.007)+0.3×0.44≈0.02+0.132=0.152, indicating a low risk of side effects. Similarly, considering the confidence level assessment, if the average confidence level of the ECG SDNN is <0.7, the correction coefficient is 1.2, and the corrected f2(x)=0.152×1.2≈0.182, thus appropriately increasing the side effect risk score.

[0068] Finally, a function f3(x) is defined to minimize the drug accumulation. Excessive drug accumulation leads to the risk of toxicity. Quantification is based on dosing parameters and dosing duration, combined with the mouse's metabolic rate. The calculation formula is f3(x) = (f × d × T) / M, where f is the pulse frequency (Hz), d is the dosage (μL / pulse), T is the prediction duration (1000ms = 1s), and M is the mouse weight (set to 25g to suit the typical size of laboratory mice). The value of f3(x) ranges from 0 to ∞. The closer the value is to 0, the lower the drug accumulation. The optimization objective is to minimize this function value, and the drug accumulation threshold is set at 0.02 μL / (g). If f3(x) > 0.02, it is considered to be an excess and the parameter combination is directly eliminated.

[0069] In the example, the parameter combination x = [1.2Hz, 0.3μL / pulse] is used to calculate: f3(x) = (1.2 × 0.3 × 1) / 25 = 0.36 / 25 = 0.0144μL / (g) The value of f3(x) is less than the threshold of 0.02, which meets the requirements. If the parameter combination is x = [2.0 Hz, 0.5 μL / pulse], then f3(x) = (2.0 × 0.5 × 1) / 25 = 1.0 / 25 = 0.04 μL / (g) If the threshold is exceeded (s), the device is directly eliminated.

[0070] By integrating three objective functions, a multi-objective optimization problem is generated. The optimization objectives are: minimize[-f1(x),f2(x),f3(x)], where minimizing -f1(x) is equivalent to maximizing f1(x). The three objectives are optimized simultaneously and are mutually restrictive. It is necessary to maximize the efficacy of the drug while minimizing the risk of side effects and the amount of drug accumulation, thus forming a complete multi-objective optimization model, which provides a clear direction for subsequent algorithm solutions.

[0071] S2032, based on the physiological response trajectory prediction results and confidence assessment matrix, construct a set of constraints, including the safe range of physiological parameters, maximum dosing rate and minimum dosing interval, and generate optimized constraints; The core of this step is to set constraints on the boundaries of parameter optimization to ensure that the optimized parameter combinations are safe, feasible, and conform to the physiological tolerance range of mice and the working characteristics of the pulse pump. The constraints need to be closely integrated with the physiological response trajectory and confidence assessment. The specific implementation method is as follows: The set of constraints is divided into three categories: physiological parameter safety constraints, drug administration parameter boundary constraints, and confidence correlation constraints. All constraints are inequality constraints to ensure that the parameter combination is within a safe and feasible range, avoiding problems such as drug overdose and abnormal physiological parameters. Each constraint has a clear parameter range and basis, which is combined with mouse experimental data and pulse pump operating parameters.

[0072] The first type of constraint is the physiological parameter safety range constraint. Based on the normal range of 6 physiological characteristics in the physiological response trajectory prediction results, it ensures that the optimized parameter combination will not cause the mouse's physiological parameters to exceed the safety range, avoiding poisoning, physiological disorders and other situations. At the same time, combined with the confidence assessment matrix, the safety range is appropriately narrowed for predicted values ​​with low confidence, thereby increasing the strictness of the constraint.

[0073] The specific physiological parameter safety constraints are as follows: The safe range for ECG SDNN is [SDNN_base×0.8, SDNN_base×1.2], i.e., [0.0056s, 0.0084s], with the constraint 0.0056≤SDNN(x)≤0.0084; the safe range for ECG MRR is [0.18s, 0.22s], with the constraint 0.18≤MRR(x)≤0.22; the safe range for EEG Pα is [25%, 45%], with the constraint... 25%≤Pα(x)≤45%; the safe range of EEG Pβ is [20%, 35%], with the constraint that 20%≤Pβ(x)≤35%; the safe range of blood oxygen saturation trend slope k is [-0.004% / ms, 0.004% / ms], with the constraint that -0.004≤k(x)≤0.004; the safe range of body temperature change rate r is [-0.004℃ / ms, 0.004℃ / ms], with the constraint that -0.004≤r(x)≤0.004.

[0074] If the mean confidence level of a certain physiological parameter in the physiological response trajectory is <0.7, it indicates that the prediction value has high uncertainty. The safety range is reduced by 20%. In the example, the mean confidence level of the blood oxygen trend slope k is 0.65 <0.7, and the safety range is reduced to [-0.0032% / ms, 0.0032% / ms]. The constraints are more stringent to ensure the physiological safety of the mice.

[0075] The second type of constraint is the boundary constraint of the administration parameter, which is set based on the mouse body size, metabolic rate and pulse pump working characteristics. It is divided into maximum administration rate constraint, minimum administration interval constraint and administration parameter range constraint to ensure that the administration parameter meets the pulse pump execution capability and mouse tolerance.

[0076] Maximum dosing rate constraint: Dosing rate v = f × d, unit μL / s, the maximum tolerated dosing rate in mice is 0.8 μL / s, constraint condition is f × d ≤ 0.8; Minimum dosing interval constraint: Dosing interval t = 1 / f, unit ms, the minimum dosing interval of the pulse pump is 500ms (i.e., maximum frequency 2Hz), constraint condition is 1 / f ≥ 500ms, i.e., f ≤ 2Hz; Dosing parameter range constraint: Pulse frequency f range is [0.5Hz, 2Hz] (0.5Hz is the minimum frequency of the pulse pump, 2Hz is the maximum frequency), constraint condition is 0.5 ≤ f ≤ 2; Dosage range d range is [0.1μL / pulse, 0.5μL / pulse] (0.1μL is the minimum single dose of the pulse pump, 0.5μL is the maximum tolerated single dose in mice), constraint condition is 0.1 ≤ d ≤ 0.5.

