A painless gastrointestinal endoscopy adaptive dosing method and system based on multi-source physiological signal fusion

Abstract

The present application relates to a painless gastrointestinal endoscopy adaptive dosing method and system based on multi-source physiological signal fusion, belonging to the field of medical artificial intelligence and automatic control technology, through real-time acquisition and fusion of patient's electroencephalogram, electrocardiogram, skin electricity and blood oxygen saturation and other multi-modal physiological signals, using artificial intelligence model to deeply extract features and quantitatively evaluate the sedation- analgesia level of the patient, relying on the pharmacokinetics- pharmacodynamics model constructed by the big data platform, dynamically predicting the drug effect according to the individual characteristics and real-time state of the patient, and automatically and accurately adjusting the infusion rate of anesthetic drugs through a closed-loop control algorithm, realizing personalized and adaptive dosing, setting up an independent safety monitoring mechanism, and real-time intervention on risk events such as respiratory depression. The present application can effectively solve the problems of traditional anesthesia relying on artificial experience, single parameter monitoring being not comprehensive and lacking individualized dynamic adjustment, and significantly improve the safety, comfort and medical efficiency of painless gastrointestinal endoscopy.

Description

Technical Field

[0001] This application relates to the fields of medical artificial intelligence and automatic control technology, and in particular to a painless gastrointestinal endoscopy adaptive drug delivery method and system based on multi-source physiological signal fusion. Background Technology

[0002] Currently, anesthetic administration during painless gastroscopy and colonoscopy relies primarily on the anesthesiologist's experience, achieved through manual injection or adjustment of the infusion pump rate. This method typically only considers limited vital sign parameters, such as heart rate, blood pressure, and blood oxygen saturation, lacking precise and quantitative assessment of the patient's analgesia and sedation depth. This administration mode, dependent on the physician's subjective judgment, is prone to under- or over-administration due to individual differences and varying operator experience, leading to risks such as intraoperative movement, circulatory depression, or respiratory arrest. Furthermore, it makes standardization and normalization of the anesthesia process difficult.

[0003] In recent years, although some studies have attempted to use closed-loop drug delivery systems based on single EEG parameters such as the bispectral index, these systems rely on only a single physiological signal and cannot comprehensively reflect the balance between noxious stimuli and analgesic levels. During gastrointestinal endoscopy, procedures such as endoscope bends, traction, or inflation can trigger strong noxious stimuli. The response of a single EEG parameter to such dynamic changes is delayed and limited, potentially leading to intraoperative awareness or incomplete analgesia. Furthermore, existing systems lack consideration for individual patient pharmacokinetic characteristics and cannot adaptively adjust to factors such as age, weight, and comorbidities, thus limiting their generalization ability and clinical applicability.

[0004] Furthermore, existing automated drug delivery systems are typically closed-loop architectures with rigid control algorithms that cannot learn and evolve from continuously accumulated clinical data. The multi-dimensional physiological signals and drug delivery records generated during anesthesia are not systematically integrated and analyzed, making it difficult for algorithms to cope with complex and ever-changing clinical scenarios and specific patient groups. Therefore, developing an intelligent adaptive drug delivery system that can integrate multi-source physiological signals, incorporate big data-driven modeling, and possess continuous optimization capabilities has become a key breakthrough direction for improving the safety and comfort of painless gastroscopy and colonoscopy. Summary of the Invention

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] According to a first aspect of the present invention, the present invention claims protection for a painless gastrointestinal endoscopy adaptive drug delivery method based on multi-source physiological signal fusion, running on a computer system communicatively connected to a drug infusion pump, characterized by comprising the following steps:

[0007] S1, real-time acquisition of physiological signals from patients undergoing gastrointestinal endoscopy through multiple sensors, including at least electroencephalogram (EEG) signals, electrocardiogram (ECG) signals, skin conductance signals, and pulse oximetry signals; noise reduction, filtering, and motion artifact removal preprocessing of the acquired raw signals; and unified sampling frequency and timestamp to generate a synchronized multimodal signal stream.

