Adaptive control method of analgesic infusion system based on multi-modal feedback

Abstract

The invention discloses a self-adaptive control method of an analgesia infusion system based on multi-modal feedback, and relates to the field of analgesia infusion automation control, and the method comprises the steps: inputting a feature vector into a multi-modal fusion decision-making layer, and calculating a comprehensive pain index PI by adopting a confidence-weighted self-adaptive fusion algorithm; taking the comprehensive pain index PI, the drug cumulant and the physiological stability index as state input, and inputting the state input into a self-adaptive controller based on reinforcement learning; the adaptive controller outputs a basic infusion rate adjustment amount, a pulse dose adjustment amount, and a locking time adjustment amount. According to the invention, through multi-source signal acquisition and confidence coefficient weighted adaptive fusion, the accuracy and robustness of pain assessment are improved; the hierarchical reinforcement learning controller realizes personalized dose optimization on the premise of safety priority, and gives consideration to both analgesic effect and physiological stability; and the safety monitoring module is combined with multi-modal cross validation and pharmacokinetic prediction, so that the safety risk is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of automated control technology for analgesic infusion, and in particular to an adaptive control method for analgesic infusion systems based on multimodal feedback. Background Technology

[0002] In postoperative analgesia or chronic pain management, intravenous administration is a commonly used method of pain relief. Traditional analgesia infusion systems often use fixed-dose infusion or simple feedback control based on a single physiological signal, which makes it difficult to adapt to individual differences among patients, dynamic changes in pain status, and the influence of complex clinical environments.

[0003] In existing technologies, single signal feedback is easily interfered with, leading to inaccurate pain assessment, and fixed-weight fusion methods cannot cope with dynamic fluctuations in the quality of signals from different modalities.

[0004] Meanwhile, existing controllers lack comprehensive consideration of drug cumulative effects, physiological stability, and individual differences in drug response, which can easily lead to safety risks such as insufficient analgesia, excessive sedation, or respiratory depression.

[0005] Furthermore, when patients use other central nervous system medications in combination, existing systems struggle to achieve precise dose compensation, further reducing the safety and effectiveness of infusion control. Based on these issues, there is an urgent need for an adaptive control scheme that can accurately sense pain status, dynamically adapt to individual differences, and ensure infusion safety. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides an adaptive control method for analgesic infusion systems based on multimodal feedback. The technical solution adopted is as follows:

[0007] An adaptive control method for analgesic infusion systems based on multimodal feedback includes the following steps:

[0008] Step 1: Collect multi-source signals from the patient using multi-source sensors. These multi-source signals include physiological signals, behavioral response signals, and subjective feedback signals.

[0009] Step 2: Preprocess and extract features from the multi-source signals to obtain feature vectors for each mode;

[0010] Step 3: Input the feature vector into the multimodal fusion decision layer, and use a confidence-weighted adaptive fusion algorithm to calculate the comprehensive pain index PI;

[0011] Step 4: The comprehensive pain index (PI), cumulative drug dosage, and physiological stability index are used as state inputs and fed into the reinforcement learning-based adaptive controller.

[0012] Step 5: The adaptive controller outputs the basic infusion rate adjustment, pulse dose adjustment, and lockout time adjustment.

[0013] Step 6: The safety monitoring module verifies the safety of the control commands and ensures the safety range through multimodal cross-validation and pharmacokinetic prediction.

[0014] Step 7: Execute the adjusted infusion parameters and record the system response data for online learning and optimization.

[0015] Optionally, in step 1, the multi-source sensor includes an electrocardiogram sensor, a skin conductance sensor, a forehead electroencephalogram sensor, a photoplethysmography (PPG) sensor, a miniature camera, a pressure-sensitive mattress, and a wireless feedback terminal.

[0016] Optionally, the multimodal fusion decision layer adopts a layered fusion architecture, including:

[0017] First layer: Feature extraction and preliminary scoring within each modality;

[0018] The second layer: dynamic weight allocation based on the signal quality index;

[0019] The third layer: Temporal context-aware trend correction, which uses a temporal convolutional network to analyze the evolution of pain states.

[0020] Optionally, the reinforcement learning-based adaptive controller adopts a hierarchical structure, including a security layer, an optimization layer, and a personalization layer;

[0021] The safety layer is based on safety interventions using hard physiological thresholds;

[0022] The optimization layer is based on dose optimization control using a deep Q-network;

[0023] The personalization layer is based on an individual feature adaptation module using transfer learning.

[0024] Optionally, in step 4, the reward function for the adaptive controller reinforcement learning is:

[0025] ;

[0026] in, It is the reward value at time t. It is the comprehensive pain index at time t. The drug increment at time t Let be the variance of physiological indicators at time t. Rate the comfort level at time t. , , and It is the weighting coefficient.

[0027] Optionally, the safety monitoring module includes a pharmacokinetic prediction unit, a multimodal consistency check, and a risk warning unit;

[0028] The pharmacokinetic prediction unit predicts plasma concentration based on a personalized pharmacokinetic model; the multimodal consistency check initiates a review when the difference in pain assessment between different modalities exceeds a threshold; and the risk warning unit predicts the risk of respiratory depression based on a logistic regression model.

[0029] Optionally, the formula for predicting plasma concentration using a personalized pharmacokinetic model is:

[0030] ;

[0031] in It is after time t plasma concentration at time, Let be the plasma concentration at time t. To eliminate the rate constant, For infusion rate, This represents the plasma clearance rate of analgesic drugs.