[0077] In the example, the parameter combination x=[2.1Hz, 0.3μL / pulse] has a frequency f=2.1Hz>2Hz, which violates the minimum dosing interval constraint and is directly eliminated; the parameter combination x=[1.5Hz, 0.6μL / pulse] has a dose d=0.6μL / pulse>0.5μL / pulse, which violates the dosing dose range constraint and is directly eliminated; the parameter combination x=[1.2Hz, 0.7μL / pulse] has a dosing rate v=1.2×0.7=0.84μL / s>0.8μL / s, which violates the maximum dosing rate constraint and is directly eliminated.

[0078] The third type of constraint is the confidence correlation constraint, which is set based on the prediction confidence evaluation matrix to ensure that parameter combinations with high prediction reliability are selected first during the optimization process. The constraint condition is: the mean prediction confidence of the 6 physiological features in the physiological response trajectory is ≥0.6. If the mean is <0.6, it indicates that the overall prediction result is unreliable, and the corresponding parameter combination is directly eliminated to avoid unreasonable optimization results due to prediction errors.

[0079] In the example, the mean confidence level of the physiological feature prediction corresponding to a certain parameter combination is 0.58 < 0.6, so it is directly eliminated even if other constraints are met; if the mean is 0.62 ≥ 0.6, the parameter combination is retained and enters the subsequent optimization and screening.

[0080] By integrating the three types of constraints, a complete set of optimization constraints is generated. All constraints are effective simultaneously. In the subsequent algorithm solution process, this set of constraints must be strictly followed. Only parameter combinations that satisfy all constraints are searched to ensure that the optimized parameter combinations are safe, feasible, and reliable, providing clear constraint boundaries for the subsequent non-dominated sorting genetic algorithm solution.

[0081] S2033 uses a non-dominated sorting genetic algorithm to solve a multi-objective optimization problem. Under optimization constraints, it searches for Pareto optimal solutions and generates a set of candidate parameter combinations. The core of this step is to use a non-dominated sorting genetic algorithm to search for the Pareto optimal solution set of a multi-objective optimization problem within the constraints, and to select candidate parameter combinations that take into account the three optimization objectives. The key is to set the algorithm parameters to fit the scenario, ensuring both solution efficiency and solution set quality. The specific implementation is as follows: Non-dominated sorting genetic algorithm is a high-efficiency multi-objective optimization algorithm. Its core advantage is that it can search for multiple Pareto optimal solutions in a single iteration without converting multiple objectives into a single objective. It is suitable for the three mutually constrained optimization objectives in this case. Its core principle is to classify individuals (parameter combinations) in the population through non-dominated sorting, and maintain the diversity of the solution set by combining crowding calculation. Through selection, crossover, and mutation operations, it gradually converges to the Pareto optimal solution set, which meets the needs of mouse drug administration parameter optimization.

[0082] The specific parameter settings of the algorithm are combined with the range of drug parameters and optimization requirements to ensure solution efficiency and rationality of the solution set: the population size is set to 50 (the number of individuals, i.e., the number of parameter combinations being optimized simultaneously), balancing solution efficiency and solution set diversity. Too many individuals will increase computation time, while too few individuals will lead to an incomplete solution set; the number of iterations is set to 30, and the iteration is terminated early when the Pareto optimal solution set remains unchanged for 5 consecutive iterations to avoid invalid iterations; the crossover probability is set to 0.8 to generate new individuals. Too high a crossover probability will lead to population instability, while too low a crossover probability will lead to slow convergence; the mutation probability is set to 0.05 to maintain population diversity and avoid the algorithm getting stuck in local optima. Too high a mutation probability will lead to disorder in the algorithm, while too low a mutation probability will lead to local optima.

[0083] The algorithm solution process is divided into four stages: population initialization, non-dominated sorting, genetic operations (selection, crossover, mutation), and convergence judgment. Each stage strictly follows the optimization constraints to ensure that the generated individuals (parameter combinations) satisfy all constraints.

[0084] Population initialization phase: 50 parameter combinations (individuals) are randomly generated. The pulse frequency f of each individual is randomly selected within the range of [0.5Hz, 2Hz] (rounded to two decimal places), and the drug dosage d is randomly selected within the range of [0.1μL / pulse, 0.5μL / pulse] (rounded to two decimal places). After generating the initial population, each individual is checked for constraints. Individuals that violate any constraint are eliminated. If the number of eliminated individuals is less than 50, randomly generated individuals that meet the constraints are added to ensure that the initial population size is 50 and that all individuals meet the constraints. In the example, the individuals in the initial population include [0.8Hz, 0.2μL / pulse], [1.5Hz, 0.3μL / pulse], [1.9Hz, 0.4μL / pulse], etc., all of which have passed constraint checks and meet all constraints.

[0085] Non-dominated sorting phase: For each individual in the initial population, calculate its three objective function values ​​[-f1(x), f2(x), f3(x)]. Sort the individuals according to the non-dominated relation, defined as follows: if individual A's three objective function values ​​are all no worse than individual B's, and at least one of its objective function values ​​is better than individual B's, then individual A dominates individual B, and individual B is eliminated. Individuals not dominated by any individual constitute the first-level non-dominated solution set (Pareto optimal front), individuals dominated by the first level but not dominated by other individuals constitute the second level, and so on, completing the non-dominated sorting. After sorting, calculate the crowding degree of each individual. Crowding degree measures the density of individuals in the solution set; the higher the crowding degree, the better the diversity of individuals, avoiding the concentration of the solution set in a certain area and ensuring the comprehensiveness of the solution set.

[0086] Genetic operation phase: The selection operation adopts the roulette wheel selection method, based on the non-dominated level and crowding degree of individuals, selecting individuals with high fitness to enter the crossover and mutation phases. Individuals in the first level of non-dominated solution set have the highest probability of being selected, followed by the second level, and so on down. At the same time, individuals with high crowding degree are selected first to maintain population diversity. The crossover operation adopts the single-point crossover method, randomly selecting two individuals, randomly selecting a crossover point in the parameter dimension, exchanging the parameters after the crossover point, generating two new individuals. After crossover, the constraints of the new individuals are checked, eliminating individuals that violate the constraints and supplementing individuals that meet the constraints. The mutation operation adopts the random mutation method, randomly selecting a parameter (f or d) of an individual, and randomly mutating it by a small value within its value range (f mutation amplitude ±0.1Hz, d mutation amplitude ±0.05μL / pulse). After mutation, the constraints are checked, and individuals that violate the constraints are eliminated.