[0008] S2 extracts time-domain, frequency-domain, and nonlinear features characterizing the patient's pain stress and sedation depth from the preprocessed multimodal signal stream and inputs them into a pre-trained sedation-analgesia depth assessment model, outputting a comprehensive quantitative analgesia depth index for real-time assessment of the patient's comfort status.

[0009] S3 calls upon a pharmacokinetic-pharmacodynamic model trained on massive historical anesthesia data. Using the patient's basic information, current analgesia depth index, and history of drug administration as input, it predicts the changes in drug concentration in the patient's body and the corresponding expected analgesia depth within a preset time period in the future, and calculates the optimal dosing rate required to achieve and maintain the target analgesia depth range.

[0010] S4 sends the optimal dosing rate command calculated in S3 to the drug infusion pump to achieve real-time and precise control of the anesthetic drug, continuously monitor the physiological signals for abnormal fluctuations, and immediately trigger the safety intervention protocol if a predefined risk event such as respiratory depression or bradycardia is detected, automatically suspending the dosing and issuing an alarm.

[0011] S5 archives all multimodal physiological signals, drug administration records, model decision logs, and postoperative recovery scores from the entire examination process to the big data platform, and periodically uses the accumulated dataset to incrementally learn and optimize the evaluation model in S2 and the PK-PD model in S3.

[0012] Furthermore, step S1 also includes:

[0013] The sensors include a forehead-mounted EEG monitor, ECG leads, a finger-clip pulse oximeter, and a wristband-type skin conductance sensor; the preprocessing includes using an independent component analysis algorithm to separate electrooculography artifacts in the EEG signal and using an adaptive filter to eliminate power frequency interference in the ECG signal.

[0014] Furthermore, step S2 also includes:

[0015] Deep features include, but are not limited to: spectral entropy, burst suppression ratio, and relative power of delta waves in EEG signals; time-frequency indices of heart rate variability in ECG signals; response frequency and amplitude of skin conductance signals; and the decreasing trend of blood oxygen saturation.

[0016] The sedation-analgesia depth assessment model is a gradient boosting decision tree ensemble learning model.

[0017] Furthermore, step S3 also includes:

[0018] The patient's basic personal information includes age, weight, height, and history of underlying medical conditions;

[0019] The PK-PD model is a three-compartment model constructed based on population pharmacokinetic principles and using a nonlinear mixed-effects modeling method. Its parameters are learned from historical big data through the expectation-maximization algorithm.

[0020] Furthermore, step S3 also includes:

[0021] Dynamic prediction is performed on a rolling basis at fixed time periods. Each prediction references the latest physiological signal characteristics and dosing data. The predicted drug concentration-time curve is integrated, and the optimization objective is to maintain the analgesia depth index within the target range. The real-time dosing rate adjustment is calculated through a proportional-integral-derivative controller.

[0022] Furthermore, step S4 also includes:

[0023] The safety intervention protocol includes automatically reducing the dosing rate when blood oxygen saturation remains below a preset threshold;

[0024] When the brain burst suppression ratio is detected to exceed the safe limit, it is determined that the sedation is too deep, and the drug administration is immediately suspended and the anesthesiologist is notified.

[0025] Furthermore, step S5 also includes:

[0026] The big data platform uses a distributed columnar database to store high-concurrency time-series physiological signal data and utilizes Spark Streaming technology to perform real-time feature calculations on the inbound data, providing data preparation for the immediate optimization of the model.

[0027] Furthermore, step S5 also includes:

[0028] During model iteration and optimization, newly added anonymized case data is extracted from the big data platform at fixed intervals to fine-tune the existing assessment model and PK-PD model, continuously absorbing new clinical patterns and achieving continuous improvement in the model's self-evolution and personalized adaptability.

[0029] Furthermore, the method also includes:

[0030] S6. At the end of the examination, based on the total amount of medication administered throughout the process, fluctuations in physiological indicators, and individual patient characteristics, the system uses recovery data from similar cases in the big data platform to predict the patient's awakening time and the probability of postoperative nausea and vomiting, and automatically generates an anesthetic effect and safety assessment report.