[0032] Optionally, a drug interaction compensation module may also be included, which automatically adjusts the dosage algorithm when it detects that the patient is using other central nervous system drugs, and calculates the compensation factor using the following formula. :

[0033] ;

[0034] Where [BZD] represents the equivalent dose of benzodiazepines, and [AH] represents the equivalent dose of antihistamines. This is equivalent to the dose of other central nervous system depressants.

[0035] An analgesic infusion system includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the steps of an adaptive control method for the analgesic infusion system based on multimodal feedback.

[0036] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of an adaptive control method for an analgesic infusion system based on multimodal feedback.

[0037] In summary, the present invention has at least one of the following beneficial technical effects:

[0038] This invention provides an adaptive control method for analgesic infusion systems based on multimodal feedback. By acquiring multi-source signals and using confidence-weighted adaptive fusion, the accuracy and robustness of pain assessment are improved. A hierarchical reinforcement learning controller enables personalized dose optimization under the premise of safety priority, taking into account both analgesic effect and physiological stability. The safety monitoring module, combined with multimodal cross-validation and pharmacokinetic prediction, effectively reduces safety risks. The drug interaction compensation module adapts to combined drug use scenarios, further improving the accuracy of dose control. The online learning optimization function enables the system to continuously adapt to changes in patient status, significantly improving the safety, effectiveness, and personalization of analgesic infusion. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the adaptive control method for analgesic infusion systems based on multimodal feedback, as described in this invention.

[0040] Figure 2 This is a schematic diagram of the operation process of the adaptive controller of the present invention.

[0041] Figure 3 This is a block diagram of the overall structure of the analgesic infusion system based on multimodal feedback according to the present invention. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings.

[0043] This invention discloses an adaptive control method for analgesic infusion systems based on multimodal feedback.

[0044] Reference Figures 1-3 Example 1, an adaptive control method for an analgesic infusion system based on multimodal feedback, includes the following steps:

[0045] Step 1: Collect multi-source signals from the patient using multi-source sensors. These multi-source signals include physiological signals, behavioral response signals, and subjective feedback signals.

[0046] Step 2: Preprocess and extract features from the multi-source signals to obtain feature vectors for each mode;

[0047] Step 3: Input the feature vector into the multimodal fusion decision layer, and use a confidence-weighted adaptive fusion algorithm to calculate the comprehensive pain index PI;

[0048] Step 4: The comprehensive pain index (PI), cumulative drug dosage, and physiological stability index are used as state inputs and fed into the reinforcement learning-based adaptive controller.

[0049] Step 5: The adaptive controller outputs the basic infusion rate adjustment, pulse dose adjustment, and lockout time adjustment.

[0050] Step 6: The safety monitoring module verifies the safety of the control commands and ensures the safety range through multimodal cross-validation and pharmacokinetic prediction.

[0051] Step 7: Execute the adjusted infusion parameters and record the system response data for online learning and optimization.

[0052] In Example 2, step 1, the multi-source sensor includes an electrocardiogram sensor, a skin conductance sensor, a frontal electroencephalogram sensor, a photoplethysmography (PPG) sensor, a miniature camera, a pressure-sensitive mattress, and a wireless feedback terminal.

[0053] By adopting the above technical solutions, a closed-loop adaptive system of perception-fusion-decision-control-verification-optimization is constructed. The accuracy of pain assessment is improved by relying on the complementarity of multimodal signals, personalized dynamic dose adjustment is achieved through reinforcement learning, and infusion safety is ensured by combining multi-layer safety mechanisms.

[0054] Specifically, pain states have multi-dimensional representational characteristics, and a single signal cannot fully and accurately reflect the true degree of pain. Therefore, we first collect pain-related information covering the three dimensions of physiological, behavioral, and subjective pain through multi-source signal acquisition. Then, we remove noise interference and extract core features through preprocessing to provide a high-quality data foundation for subsequent fusion. The confidence-weighted adaptive fusion algorithm can dynamically allocate weights according to the quality of each modality signal and combine time context awareness correction to achieve dynamic and accurate assessment of pain state and output a comprehensive pain index. The reinforcement learning adaptive controller takes the comprehensive pain index and cumulative drug dose as state inputs. By designing a reward function that takes into account analgesic effect, drug dosage, physiological stability, and comfort, the controller is driven to learn the optimal infusion parameters autonomously. At the same time, the hierarchical structure ensures personalized optimization under the premise of safety priority. The safety monitoring module avoids misjudgment of a single signal through multimodal cross-validation and avoids the risk of drug overdose in advance by combining pharmacokinetic prediction. Finally, through execution feedback and online learning, the system continuously adapts to individual differences in patients and changes in pain state, forming a closed-loop optimization.