[0087] Convergence judgment phase: After completing one genetic operation, a new population is obtained. The non-dominated sorting and genetic operations are repeated until the number of iterations reaches 30, or the first-level non-dominated solution set remains unchanged for 5 consecutive iterations. The algorithm is considered to have converged. At this time, the first-level non-dominated solution set is the Pareto optimal solution set. This solution set contains multiple parameter combinations. Each combination cannot improve any objective without worsening the other two objectives. That is, each combination takes into account the three optimization objectives and there is no obvious superiority or inferiority.

[0088] In the example, after 30 iterations, the converged Pareto optimal solution set contains eight candidate parameter combinations: [0.9Hz, 0.3μL / pulse], [1.1Hz, 0.25μL / pulse], [1.2Hz, 0.3μL / pulse], [1.3Hz, 0.35μL / pulse], [1.4Hz, 0.2μL / pulse], [1.6Hz, 0.25μL / pulse], [1.7Hz, 0.3μL / pulse], and [1.8Hz, 0.2μL / pulse]. All combinations satisfy all constraints and each has a different emphasis on the three objectives. For example, [1.2Hz, 0.3μL / pulse] has better efficacy, lower side effects, and moderate cumulative dose; [1.8Hz, 0.2μL / pulse] has lower cumulative dose and extremely low side effects, but slightly weaker efficacy. This set of candidate parameter combinations is generated, providing a basis for subsequent optimal parameter selection.

[0089] S2034, based on the real-time optimization decision criteria, select the pulse frequency and dosage combination with the highest comprehensive score from the candidate parameter combination set to obtain the optimal pulse frequency and dosage parameter combination.

[0090] The core of this step is to establish a real-time optimization decision criterion, comprehensively score the candidate parameter combinations in the Pareto optimal solution set, and select the comprehensive optimal parameter combination. The key is that the scoring criterion aligns with actual drug administration needs, taking into account the three optimization objectives and prediction reliability. The specific implementation method is as follows: The real-time optimization decision-making criteria adopt a weighted comprehensive scoring method. Combining the core needs of mouse pulse pump administration (prioritizing efficacy, then reducing the risk of side effects, and finally reducing the cumulative drug amount), the scoring weights of three objective functions are set. The weight of efficacy index score is W1=0.5, the weight of side effect risk score is W2=0.3, and the weight of cumulative drug amount score is W3=0.2. The sum of the weights is 1, which ensures that the scoring focuses on key points and meets actual needs.

[0091] The calculation of the comprehensive score consists of three steps: First, the three objective function values ​​of each candidate parameter combination are standardized and converted into a score of 0-100 to eliminate the difference in the dimensions of the objective function values; then, the comprehensive score is calculated according to the weights; finally, the comprehensive score is corrected by combining the prediction confidence assessment to obtain the final score. The candidate parameter combination with the highest final score is the optimal parameter combination.

[0092] The first step is to standardize the objective function scores: The efficacy score S1 = f1(x) × 100, where f1(x) is the efficacy function value (0-1) for this parameter combination. The value of S1 ranges from 0 to 100; a higher score indicates better efficacy. The side effect risk score S2 = (1-f2(x) / f2_max) × 100, where f2_max is the maximum value of the side effect risk function for all individuals in the candidate solution set. The value of S2 ranges from 0 to 100; a higher score indicates lower side effect risk. The cumulative drug amount score S3 = (1-f3(x) / f3_max) × 100, where f3_max is the maximum value of the cumulative drug amount function for all individuals in the candidate solution set. The value of S3 ranges from 0 to 100; a higher score indicates lower cumulative drug amount.

[0093] The second step is to calculate the comprehensive score: the comprehensive score S = W1×S1 + W2×S2 + W3×S3, where W1 = 0.5, W2 = 0.3, W3 = 0.2. The comprehensive score S ranges from 0 to 100 points. The higher the score, the better the overall performance of the parameter combination.

[0094] The third step is score correction: Combining the prediction confidence assessment matrix of the physiological response trajectory, the mean prediction confidence of the 6 physiological characteristics corresponding to this parameter combination is calculated as C(0-1), and the correction coefficient is C×0.1+0.9. The corrected final score S_final=S×(C×0.1+0.9). If C≥0.8, the correction coefficient is 1.0 (no correction is needed); if C<0.6, the correction coefficient is 0.95, further reducing the score to reflect the impact of prediction uncertainty and ensure the reliability of the optimal parameter combination.

[0095] The example uses two candidate parameter combinations from the Pareto optimal solution set to illustrate the scoring process in detail: Candidate combination 1: x1 = [1.2Hz, 0.3μL / pulse], calculated as f1(x1) = 0.78, f2(x1) = 0.152, f3(x1) = 0.0144μL / (g) Candidate combination 2: x2 = [1.8 Hz, 0.2 μL / pulse], calculated f1(x2) = 0.65, f2(x2) = 0.10, f3(x2) = 0.0128 μL / (g) s).

[0096] First, determine f2_max = 0.2 (the side effect risk function value of a certain candidate combination) and f3_max = 0.018μL / (g) in the candidate solution set. s) (the cumulative drug amount function value of a candidate combination).

[0097] Standardized scores for candidate combination 1: S1 = 0.78 × 100 = 78 points, S2 = (1 - 0.152 / 0.2) × 100 = (1 - 0.76) × 100 = 24 points, S3 = (1 - 0.0144 / 0.018) × 100 = (1 - 0.8) × 100 = 20 points; comprehensive score S = 0.5 × 78 + 0.3 × 24 + 0.2 × 20 = 39 + 7.2 + 4 = 50.2 points; the mean confidence level of this combination C1 = 0.85 ≥ 0.8, the correction coefficient is 1.0, and the final score S_final1 = 50.2 × 1.0 = 50.2 points.