[0031] The method operates on an architecture that combines edge computing nodes and a cloud-based big data platform. The real-time signal processing and drug delivery control tasks in steps S1 to S4 are completed at the edge, while the data archiving and model optimization tasks in step S5 are completed in the cloud to leverage powerful distributed computing capabilities.

[0032] According to a second aspect of the present invention, the present invention claims protection for a painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion, comprising:

[0033] One or more processors;

[0034] A memory storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the aforementioned painless gastrointestinal endoscopy adaptive drug delivery method based on multi-source physiological signal fusion.

[0035] This invention relates to a painless gastrointestinal endoscopy adaptive drug delivery method and system based on multi-source physiological signal fusion, belonging to the field of medical artificial intelligence and automatic control technology. It involves real-time acquisition and fusion of multimodal physiological signals from the patient, such as electroencephalogram (EEG), electrocardiogram (ECG), skin conductance, and blood oxygen saturation. Using an artificial intelligence model, it deeply extracts features and quantitatively assesses the patient's sedation-analgesia level. Based on a pharmacokinetic-pharmacodynamic model constructed using a big data platform, it dynamically predicts drug effects according to the patient's individual characteristics and real-time status. A closed-loop control algorithm automatically and precisely adjusts the infusion rate of anesthetic drugs, achieving personalized and adaptive drug delivery. An independent safety monitoring mechanism is established to intervene in real-time for risk events such as respiratory depression. This invention effectively solves the problems of traditional anesthesia relying on human experience, incomplete monitoring of single parameters, and lack of individualized dynamic adjustment, significantly improving the safety, comfort, and medical efficiency of painless gastrointestinal endoscopy. Attached Figure Description

[0036] Figure 1 A flowchart illustrating the workflow of a painless gastrointestinal endoscopy adaptive drug delivery method based on multi-source physiological signal fusion, as claimed in this application embodiment.

[0037] Figure 2 The diagram shows the structural modules of a painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion, as claimed in the embodiments of this application. Detailed Implementation

[0038] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0039] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0040] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0041] According to a first embodiment of the present invention, the present invention claims protection for a painless gastrointestinal endoscopy adaptive drug delivery method based on multi-source physiological signal fusion, which runs on a computer system communicatively connected to a drug infusion pump, characterized in that, referring to Figure 1 This includes the following steps:

[0042] S1, real-time acquisition of physiological signals from patients undergoing gastrointestinal endoscopy through multiple sensors, including at least electroencephalogram (EEG) signals, electrocardiogram (ECG) signals, skin conductance signals, and pulse oximetry signals; noise reduction, filtering, and motion artifact removal preprocessing of the acquired raw signals; and unified sampling frequency and timestamp to generate a synchronized multimodal signal stream.

[0043] S2 extracts time-domain, frequency-domain, and nonlinear features characterizing the patient's pain stress and sedation depth from the preprocessed multimodal signal stream and inputs them into a pre-trained sedation-analgesia depth assessment model, outputting a comprehensive quantitative analgesia depth index for real-time assessment of the patient's comfort status.

[0044] S3 calls upon a pharmacokinetic-pharmacodynamic model trained on massive historical anesthesia data. Using the patient's basic information, current analgesia depth index, and history of drug administration as input, it predicts the changes in drug concentration in the patient's body and the corresponding expected analgesia depth within a preset time period in the future, and calculates the optimal dosing rate required to achieve and maintain the target analgesia depth range.

[0045] S4 sends the optimal dosing rate command calculated in S3 to the drug infusion pump to achieve real-time and precise control of the anesthetic drug, continuously monitor the physiological signals for abnormal fluctuations, and immediately trigger the safety intervention protocol if a predefined risk event such as respiratory depression or bradycardia is detected, automatically suspending the dosing and issuing an alarm.

[0046] S5 archives all multimodal physiological signals, drug administration records, model decision logs, and postoperative recovery scores from the entire examination process to the big data platform, and periodically uses the accumulated dataset to incrementally learn and optimize the evaluation model in S2 and the PK-PD model in S3.