[0055] The principle of multi-source sensor selection is based on the physiological, behavioral, and subjective multi-dimensional response characteristics triggered by pain. Sensors capable of accurately capturing corresponding signals are selected to achieve comprehensive and synchronous acquisition of multi-dimensional signals. Specifically, an electrocardiogram (ECG) sensor collects the patient's ECG signals, analyzing features such as heart rate variability to reflect the autonomic nervous system stress state triggered by pain; a skin conductance sensor captures changes in skin conductivity, which are directly related to sympathetic nerve activation triggered by pain; a frontal electroencephalogram (EEG) sensor collects brainwave signals, reflecting changes in central nervous system activity related to pain; a photoplethysmography (PPG) sensor acquires pulse wave signals, extracting physiological indicators such as pulse rate and blood oxygen saturation, indirectly reflecting circulatory system fluctuations triggered by pain; a miniature camera captures real-time behavioral response signals such as facial expressions and limb movements, which are important intuitive bases for pain assessment; a pressure-sensitive mattress senses changes in patient position and limb movement amplitude to assist in judging abnormal physical activity caused by pain; and a wireless feedback terminal provides patients with a direct input channel for subjective pain perception, achieving accurate acquisition of subjective feedback signals. The combination of multiple sensors can cover different dimensions of pain response, ensuring the complementarity and comprehensiveness of the collected signals, and providing solid data support for subsequent multimodal fusion evaluation.

[0056] Example 3: The multimodal fusion decision layer adopts a layered fusion architecture, including:

[0057] First layer: Feature extraction and preliminary scoring within each modality;

[0058] The second layer: dynamic weight allocation based on the signal quality index;

[0059] The third layer: Temporal context-aware trend correction, which uses a temporal convolutional network to analyze the evolution of pain states.

[0060] Example 4: The reinforcement learning-based adaptive controller adopts a hierarchical structure, including a security layer, an optimization layer, and a personalization layer.

[0061] The safety layer is based on safety interventions using hard physiological thresholds;

[0062] The optimization layer is based on dose optimization control using a deep Q-network;

[0063] The personalization layer is based on an individual feature adaptation module using transfer learning.

[0064] By adopting the above technical solution, the multimodal fusion decision layer employs a hierarchical fusion architecture based on the heterogeneity of multimodal signals, the dynamic fluctuation of signal quality, and the temporal evolution of pain states. Through a hierarchical and progressive processing logic, it gradually improves the accuracy and reliability of pain assessment. The first layer involves feature extraction and preliminary scoring within each modality. Its core purpose is to establish a basic assessment basis for each modality signal. Targeted feature extraction algorithms are used to mine core information related to pain in each modality. Simultaneously, preliminary scoring is performed to screen effective features and eliminate invalid or interfering features, providing high-quality single-modality foundational data for subsequent fusion and preventing low-quality single-modality features from affecting the overall fusion effect. The second layer uses dynamic weight allocation based on signal quality indices. This is because different modal signals are susceptible to interference in different clinical scenarios, and signal quality varies. Fixed weights cannot adapt to dynamic changes. By calculating the signal quality index for each modality, higher weights are assigned to high-quality signals, while lower weights are assigned to low-quality signals, or they are even temporarily eliminated. This ensures that the fusion result relies more on reliable signal sources, improving the robustness of the fusion assessment. The third layer, temporal context-aware trend correction, is based on the principle that pain states are continuous and evolving. Signal assessment at a single moment is easily affected by transient interference. Using a temporal convolutional network can effectively capture the temporal dependencies of pain signals, analyze the evolutionary trend of pain states, and correct the current fusion result. This avoids misjudgments caused by transient signal fluctuations, making the comprehensive pain index more closely reflect the dynamic changes in the patient's actual pain. The three-layer architecture works synergistically to achieve progressive optimization from single-modal basic processing to dynamic weight fusion, and then to temporal trend correction, comprehensively improving the accuracy and stability of multimodal fusion assessment.

[0065] The hierarchical structure of the reinforcement learning-based adaptive controller revolves around the core requirements of safety priority, efficacy optimization, and individualized adaptation in analgesic infusion control. By dividing functional boundaries into layers, each layer focuses on specific objectives and works collaboratively, ensuring the safety, effectiveness, and personalization of the control process. The safety layer is based on hard physiological thresholds for safety intervention. Its core principle is to prioritize patient safety, pre-setting safety thresholds for key physiological indicators such as respiratory rate, blood oxygen saturation, and blood pressure. When these thresholds are exceeded, emergency intervention commands are triggered directly, such as pausing infusion or reducing the dosage, quickly mitigating risks to patient safety and providing a safety boundary for subsequent optimized control. The optimization layer is based on dose optimization control using deep Q-networks. Within the safety boundaries defined by the safety layer, it utilizes the reinforcement learning capabilities of deep Q-networks to continuously learn the mapping relationship between the overall pain index, cumulative drug dosage, and infusion parameter adjustments. Guided by a pre-set reward function, it autonomously explores the optimal infusion dose adjustment strategy, achieving a balance between analgesic effect and drug dosage optimization. The personalized layer is based on an individual feature adaptation module using transfer learning. The principle is that individual differences in age, weight, underlying diseases, and drug metabolism capabilities among patients lead to varying responses to analgesics. Transfer learning technology is used to train a general model from a large amount of existing patient data, which is then quickly adapted to the individual characteristics of new patients. This eliminates the need to retrain the model for each patient, shortening the model adaptation time and improving the accuracy of infusion parameter adaptation to individuals. This three-layer structure forms a progressive control logic of "safety safety net - efficacy optimization - individual adaptation," ensuring both infusion safety and personalized analgesic effect optimization.

[0066] In Example 5, step 4, the reward function for the adaptive controller reinforcement learning is:

[0067] ;

[0068] in, It is the reward value at time t. It is the comprehensive pain index at time t. The drug increment at time t Let be the variance of physiological indicators at time t. Rate the comfort level at time t. , , and It is the weighting coefficient.