[0098] Standardized scores for candidate combination 2: S1 = 0.65 × 100 = 65 points, S2 = (1 - 0.10 / 0.2) × 100 = 50 points, S3 = (1 - 0.0128 / 0.018) × 100 ≈ (1 - 0.711) × 100 = 28.9 points; comprehensive score S = 0.5 × 65 + 0.3 × 50 + 0.2 × 28.9 = 32.5 + 15 + 5.78 = 53.28 points; the mean confidence level of this combination C2 = 0.82 ≥ 0.8, the correction coefficient is 1.0, and the final score S_final2 = 53.28 × 1.0 = 53.28 points.

[0099] The final scores of the other 6 candidate combinations were calculated. For example, candidate combination 3: x3 = [1.3Hz, 0.35μL / pulse], with a final score of 55.6 points; candidate combination 4: x4 = [1.1Hz, 0.25μL / pulse], with a final score of 49.8 points. After comparison, candidate combination 3 had the highest final score (55.6 points), which is the optimal combination of pulse frequency and drug dosage parameters.

[0100] After scoring, the optimal parameter combination is finally verified. Verification includes: whether all optimization constraints are met, whether the three objective function values ​​are within a reasonable range, and whether the mean confidence level is ≥0.6, ensuring the optimal parameter combination is safe, feasible, and reliable. In the example, candidate combination 3 [x3=[1.3Hz, 0.35μL / pulse]] shows that it meets all constraints, f1(x3)=0.82 (good efficacy), f2(x3)=0.18 (low side effects), and f3(x3)=0.0182μL / (g) (s) (not exceeding the threshold), confidence mean C3=0.83≥0.6, verification passed.

[0101] If multiple candidate combinations have the same final score (difference ≤ 0.5 points), the combination with the higher efficacy index score will be selected first; if the efficacy index scores are also the same, the combination with the higher side effect risk score will be selected first; if they are still the same, the combination with the higher cumulative drug dose score will be selected first to ensure the uniqueness of the selection.

[0102] Ultimately, by optimizing decision criteria in real time, the optimal combination of pulse frequency and dosage parameters was obtained. This combination maximizes efficacy, minimizes the risk of side effects and minimizes drug accumulation, and has high predictive reliability. It can adapt to the individual differences and physiological state of the mice, providing core parameter support for the subsequent generation of real-time control commands and the realization of dynamic regulation of pulse pump drug delivery.

[0103] S204, based on the optimal combination of pulse frequency and drug dosage parameters, a real-time control command is generated and sent to the pulse pump actuator to realize dynamic intelligent adjustment of pulse pump drug delivery.

[0104] Specifically, the optimal combination of pulse frequency and dosage parameters can be analyzed to calculate the specific control parameters of the pulse pump, including pulse width, pulse interval and single dose volume, and generate a set of pulse pump control parameters. The core of this step is to convert the optimal dosing parameters obtained from multi-objective optimization into specific control parameters that the pulse pump actuator can directly identify, and to establish the correlation between the optimal parameters and the pump's operating parameters, ensuring that the parameter calculations are accurate and closely match the pump's operating characteristics. The specific implementation method is as follows: The optimal pulse frequency and dosage parameter combination is the foundation of this step, with the format [x=[f,d]], where f is the optimal pulse frequency (in Hz) and d is the optimal dosage (in μL / pulse). This combination has been verified for safety and feasibility, and is consistent with the physiological tolerance range of individual mice and the pump's execution capability. The analysis process first extracts two core parameters from this combination, clarifies their physical meaning, and then, combined with the inherent operating parameters of the pulse pump, derives three key control parameters—pulse width, pulse interval, and single-dose volume—through fixed calculation formulas. These three parameters, in conjunction with the optimal parameters, determine the pulse pump's dosing rhythm and dosage.

[0105] First, calculate the pulse interval, which is the time interval (in milliseconds) between two adjacent administration pulses. It is inversely proportional to the pulse frequency and is calculated using the formula T_interval = 1000 / f, where 1000 is a unit conversion factor (converting seconds to milliseconds), and f is the optimal pulse frequency (Hz). The pulse interval value must strictly match the optimal frequency to ensure the administration rhythm is consistent with the optimization objective. Simultaneously, it must meet the minimum interval constraint of the pulse pump (not less than 500 ms). If the calculated result is less than the minimum interval, the minimum interval value should be used, and the reason for adjustment should be noted to ensure stable pump operation.

[0106] In the example, the optimal parameter combination is [x=[1.3Hz, 0.35μL / pulse]]. Substituting these parameters into the formula, the pulse interval is calculated as: T_interval=1000 / 1.3≈769.23ms. This value is greater than the minimum pulse pump interval of 500ms, meeting the constraint requirements. Therefore, the pulse interval is determined to be 769.23ms (rounded to two decimal places to ensure control accuracy). If the optimal frequency is 2.0Hz, the calculated T_interval=500ms, which is exactly equal to the minimum interval, requiring no adjustment. If the optimal frequency is 2.1Hz, the calculated T_interval≈476.19ms, which is less than the minimum interval, requiring adjustment to 500ms. An adjustment log is recorded for future reference.

[0107] Next, determine the single-dose volume. The single-dose volume is completely consistent with the optimal dose d, i.e., V_single = d, in μL / pulse. The core reason is that the optimal dose is defined as the dose delivered by the pulse pump in a single pulse; the two have the same physical meaning and require no additional calculation. It is only necessary to confirm that its value is within the single-dose volume range of the pulse pump (0.1-0.5 μL / pulse), consistent with the constraints in step four. In the example, the optimal dose d = 0.35 μL / pulse, therefore the single-dose volume V_single = 0.35 μL / pulse, which conforms to the pump's single-dose range, allowing for direct determination of this parameter.

[0108] Finally, the pulse width is calculated. Pulse width refers to the duration (in milliseconds) of a single drug delivery pulse. Its value is determined by the single-dose volume and the inherent drug delivery rate of the pulse pump. The drug delivery rate is a core inherent parameter of the pulse pump, set based on the characteristics of the pump's micro-drive mechanism, and fixed at 0.5 μL / ms (adapting to the micro-dose administration needs of mice, ensuring stable and controllable flow rate). The calculation formula is T_width = V_single / v, where V_single is the single-dose volume (μL / pulse), and v is the drug delivery rate (μL / ms). The pulse width must be controlled within the pump's allowable range (0.2-1.0 ms) to ensure a stable drug delivery rate and avoid dosage deviations due to excessively wide or narrow pulses.

[0109] In the example, the single-dose volume V_single = 0.35 μL / pulse, and the flow rate v = 0.5 μL / ms, substituting these values ​​into the formula, the pulse width is calculated as T_width = 0.35 / 0.5 = 0.7 ms. This value is within the pump's allowable range of 0.2-1.0 ms, meeting the requirements, and the pulse width is determined to be 0.7 ms. If the single-dose volume is 0.5 μL / pulse, the calculated T_width = 1.0 ms, reaching the pump's maximum pulse width; if the single-dose volume is 0.1 μL / pulse, the calculated T_width = 0.2 ms, which is the pump's minimum pulse width, and no adjustment is needed in either case. If flow rate fluctuations are caused by pump aging, the flow rate needs to be calibrated in real time and the pulse width recalculated.

[0110] After the three specific control parameters are calculated, they are integrated to form a pulse pump control parameter set. This set contains five core parameters: optimal pulse frequency f = 1.3Hz, optimal dosage d = 0.35μL / pulse, pulse interval T_interval = 769.23ms, pulse width T_width = 0.7ms, and single-dose volume V_single = 0.35μL / pulse. Simultaneously, parameter calculation timestamps (1ms accuracy) and verification identifiers (indicating whether parameters conform to pump constraints) are added to ensure the parameter set is complete and traceable. After the parameter set is generated, a final verification is performed to confirm that all parameter values ​​are within the pump's operating range and that the calculation process is error-free. Once verification is successful, the set is used for subsequent control command conversion.

[0111] The pulse pump control parameter set is converted into a control command format that the pulse pump can recognize, including start command, parameter setting command and start command, to generate a standardized control command sequence; The core of this step is to convert the structured set of control parameters into binary control instructions that the pulse pump actuator can recognize, clarifying the instruction format, instruction type, and sequence order to ensure accurate instruction transmission and error-free execution. The specific implementation method is as follows: The pulse pump actuator uses a fixed instruction format. The instructions are in binary string form, consisting of five parts: instruction header, instruction code, parameter values, checksum, and instruction tail. The total length is fixed at 16 bytes. Each part has a clear function and a fixed length to avoid instruction recognition confusion. The instruction header is fixed at 0xAA (1 byte) and is used to identify the start of the instruction, informing the actuator to prepare to receive the instruction. The instruction code (1 byte) is used to distinguish the instruction type; different instructions correspond to different instruction codes. The parameter values ​​(12 bytes) are used to store specific control parameters, arranged in a fixed order. The checksum (1 byte) is used to verify whether errors have occurred during instruction transmission. The instruction tail is fixed at 0x55 (1 byte) and is used to indicate the end of the instruction, informing the actuator that the instruction has been received completely.

[0112] The control commands are divided into three categories: start command, parameter setting command, and start command. These three commands are arranged in the order of "start command → parameter setting command → start command" to form a standardized control command sequence, ensuring that the pulse pump executes the drug delivery operation according to the correct procedure. The start command is used to wake up the pulse pump actuator, switching it from dormant to working state, ready to receive parameters and start commands. The parameter setting command is used to transmit the various parameters in the control parameter set to the pump controller to complete the parameter configuration. The start command is used to trigger the pulse pump to start executing the drug delivery operation according to the configured parameters. The three types of commands work together to ensure that the drug delivery process proceeds in an orderly manner.

[0113] First, a start instruction is constructed. The start instruction has no specific parameter values; its instruction code is fixed at 0x01. The instruction format is: instruction header (0xAA) + instruction code (0x01) + parameter value (12 bytes 0x00) + checksum + instruction tail (0x55). The checksum is calculated using an XOR checksum method. The formula is: checksum = instruction header ⊕ instruction code ⊕ each byte of parameter value ⊕ instruction tail. XOR checksum is simple to calculate, has strong real-time performance, and is suitable for the instruction verification requirements of micro pulse pumps, effectively detecting bit flipping errors during instruction transmission. In the example, the binary string of the start instruction is 0xAA0x010x000x00...0x00 (12 0x00s) 0xAA⊕0x01⊕0x55=0xFF0x55, meaning the complete start instruction is 0xAA0100000000000000000000000FF55.

[0114] Next, the parameter setting instructions are constructed. The instruction code for the parameter setting instructions is fixed at 0x02. The parameter value part stores five parameters from the control parameter set in a fixed order. Each parameter occupies a different byte length to ensure that the pump controller can correctly parse them: the pulse frequency f (1.3Hz) occupies 2 bytes, stored in decimal to hexadecimal format, retaining one decimal place, with a storage format of 0x0103 (corresponding to 1.3); the dosage d (0.35μL / pulse) occupies 2 bytes, with a storage format of 0x0035 (corresponding to 0.35); the pulse interval T_interval (769.23ms) occupies 3 bytes, with a storage format of 0x02FF03 (corresponding to 769.23); the pulse width T_width (0.7ms) occupies 2 bytes, with a storage format of 0x0007 (corresponding to 0.7); and the single dose volume V_single (0.35μL / pulse) occupies 3 bytes, with a storage format of 0x000035 (corresponding to 0.35). After the parameter value is concatenated, the checksum is calculated using the XOR check method. In the example, the checksum of the parameter setting instruction is 0xAA⊕0x02⊕each byte of the parameter value⊕0x55=0xEE. The complete parameter setting instruction is 0xAA020103003502FF030007000035EE55.

[0115] Finally, the startup instruction is constructed. The startup instruction has no specific parameter values, and the instruction code is fixed at 0x03. The instruction format is the same as the start instruction: instruction header (0xAA) + instruction code (0x03) + parameter value (12 bytes 0x00) + checksum + instruction tail (0x55). The checksum is calculated in the same way as the start instruction. In the example, the checksum of the startup instruction = 0xAA⊕0x03⊕0x55 = 0xFC. The complete startup instruction is 0xAA03000000000000000000000000FC55.