[0069] By adopting the above technical solution, based on the multi-objective optimization requirements of analgesic infusion control, a composite function with multi-dimensional reward and punishment components is constructed to guide the reinforcement learning controller to autonomously learn the optimal infusion strategy that balances analgesic effect, pharmacoeconomics, physiological stability, and patient comfort. The core logic is to reinforce desired behavior with positive rewards and constrain undesirable behavior with negative punishments, so that the controller gradually approaches the optimal control strategy that balances multiple objectives through iterative learning.

[0070] Specifically, 1 is the product of the sub-items and weighting coefficients of the comprehensive pain index. The principle is that the lower the comprehensive pain index, the better the analgesic effect. This sub-item is output by the positive reward-guided controller to reduce the pain index. At the same time, the weighting coefficient can adjust the importance of this target according to the clinical analgesia priority.

[0071] The product of the drug increment item and the weighting coefficient is a negative penalty item. The principle is that excessive infusion of analgesics will increase safety risks and medical costs. By imposing a penalty on the drug increment, the controller is prevented from excessively increasing the drug dosage in pursuit of short-term analgesia.

[0072] The product of the variance component of the physiological indicators and the weighting coefficient is also a negative penalty term. The principle is that drastic fluctuations in physiological indicators reflect the instability of the patient's physical state, which may be caused by adverse drug reactions or pain stress. By penalizing this component, the controller is guided to maintain the stability of the patient's physiological state. The product of the comfort score component and the weighting coefficient is a positive reward term. The principle is that patient comfort is one of the important goals of analgesia treatment and directly reflects the treatment experience. By rewarding the control behavior that improves comfort, the treatment plan can be more in line with the patient's subjective needs.

[0073] The weighting coefficients are set based on clinical treatment priorities, balancing the relative importance of four objectives: analgesic effect, drug safety, physiological stability, and comfort. Ultimately, the multi-objective optimization requirements are transformed into reward and punishment signals that the reinforcement learning controller can perceive through this composite reward function, driving the controller to continuously learn and output the optimal infusion parameter adjustment strategy.

[0074] Example 6: The safety monitoring module includes a pharmacokinetic prediction unit, a multimodal consistency check, and a risk warning unit;

[0075] The pharmacokinetic prediction unit predicts plasma concentration based on a personalized pharmacokinetic model; the multimodal consistency check initiates a review when the difference in pain assessment between different modalities exceeds a threshold; and the risk warning unit predicts the risk of respiratory depression based on a logistic regression model.

[0076] Example 7: The formula for predicting plasma concentration using a personalized pharmacokinetic model is as follows:

[0077] ;

[0078] in It is after time t plasma concentration at time, Let be the plasma concentration at time t. To eliminate the rate constant, For infusion rate, This represents the plasma clearance rate of analgesic drugs.

[0079] By adopting the above technical solutions, a multi-dimensional collaborative safety assurance system is constructed. Through a three-pronged approach of advance prediction, cross-validation, and risk warning, safety risks during analgesia infusion are comprehensively mitigated, providing a reliable safety boundary for adaptive control. The core logic involves setting up dedicated monitoring units for three key risk points during analgesia infusion: drug overdose, misjudgment of pain assessment, and critical complications, enabling early detection and intervention of risks.

[0080] Specifically, the pharmacokinetic prediction unit operates on the principle of drug metabolism in the body. Considering individual differences in drug metabolism among patients, a general pharmacokinetic model cannot accurately predict individual plasma drug concentrations. Therefore, an individualized model is used to predict changes in plasma concentration in advance, avoiding adverse reactions caused by drug concentrations exceeding safe limits. The multimodal consistency check acknowledges the complementarity of multimodal signals but also recognizes potential assessment biases. When the difference in pain assessment results across different modalities exceeds a preset threshold, it indicates potential interference or misjudgment in the current single-modal assessment. Initiating a review process allows for re-analysis of the original multimodal signals, improving the reliability of pain assessment and preventing dosage adjustment errors due to assessment bias. The risk warning unit recognizes respiratory depression as one of the most critical safety hazards of analgesics. A logistic regression model trained on clinical data can effectively fit the correlation between respiratory depression and related influencing factors. By inputting physiological indicators such as drug concentration, the risk of respiratory depression is predicted, issuing early warning signals and buying time for clinical intervention. These three units work synergistically to form a complete safety monitoring chain, from drug concentration prediction to assessment result verification and critical risk warning, ensuring the safety of the infusion process.

[0081] Personalized pharmacokinetic models predict plasma concentrations based on the first-order elimination kinetics of drugs in vivo. They dynamically track the balance between drug infusion and elimination processes, enabling accurate prediction of individual plasma drug concentrations. The core logic is that plasma drug concentration is determined by two components: the residual drug concentration after metabolic elimination and the concentration increment resulting from the newly infused drug. Concentration prediction is achieved by quantifying the contributions of these two components.