[0116] After the three types of instructions are constructed, they are concatenated in the order of "start instruction → parameter setting instruction → start instruction" to generate a standardized control instruction sequence. Instruction intervals (10ms between each instruction) are added to the sequence to prevent excessively dense instruction transmission from causing parsing errors in the pump controller. Simultaneously, an instruction identifier and timestamp are added to each instruction to facilitate subsequent monitoring of instruction execution status and traceability. In the example, the standardized control instruction sequence contains three instructions arranged sequentially, each labeled with a corresponding identifier and timestamp, ensuring a clear and identifiable sequence structure. The final standardized control instruction sequence can be directly used for subsequent wireless transmission.

[0117] The standardized control command sequence is sent to the pulse pump actuator via a wireless communication protocol, and the confirmation feedback from the actuator is received. The core of this step is to achieve wireless transmission and feedback reception of control commands, ensuring that commands can be accurately and quickly sent to the pulse pump, while simultaneously obtaining reception confirmation from the pump's actuator, forming a closed loop for command transmission and preventing command loss or transmission errors. The specific implementation method is as follows: Considering the miniaturization and low power consumption requirements of the mouse pulse pump, a low-power wireless communication protocol was selected. This protocol features low power consumption, moderate transmission distance (1-5 meters, suitable for experimental scenarios), and strong anti-interference capability (resistant to electromagnetic interference in the experimental environment), making it suitable for the communication needs of the mouse wearable miniature pulse pump. Communication parameters were set to match the communication capabilities of the pump controller: the baud rate was set to 9600bps to ensure that the command transmission rate matches the pump's parsing rate, avoiding untimely parsing; the communication frequency was set to 2.4GHz, which is an industrial unlicensed frequency band, requiring no additional license and reducing frequency interference with other experimental equipment; the data bits were 8 bits, the stop bits were 1 bit, and the parity bit was none, matching the binary format of the pulse pump control commands.

[0118] The establishment of a wireless communication link consists of three steps: First, the control terminal (the device used to generate instructions) activates the wireless communication module and sends a link establishment request signal, which includes the terminal identifier and the pulse pump's unique identifier (each pulse pump is assigned a unique identifier to avoid incorrect instruction delivery); second, the wireless communication module of the pulse pump actuator is in real-time monitoring mode. After receiving the link establishment request signal, it checks its own identifier against the identifier in the signal. If they match, it sends a link establishment confirmation signal; finally, after receiving the confirmation signal, the control terminal confirms that the wireless communication link has been successfully established and begins to issue a standardized control command sequence. If no confirmation signal is received, the link establishment request is sent repeatedly, up to a maximum of three times. If the link is still not established, a communication error message is triggered, instruction delivery is suspended, and communication faults (such as excessive distance or electromagnetic interference) are investigated.

[0119] The command issuance process strictly follows the standardized control command sequence. Each command is issued individually, and after issuance, a confirmation feedback from the pulse pump actuator is awaited. The feedback waiting time is set to 50ms. If no feedback is received within 50ms, the command is reissued, up to a maximum of two times, to ensure accurate command reception. The confirmation feedback signal is a 4-byte binary string in the format 0xAA + feedback code + checksum + 0x55. The feedback code indicates the command reception status: 0x00 indicates successful command reception and parsing, 0x01 indicates successful command reception but parsing error, and 0x02 indicates command reception failure. The checksum uses an XOR checksum, consistent with the control command verification method, to verify the accuracy of the feedback signal.

[0120] In the example, after establishing a wireless communication link, the control terminal first sends a start command. After sending the command, it waits for feedback. Within 50ms, it receives a feedback signal of 0xAA00A955 (feedback code 0x00, checksum 0xA9), confirming successful reception of the start command. Next, it sends a parameter setting command and receives a feedback signal of 0xAA00AB55, confirming successful reception and correct parsing. Finally, it sends a start command and receives a feedback signal of 0xAA00AD55, confirming successful reception. After all three commands are successfully received, the control terminal records a command sending log, including the sending time, command identifier, and feedback status, completing the command sending and feedback receiving process. If, after sending the parameter setting command, a feedback signal of 0xAA01AC55 (feedback code 0x01) is received, it indicates a command parsing error. The control terminal resends the parameter setting command, and upon resending, receives a feedback code of 0x00, completing command reception. If two resending attempts still result in parsing errors, a parameter error message is triggered, and the control parameter set and command format are re-verified.

[0121] Once the command is issued and all feedback is received successfully, the wireless communication link remains connected for subsequent pulse pump execution status feedback transmission. If a communication interruption occurs during command issuance, the control terminal re-establishes the link and continues to issue the incomplete command, ensuring the complete issuance of the entire command sequence and providing accurate command support for the pulse pump to perform drug delivery operations.

[0122] By monitoring the pulse pump's execution status and feedback from multi-channel biosensors in real time, a closed-loop control circuit is formed, and subsequent control commands are dynamically adjusted to achieve dynamic and intelligent regulation of pulse pump drug delivery.

[0123] The core of this step is to obtain the pulse pump execution status and mouse physiological feedback in real time through monitoring, compare it with preset standards, and adjust the control commands in a timely manner to form a closed-loop control of "parameter optimization → command issuance → execution monitoring → parameter adjustment". This ensures that the drug administration process conforms to the dynamic changes in the mouse's physiological state. The specific implementation method is as follows: Real-time monitoring consists of two parts: pulse pump execution status monitoring and multi-channel biosensor feedback monitoring. The two are performed simultaneously, with a monitoring frequency set at 10Hz (monitoring once every 100ms) to ensure timely capture of pump execution abnormalities and changes in the physiological state of mice. The monitoring data is transmitted to the control terminal in real time for subsequent adjustments and judgments.

[0124] The pulse pump's execution status is monitored via a wireless communication link. The pulse pump actuator collects its own operating parameters in real time and sends an execution status feedback signal every 100ms. The feedback signal includes three core components: pump operating status, current execution parameters, and fault indicators. The operating status is divided into three types: sleep, working, and fault, corresponding to feedback codes 0x00, 0x01, and 0x02, respectively. The current execution parameters include the actual pulse frequency, actual single-dose volume, and actual pulse interval, used to compare with preset control parameters to determine if there is any deviation in execution. The fault indicator is used to identify the type of pump fault (such as motor failure or abnormal flow rate). Feedback code 0x00 indicates no fault, while other codes correspond to specific fault types. After receiving the feedback signal, the control terminal compares the current execution parameters with the preset control parameters. If the deviation is ≤5%, it is considered normal execution; if the deviation is >5%, it is considered execution deviation, triggering a parameter calibration command to adjust the pump's execution parameters to ensure consistency with the preset parameters.

[0125] In the example, the preset pulse frequency is 1.3Hz. The actual execution frequency of the pulse pump received by the control terminal is 1.26Hz, with a deviation of (1.3-1.26) / 1.3≈3.08%≤5%, which is considered normal execution. If the actual execution frequency is 1.1Hz, the deviation is approximately 15.38%>5%, and the control terminal generates a parameter calibration command, which is sent to the pulse pump to adjust the speed of the pump drive motor, restoring the actual frequency to around 1.3Hz. If the received operating status feedback code is 0x02 (fault) and the fault identifier is 0x03 (abnormal flow rate), the control terminal triggers a fault alarm, suspends the drug delivery operation, and restarts after troubleshooting.

[0126] The multi-channel biosensor feedback monitoring follows the same sensor acquisition process as in step one. The multi-channel biosensor array acquires real-time signals from the mouse's electrocardiogram (ECG), electroencephalogram (EEG), blood oxygen saturation, and body temperature. Signal acquisition, denoising, and feature extraction are performed every 100ms, generating structured real-time physiological characteristic data, which is then transmitted to the control terminal. The control terminal compares the real-time physiological characteristic data with preset safe ranges for physiological parameters. Simultaneously, it combines this with the physiological response trajectory prediction results generated in step three to determine whether the mouse's physiological state is normal and whether drug side effects have occurred (such as ECG SDNN exceeding the safe range or abnormally high body temperature change rate).

[0127] If the real-time physiological characteristic data are all within the safe range and the deviation from the predicted physiological response trajectory is ≤10%, it indicates that the current drug administration parameters are suitable for the mouse's physiological state, and no adjustment of the control command is required. The pulse pump continues to perform the drug administration operation according to the current parameters. If the real-time physiological characteristic data exceed the safe range, or the deviation from the predicted trajectory is >10%, it indicates that the current drug administration parameters need to be adjusted. The control terminal calls the multi-objective optimization algorithm in step four again, and based on the real-time physiological characteristic data, redetermines the optimal combination of pulse frequency and drug dosage parameters. The four steps in step five are repeated to generate new control commands, which are sent to the pulse pump actuator to adjust the drug administration parameters and compensate for the impact of changes in the mouse's physiological state.

[0128] In the example, 500ms after the pulse pump administered the drug according to the optimal parameters, the control terminal detected a body temperature change rate of r = 0.0045℃ / ms in the mouse, exceeding the preset safety range (-0.004-0.004℃ / ms), and a deviation from the predicted trajectory of 12% > 10%, indicating that the current dosage was too high and might cause mild side effects. The control terminal immediately restarted the multi-objective optimization algorithm, recalculated a new optimal parameter combination [x = [1.2Hz, 0.3μL / pulse]] based on the real-time collected physiological characteristic data, and repeated the parameter parsing, instruction conversion, and issuance process in step five to generate new control instructions, adjusting the pulse frequency of the pulse pump to 1.2Hz and the single administration volume to 0.3μL / pulse, reducing the administration rate and decreasing the cumulative drug amount. After adjustment, real-time monitoring continued. If the body temperature change rate returned to the safe range, the new control parameters were maintained; if it remained abnormal, adjustments continued until the mouse's physiological state stabilized.

[0129] By combining pulse pump execution status monitoring and biosensor feedback monitoring, a complete closed-loop control circuit is formed, enabling dynamic and intelligent adjustment of drug administration parameters. This ensures that the pulse pump's administration rhythm and dosage are always adapted to the mouse's physiological state throughout the entire drug administration process, maximizing efficacy while effectively reducing the risk of side effects. It also avoids drug administration deviations caused by pulse pump malfunctions, ultimately achieving precise, safe, and intelligent regulation of pulse pump drug administration in mice.

[0130] Another embodiment of the present invention provides a dynamic parameter intelligent adjustment system for pulse pump drug delivery in mice, see [link to relevant documentation]. Figure 5 The system may include: The acquisition module 501 is used to acquire physiological feedback signals of drug-treated mice in real time through a multi-channel biosensor, and to denoise and extract features from the physiological feedback signals to generate structured real-time physiological feature data. The input module 502 is used to input the real-time physiological feature data into a pre-trained individualized drug response model and output the predicted physiological response trajectory and the corresponding confidence assessment. The determination module 503 is used to dynamically determine the current optimal combination of pulse frequency and drug dosage parameters based on the physiological response trajectory and the confidence assessment through a multi-objective optimization algorithm. The adjustment module 504 is used to generate real-time control commands based on the optimal combination of pulse frequency and drug dosage parameters and send them to the pulse pump actuator to realize dynamic intelligent adjustment of pulse pump drug delivery.

[0131] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0132] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0133] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for intelligent adjustment of dynamic parameters for pulse pump drug delivery in mice, characterized in that, The method includes: Physiological feedback signals of drug-treated mice are collected in real time using multi-channel biosensors, and the physiological feedback signals are denoised and feature extracted to generate structured real-time physiological feature data. The real-time physiological feature data is input into a pre-trained individualized drug response model, which outputs the predicted physiological response trajectory and the corresponding confidence assessment. Based on the physiological response trajectory and the confidence assessment, the current optimal combination of pulse frequency and drug dosage parameters is dynamically determined through a multi-objective optimization algorithm. Based on the optimal combination of pulse frequency and dosage parameters, a real-time control command is generated and sent to the pulse pump actuator to achieve dynamic and intelligent adjustment of pulse pump drug delivery.