[0082] Specifically, the principle behind the first term in the formula is that after a drug enters the body, it will naturally disappear over time, following an exponential decay law. The plasma concentration Cpt at time t will, after a time interval Δt, be Cpt multiplied by the negative of e. The power, where The elimination rate constant reflects an individual's ability to eliminate drugs. The larger the plasma concentration, the faster the drug is eliminated; the principle behind the second point is that continuous drug infusion over a time interval Δt will increase plasma concentration, and the infusion rate... The rate at which a drug enters the body is determined by plasma clearance (CL), which reflects the rate at which the drug is eliminated from the body. The ratio of the two is... This represents the theoretical increment of drug concentration under steady-state conditions, while This corrects for the effect of incomplete drug elimination within the time interval Δt, specifically the concentration increase resulting from the uneliminated portion of the newly infused drug. The entire formula integrates the two core processes of drug elimination and infusion, combined with individualized... The CL parameter can accurately predict plasma concentrations at different time points, providing a quantitative basis for the safety monitoring module to determine whether drug concentrations exceed the safe range. It also adapts to individual metabolic differences among different patients, improving the accuracy of predictions.

[0083] Example 8 also includes a drug interaction compensation module, which automatically adjusts the dosage algorithm when it detects that the patient is using other central nervous system drugs, and calculates the compensation factor using the following formula. :

[0084] ;

[0085] Where [BZD] represents the equivalent dose of benzodiazepines, and [AH] represents the equivalent dose of antihistamines. This is equivalent to the dose of other central nervous system depressants.

[0086] By employing the above technical solution, the drug interaction compensation module addresses scenarios where patients may use multiple central nervous system drugs in combination in clinical settings. It mitigates the safety risks associated with synergistic effects of different drugs and achieves precise analgesic dosage matching. The core logic is that central nervous system drugs and analgesics may have synergistic effects, potentially increasing the risk of adverse reactions such as sedation and respiratory depression. Therefore, it is necessary to monitor the combined use of these drugs and dynamically adjust the analgesic dosage to counteract the adverse effects of drug interactions.

[0087] Specifically, the module first monitors the patient's medication information in real time. When it detects the use of other central nervous system drugs, it activates the dosage adjustment algorithm. The core of this algorithm is to quantify the strength of drug interactions through compensation factors, and then correct the infusion dosage.

[0088] The compensation factor formula is based on the differences in the synergistic effects of different types of central nervous system (CNS) drugs and analgesics. It quantifies the total synergistic effect through weighted summation and then converts it into a dose adjustment coefficient. The core logic of the denominator in the formula is to assign corresponding weights to the equivalent doses of different types of CNS drugs according to their synergistic effect strength. The weight coefficients are set based on the risk of various drugs enhancing the adverse reactions of analgesics in clinical data. Benzodiazepines have the strongest synergistic effect, hence their highest weight coefficient. The weight coefficients for antihistamines and other CNS depressants correspond to their synergistic effect strengths in that order. By summing the products of the equivalent doses of various drugs and their corresponding weights, the total synergistic effect of combined medication can be quantified. The larger the sum, the higher the synergistic risk; the larger the denominator, the smaller the compensation factor. As an adjustment coefficient for analgesic drug dosage, a smaller compensation factor means a greater reduction in the analgesic drug dosage. Thus, through precise dose compensation, the safety risks of excessive sedation and respiratory depression caused by combined medication can be effectively avoided while ensuring analgesic efficacy. Meanwhile, using equivalent doses for calculation can unify the dosage measurement standards for different types and specifications of drugs, ensure the accuracy of synergistic effect quantification, and improve the universality and adaptability of the compensation algorithm.

[0089] Example 9: An analgesic infusion system includes a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the steps of an adaptive control method for an analgesic infusion system based on multimodal feedback.

[0090] Example 10: A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of an adaptive control method for an analgesic infusion system based on multimodal feedback.

[0091] The following describes the implementation principle of the present invention using specific embodiments:

[0092] This study is designed for patient-controlled analgesia (PCA) administration in patients following laparoscopic cholecystectomy. The patient is 55 years old, weighs 60 kg, and is ASA class II. Postoperatively, the patient requires continuous analgesia for 48 hours. The patient has no history of central nervous system disease, did not use sedatives preoperatively, and experiences moderate postoperative pain. The pain level dynamically changes with the recovery process, requiring adaptive control to achieve a balance between analgesic efficacy and medication safety.

[0093] like Figure 3 As shown, the system configuration is as follows:

[0094] The multi-source sensor configuration is as follows: an electrocardiogram (ECG) sensor with a sampling frequency of 250Hz and lead II for capturing heart rate and heart rate variability-related signals; a skin conductance sensor with a sampling frequency of 100Hz, with electrodes attached to the inside of the patient's palm to collect changes in skin conductance; an 8-lead frontal EEG sensor to collect EEG signals from the frontal region; a photoplethysmography (PPG) sensor with a sampling frequency of 125Hz, worn on the tip of the patient's index finger to collect pulse wave signals and blood oxygen saturation data; a miniature camera mounted on the crossbar of the IV stand, aimed at the patient's face and upper body to capture facial expressions and limb movements; a pressure-sensitive mattress with a 20×20 point sensor array, placed on the patient's bed to sense changes in body position and limb movement; and a wireless feedback terminal with a touch-screen control panel, allowing for a pain rating scale of 0-10 points, supporting subjective feedback input from the patient at least once every 5 minutes.

[0095] The processor is an industrial-grade processor that supports real-time data storage and algorithm computation. The system power supply adopts a dual-redundant design to ensure continuous operational stability.

[0096] like Figure 2 The specific operating procedure is shown below:

[0097] Step 1: Multi-source signal acquisition:

[0098] After the system starts, all sensors synchronously begin data acquisition, with an acquisition cycle of 1 second, continuously acquiring data without interruption. The electrocardiogram (ECG) sensor captures the patient's ECG signals in real time, outputting a set of heart rate data and raw waveforms every second; the skin conductance sensor outputs skin conductance values ​​every second, ranging from 0-50 μS; the frontal EEG sensor outputs the amplitude of 8-lead EEG signals every second, covering the frequency range of delta waves, theta waves, alpha waves, and beta waves; the photoplethysmography (PPG) sensor outputs pulse rate, blood oxygen saturation, and pulse wave amplitude data every second; the miniature camera captures images of the patient's facial expressions and upper body postures frame by frame, extracting key behavioral features every second; the pressure-sensitive mattress outputs body position distribution data and quantified values ​​of limb movement amplitude every second; the wireless feedback terminal records the subjective pain score in real time after the patient triggers input, retains the previous input value when not triggered, and automatically prompts the patient to update every 5 minutes.

[0099] Step 2, Preprocessing and Feature Extraction:

[0100] The preprocessing procedure is as follows: For electrocardiogram (ECG) signals, a 5Hz high-pass filter and a 50Hz notch filter are used to remove baseline drift and power frequency interference. The time-domain feature of heart rate variability, the standard deviation of the RR interval, is extracted (range 50-150ms); the frequency-domain feature, the power ratio of the low-frequency band to the high-frequency band, is extracted (range 0.5-3.0). For electrodermal conductance (EDA) signals, a 1Hz low-pass filter is used to remove high-frequency noise. The mean and rate of change of EDA are extracted (range -5-5μS / s). For electroencephalogram (EEG) signals, a 0.5Hz high-pass filter and a 30Hz low-pass filter are used to extract the power spectral density of each lead. The power ratio of the β wave to the α wave is calculated (range 1.2-4.5). Pulse wave signals were filtered using a 2Hz high-pass filter to extract pulse rate (range 60-120 beats / min), blood oxygen saturation (range 95%-100%), and pulse wave conduction time (range 100-300ms). Behavioral signals were extracted using a facial feature point detection algorithm to extract 68 facial key points and calculate feature values ​​for pain-related expressions such as frowning, teeth clenching, and eyelid closure (range 0-1). Limb movements were extracted using a posture recognition algorithm to extract range of motion (range 0-10cm). Body position changes were calculated using pressure-sensitive data to determine center of gravity shift (range 0-20cm). Subjective feedback signals were directly extracted as raw scores (range 0-10 points) without preprocessing.

[0101] Step 3: Multimodal fusion decision-making and comprehensive pain index calculation:

[0102] The multimodal fusion decision layer performs calculations according to a hierarchical architecture: The first layer extracts and preliminarily scores features within each modality: In the physiological modality, ECG features are based on the standard deviation of the RR interval and the low-frequency / high-frequency power ratio to output a physiological pain score, ranging from 0 to 10; skin conductance features are based on the mean and rate of change of conductivity to output a physiological pain score, ranging from 0 to 10; EEG features are based on the β / α power ratio to output a physiological pain score, ranging from 0 to 10; pulse wave features are based on pulse rate and conduction time to output a physiological pain score, ranging from 0 to 10. In the behavioral modality, facial expression features are based on pain-related expression feature values ​​to output a behavioral pain score, ranging from 0 to 10; limb activity features are based on the range of activity and changes in body position to output a behavioral pain score, ranging from 0 to 10; the subjective modality directly uses the patient's input score as the subjective pain score, ranging from 0 to 10.

[0103] The second layer is based on dynamic weight allocation of the signal quality index: calculate the signal quality index of each mode, ranging from 0 to 1, where the signal quality index = 1 - noise amplitude / signal amplitude. The weighting for each signal quality index is as follows: ECG signal quality index below 0.6: 0.1; 0.2 between 0.6 and 0.8: 0.3; above 0.8: 0.3. For skin conductance signal quality index: below 0.6: 0.1; 0.15 between 0.6 and 0.8: 0.2. For electroencephalogram (EEG): below 0.6: 0.1; 0.15 between 0.6 and 0.8: 0.2. For pulse wave signal quality index: below 0.6: 0.05; 0.1 between 0.6 and 0.8: 0.15; above 0.8: 0.15. For facial expression signal quality index: below 0.6: 0.05; 0.1 between 0.6 and 0.8: 0.15; above 0.8: 0.15. For limb movement signal quality index: below 0.6: 0.05; 0.1 between 0.6 and 0.8: 0.15; above 0.8: 0.15. The subjective feedback signal quality index is fixed at 0.9, with a weighting of 0.3.

[0104] In this embodiment, the quality index of each modality signal is higher than 0.8, and the weights are allocated as follows: ECG 0.3, skin conductance response 0.2, EEG 0.2, pulse wave 0.15, facial expression 0.15, limb movement 0.15, and subjective feedback 0.3. The sum of the modality weights is 1.45. After normalization, the final weights are: ECG 0.207, skin conductance response 0.138, EEG 0.138, pulse wave 0.103, facial expression 0.103, limb movement 0.103, and subjective feedback 0.207.

[0105] The third layer, temporal context-aware trend correction, employs a temporal convolutional network with a kernel size of 3 and 3 layers. It inputs preliminary modality scores and weights from the past 10 time points and outputs a trend correction coefficient ranging from 0.8 to 1.2. In this embodiment, the patient's pain showed a decreasing trend within one hour post-surgery, with a trend correction coefficient of 0.95.