2. The method according to claim 1, characterized in that, The process involves real-time acquisition of physiological feedback signals from drug-treated mice using multi-channel biosensors, followed by denoising and feature extraction of these signals to generate structured real-time physiological feature data, including: Multimodal raw physiological signals, including at least electrocardiogram (ECG), electroencephalogram (EEG), blood oxygen saturation, and body temperature data, were simultaneously acquired from drug-treated mice using a multi-channel biosensor array. The multimodal physiological raw signals were preprocessed, and wavelet transform algorithm was used to remove power frequency interference and electromyographic noise. Kalman filtering was used to eliminate motion artifacts and generate denoised physiological signals. Feature extraction is performed on the denoised physiological signals, and the RR interval variability of the electrocardiogram signal, the power spectrum characteristics of the electroencephalogram signal, the trend slope of blood oxygen saturation and the rate of change of body temperature are calculated to generate a multidimensional physiological feature vector. The multidimensional physiological feature vectors are standardized and time-aligned to construct a structured data matrix containing timestamps and feature identifiers, thereby generating structured real-time physiological feature data.

3. The method according to claim 2, characterized in that, The step of inputting the real-time physiological feature data into a pre-trained individualized drug response model and outputting the predicted physiological response trajectory and corresponding confidence assessment includes: Load a pre-trained individualized drug response model, which is based on a long short-term memory network architecture and trained using historical drug administration data and physiological feedback signals; Structured real-time physiological feature data is input into the loaded individualized drug response model, and the predicted values ​​of physiological state at multiple future time points are obtained through forward propagation calculation, generating a preliminary physiological response prediction sequence. The uncertainty estimation module based on the individualized drug response model uses the Monte Carlo dropout method to calculate the confidence interval of the predicted values ​​at each time point and generate a prediction confidence assessment matrix. By integrating the preliminary physiological response prediction sequence and the prediction confidence assessment matrix, a continuous physiological response trajectory with confidence intervals is constructed, generating a complete physiological response trajectory prediction result.

4. The method according to claim 3, characterized in that, The process of dynamically determining the optimal combination of pulse frequency and drug dosage parameters based on the physiological response trajectory and the confidence assessment using a multi-objective optimization algorithm includes: Construct a multi-objective optimization function, including maximizing efficacy indicators, minimizing side effect risk, and minimizing drug accumulation, and generate the corresponding multi-objective optimization problem; Based on the physiological response trajectory prediction results and confidence assessment matrix, a set of constraints is constructed, including the safe range of physiological parameters, the maximum dosing rate, and the minimum dosing interval, and optimization constraints are generated. A non-dominated sorting genetic algorithm is used to solve a multi-objective optimization problem. Under optimization constraints, a Pareto optimal solution set is searched to generate a set of candidate parameter combinations. Based on the real-time optimization decision-making criteria, the pulse frequency and dosage combination with the highest comprehensive score is selected from the candidate parameter combination set to obtain the optimal pulse frequency and dosage parameter combination.

5. The method according to claim 4, characterized in that, The step of generating real-time control commands based on the optimal pulse frequency and dosage parameter combination and sending them to the pulse pump actuator to achieve dynamic intelligent adjustment of pulse pump drug delivery includes: The optimal combination of pulse frequency and dosage parameters is analyzed, and the specific control parameters of the pulse pump are calculated, including pulse width, pulse interval and single dose volume, to generate a set of pulse pump control parameters. The pulse pump control parameter set is converted into a control command format that the pulse pump can recognize, including start command, parameter setting command and start command, to generate a standardized control command sequence; The standardized control command sequence is sent to the pulse pump actuator via a wireless communication protocol, and the confirmation feedback from the actuator is received. By monitoring the pulse pump's execution status and feedback from multi-channel biosensors in real time, a closed-loop control circuit is formed, and subsequent control commands are dynamically adjusted to achieve dynamic and intelligent regulation of pulse pump drug delivery.

6. A dynamic parameter intelligent adjustment system for pulse pump drug delivery in mice, characterized in that, The system includes: The acquisition module is used to acquire physiological feedback signals of drug-treated mice in real time through multi-channel biosensors, and to denoise and extract features from the physiological feedback signals to generate structured real-time physiological feature data. The input module is used to input the real-time physiological feature data into the pre-trained individualized drug response model and output the predicted physiological response trajectory and the corresponding confidence assessment. The determination module is used to dynamically determine the current optimal combination of pulse frequency and drug dosage parameters based on the physiological response trajectory and the confidence assessment through a multi-objective optimization algorithm. The adjustment module is used to generate real-time control commands based on the optimal combination of pulse frequency and drug dosage parameters and send them to the pulse pump actuator to realize dynamic intelligent adjustment of pulse pump drug delivery.

7. The system according to claim 6, characterized in that, The acquisition module is specifically used for: Multimodal raw physiological signals, including at least electrocardiogram (ECG), electroencephalogram (EEG), blood oxygen saturation, and body temperature data, were simultaneously acquired from drug-treated mice using a multi-channel biosensor array. The multimodal physiological raw signals were preprocessed, and wavelet transform algorithm was used to remove power frequency interference and electromyographic noise. Kalman filtering was used to eliminate motion artifacts and generate denoised physiological signals. Feature extraction is performed on the denoised physiological signals, and the RR interval variability of the electrocardiogram signal, the power spectrum characteristics of the electroencephalogram signal, the trend slope of blood oxygen saturation and the rate of change of body temperature are calculated to generate a multidimensional physiological feature vector. The multidimensional physiological feature vectors are standardized and time-aligned to construct a structured data matrix containing timestamps and feature identifiers, thereby generating structured real-time physiological feature data.

8. The system according to claim 7, characterized in that, The input module is specifically used for: Load a pre-trained individualized drug response model, which is based on a long short-term memory network architecture and trained using historical drug administration data and physiological feedback signals; Structured real-time physiological feature data is input into the loaded individualized drug response model, and the predicted values ​​of physiological state at multiple future time points are obtained through forward propagation calculation, generating a preliminary physiological response prediction sequence. The uncertainty estimation module based on the individualized drug response model uses the Monte Carlo dropout method to calculate the confidence interval of the predicted values ​​at each time point and generate a prediction confidence assessment matrix. By integrating the preliminary physiological response prediction sequence and the prediction confidence assessment matrix, a continuous physiological response trajectory with confidence intervals is constructed, generating a complete physiological response trajectory prediction result.

9. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images