[0106] The overall pain index (PI) is calculated as follows: PI = Σ (preliminary scores for each modality × normalized weights) × trend correction coefficient. The preliminary scores for each modality are: ECG 7.2, skin conductance 6.8, EEG 7.0, pulse wave 6.5, facial expression 6.6, limb movement 6.4, and subjective feedback 7.5. Therefore, PI = 6.6.

[0107] Step 4: Reinforcement learning adaptively controls input and computation:

[0108] The status input parameters include: comprehensive pain index PI=6.6, cumulative drug dose (cumulative infusion of sufentanil 15μg in the first hour after surgery), physiological stability indicators (heart rate 85 bpm, blood pressure 125 / 80 mmHg, respiratory rate 18 breaths / min, blood oxygen saturation 98%), and physiological indicator variances (heart rate variance 5.2, blood pressure variance 8.3, respiratory rate variance 1.5).

[0109] The reinforcement learning adaptive controller employs a hierarchical structure: the safety layer presets physiological safety thresholds: respiratory rate below 10 breaths / min or above 25 breaths / min, blood oxygen saturation below 92%, systolic blood pressure below 90 mmHg or above 160 mmHg, and diastolic blood pressure below 60 mmHg or above 100 mmHg. In this embodiment, all physiological indicators are within the safety range, and the safety layer does not trigger intervention.

[0110] The optimization layer employs a deep Q-network with an input layer of 6 neurons, 3 hidden layers (32 neurons per layer) using ReLU activation, and an output layer of 3 neurons, corresponding to adjustments for the base infusion rate, pulse dose, and lockout time. The reward function parameters are set as follows: =0.4, =0.3, =0.2, =0.1, the comfort score is collected by the subjective feedback terminal, and in this embodiment it is 6 points. The reward value Rt=-0.49.

[0111] The personalization layer is based on transfer learning. The basic model is trained using historical data from 1,000 patients undergoing postoperative analgesia via laparoscopy. The input data includes the current patient's age, weight, ASA classification, and other individual characteristics. The model transfer adaptation time is 30 seconds, and the output individual adaptation coefficient is 1.1.

[0112] Step 5: Control parameter output:

[0113] Based on the reinforcement learning calculation results and combined with the individual adaptation coefficient, the following control parameters are output: Baseline infusion rate adjustment: The original baseline rate was 2 ml / h, and the adjusted rate is 2.5 ml / h; Pulse dose adjustment: The single pulse dose is 0.5 ml, containing 5 μg of sufentanil; Lockout time adjustment: The original lockout time was 15 minutes, and the adjusted time is 12 minutes.

[0114] Step 6: Security Monitoring Verification

[0115] The pharmacokinetic prediction unit employs a personalized pharmacokinetic model, with an individual patient elimination rate constant ke = 0.12 / h, plasma clearance CL = 1.8 L / h, and the current plasma concentration at time t. =0.3ng / ml, predicting plasma concentration after Δt=5 minutes. . =0.3109 ng / ml. This value is below the safety threshold of 0.5 ng / ml and meets the safety requirements.

[0116] Multimodal consistency check: Calculate the difference between the pain score of each modality and the overall pain index. The maximum difference is 0.7, which is lower than the preset threshold of 0.8, so no review is required.

[0117] The risk warning unit uses a logistic regression model, inputting physiological indicators, drug concentrations, and other data, to predict a respiratory depression risk probability of 2.3%, which is lower than the warning threshold of 5%, so no warning signal is output.

[0118] Step 7 Execution and Online Learning Optimization

[0119] The infusion pump executes the infusion according to the adjusted parameters: a continuous infusion rate of 2.5 ml / h at the base rate, with a single infusion of 0.5 ml when the patient triggers a pulse dose, and a lockout time of 12 minutes. The system records system response data every 5 minutes, including adjusted infusion parameters, newly acquired multi-source signals, comprehensive pain index, physiological indicators, and patient subjective feedback. The online learning module updates the reinforcement learning model parameters every 30 minutes, optimizing the reward function weights and network parameters based on the newly acquired response data, ensuring the control strategy continuously adapts to changes in the patient's pain state.

[0120] Examples of Drug Interaction Compensation Applications

[0121] If a patient uses midazolam (equivalent dose BZD = 2 mg) 4 hours post-surgery due to insomnia, and no other central nervous system medications are used, the drug interaction compensation module will be activated. Compensation factor. =0.417. Based on the compensation factor, the control parameters were adjusted: the basal infusion rate was adjusted to 2.5 × 0.417 ≈ 1.04 ml / h, the pulse dose was adjusted to 0.5 × 0.417 ≈ 0.21 ml, and the lockout time was maintained at 12 minutes. The pharmacokinetic prediction unit recalculated the plasma concentration to ensure that the adjusted predicted concentration remained within the safe range.

[0122] Closed-loop optimization effect:

[0123] Two hours post-surgery, the patient's overall pain index decreased to 4.2, and physiological indicators stabilized: heart rate 78 bpm, respiratory rate 16 breaths / min, and blood oxygen saturation 99%. The reward value Rt = -0.0233. After online system learning, the baseline infusion rate was adjusted to 2.2 ml / h, the pulse dose to 0.4 ml, and the lockout time to 13 minutes, continuously optimizing the balance between analgesic effect and medication dosage.

[0124] An analgesic infusion system includes a processor and a memory. The processor is an industrial-grade quad-core CPU, and the memory capacity is 256GB. The memory stores a computer program, which, when executed by the processor, implements all the steps of the aforementioned adaptive control method for analgesic infusion systems based on multimodal feedback. The system also includes a sensor interface module, an infusion control module, a safety monitoring module, and a wireless communication module. The sensor interface module supports synchronous data acquisition and transmission from multiple sensors; the infusion control module connects to the infusion pump to achieve precise control of the infusion rate, pulse dose, and lockout time; the safety monitoring module operates independently of the control module to ensure the priority of safety interventions; and the wireless communication module supports data interaction with hospital information systems.

[0125] A computer-readable storage medium, a USB flash drive, with a storage capacity of 64GB, stores a computer program. When executed by a processor, the computer program implements all the steps of the aforementioned adaptive control method for analgesic infusion systems based on multimodal feedback. This storage medium supports hot-swapping and can be used for system program updates, data backups, and program portability between multiple devices. It has a data transfer rate of at least 100MB / s, supports long-term data storage, and eliminates the risk of data loss.

[0126] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. An adaptive control method for analgesic infusion systems based on multimodal feedback, characterized in that, Includes the following steps: Step 1: Collect multi-source signals from the patient using multi-source sensors. These multi-source signals include physiological signals, behavioral response signals, and subjective feedback signals. Step 2: Preprocess and extract features from the multi-source signals to obtain feature vectors for each mode; Step 3: Input the feature vector into the multimodal fusion decision layer, and use a confidence-weighted adaptive fusion algorithm to calculate the comprehensive pain index PI; Step 4: The comprehensive pain index (PI), cumulative drug dosage, and physiological stability index are used as state inputs and fed into the reinforcement learning-based adaptive controller. Step 5: The adaptive controller outputs the basic infusion rate adjustment, pulse dose adjustment, and lockout time adjustment. Step 6: The safety monitoring module verifies the safety of the control commands and ensures the safety range through multimodal cross-validation and pharmacokinetic prediction. Step 7: Execute the adjusted infusion parameters and record the system response data for online learning and optimization.

2. The adaptive control method for analgesic infusion system based on multimodal feedback according to claim 1, characterized in that, In step 1, the multi-source sensors include an electrocardiogram sensor, a skin conductance sensor, a forehead electroencephalogram sensor, a photoplethysmography (PPG) sensor, a miniature camera, a pressure-sensitive mattress, and a wireless feedback terminal.

3. The adaptive control method for analgesic infusion system based on multimodal feedback according to claim 2, characterized in that, The multimodal fusion decision layer adopts a layered fusion architecture, including: First layer: Feature extraction and preliminary scoring within each modality; The second layer: dynamic weight allocation based on the signal quality index; The third layer: Temporal context-aware trend correction, which uses a temporal convolutional network to analyze the evolution of pain states.

4. The adaptive control method for analgesic infusion system based on multimodal feedback according to claim 3, characterized in that, The reinforcement learning-based adaptive controller adopts a hierarchical structure, including a security layer, an optimization layer, and a personalization layer; The safety layer is based on safety interventions using hard physiological thresholds; The optimization layer is based on dose optimization control using a deep Q-network; The personalization layer is based on an individual feature adaptation module using transfer learning.

5. The adaptive control method for analgesic infusion system based on multimodal feedback according to claim 4, characterized in that, In step 4, the reward function for the adaptive controller reinforcement learning is: ; in, It is the reward value at time t. It is the comprehensive pain index at time t. The drug increment at time t Let be the variance of physiological indicators at time t. Rate the comfort level at time t. , , and It is the weighting coefficient.

6. The adaptive control method for analgesic infusion system based on multimodal feedback according to claim 5, characterized in that, The safety monitoring module includes a pharmacokinetic prediction unit, a multimodal consistency check, and a risk warning unit. The pharmacokinetic prediction unit predicts plasma concentration based on a personalized pharmacokinetic model; the multimodal consistency check initiates a review when the difference in pain assessment between different modalities exceeds a threshold; and the risk warning unit predicts the risk of respiratory depression based on a logistic regression model.

7. The adaptive control method for analgesic infusion system based on multimodal feedback according to claim 6, characterized in that, The formula for predicting plasma concentration using a personalized pharmacokinetic model is: ; in It is after time t plasma concentration at time, Let be the plasma concentration at time t. To eliminate the rate constant, For infusion rate, This represents the plasma clearance rate of analgesic drugs.

8. The adaptive control method for analgesic infusion system based on multimodal feedback according to claim 7, characterized in that, It also includes a drug interaction compensation module, which automatically adjusts the dosage algorithm when it detects that the patient is using other central nervous system drugs, and calculates the compensation factor using the following formula. : ； Where [BZD] represents the equivalent dose of benzodiazepines, and [AH] represents the equivalent dose of antihistamines. This is equivalent to the dose of other central nervous system depressants.

9. An analgesic infusion system, characterized in that, It includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the steps of the adaptive control method for an analgesic infusion system based on multimodal feedback as described in claim 8.

10. A computer-readable storage medium, characterized in that, The system contains a computer program that, when executed by a processor, implements the steps of the adaptive control method for analgesic infusion systems based on multimodal feedback as described in claim 8.

Images