[0047] In this embodiment, S1, real-time acquisition and preprocessing of multimodal physiological signals: The system acquires the patient's electroencephalogram (EEG) signals via a forehead-mounted sensor, electrocardiogram (ECG) signals via chest leads, pulse oximetry signals via a finger-clip photoelectric sensor, and skin conductance signals via a wrist sensor. The system performs preliminary processing on the acquired raw signals, including eliminating fixed-frequency interference from the power supply, filtering out signal abrupt noise caused by slight patient movements, and aligning the timing of all signals to form a synchronized multimodal physiological data stream.

[0048] S2. Deep Feature Extraction and Analgesia Depth Assessment: The system automatically calculates a series of feature indicators from the synchronous data stream, including waveform disorder indicators analyzed from EEG signals, autonomic nervous system activity indicators analyzed from heart rate fluctuations, the frequency and intensity of instantaneous stress responses analyzed from electrodermal signals, and the stability of blood oxygen saturation. These feature indicators are combined into a comprehensive feature vector, which is input into a machine learning model pre-trained with a large amount of clinical data. The model analyzes this feature vector and outputs a single quantitative value between 0 and 100, which comprehensively represents the patient's current state from awake to sedation to deep sedation; the lower the value, the more sedated the patient.

[0049] S3. Dynamic Prediction Based on Big Data-Driven Pharmacokinetic Model: The system receives the real-time status values ​​generated in step S2 and combines them with the patient's basic information (such as age and weight) and accurate dosing records to date. Internally, the system calls upon a mathematical model built on massive amounts of population anesthesia data. This model simulates the distribution, metabolism, and efficacy of the drug in the human body. Using this model, the system predicts how the patient's status values ​​will change in the next few minutes at the current dosing rate. Then, with the goal of stabilizing the patient's status values ​​within a preset ideal range (e.g., 60-80), the system reverse-calculates the optimal dosing rate adjustment recommendation.

[0050] S4. Closed-Loop Adaptive Dosing Control and Safety Monitoring: The system sends the optimal dosing rate command calculated in step S3 to the intelligent infusion pump in real time, thereby achieving automatic and precise adjustment of the infusion rate of anesthetic drugs (such as propofol). Simultaneously, the system establishes an independent safety monitoring thread to continuously scan physiological signals for life-threatening abnormal patterns, such as a sustained rapid decline in blood oxygen levels or severe bradycardia. Once such a predefined critical event is detected, the system will immediately override conventional control logic, forcibly command the infusion pump to pause dosing, and activate audible and visual alarms to notify the anesthesiologist for intervention.

[0051] S5. Data Archiving and Model Iteration Optimization: All data generated during this gastroscopy and colonoscopy, including physiological signals every second, every dosing instruction, all model prediction results, and any manual intervention records from the physician, are completely and anonymized before being permanently stored on a cloud-based big data platform. The platform regularly (e.g., weekly) uses the newly added data to retrain and fine-tune the state assessment model in step S2 and the pharmacokinetic model in step S3, enabling the models to continuously learn new patterns and thus provide more accurate personalized dosing recommendations for a more diverse patient population in the future.

[0052] Furthermore, step S1 also includes:

[0053] The sensors include a forehead-mounted EEG monitor, ECG leads, a finger-clip pulse oximeter, and a wristband-type skin conductance sensor; the preprocessing includes using an independent component analysis algorithm to separate electrooculography artifacts in the EEG signal and using an adaptive filter to eliminate power frequency interference in the ECG signal.

[0054] In this embodiment, in step S1, eliminating power interference involves filtering out 50Hz or 60Hz power frequency interference components using a digital filter; filtering out motion artifacts involves identifying and removing sudden high-amplitude fluctuations in the signal that are unrelated to the electrocardiogram or respiratory rhythm using an algorithm.

[0055] Furthermore, step S2 also includes:

[0056] Deep features include, but are not limited to: spectral entropy, burst suppression ratio, and relative power of delta waves in EEG signals; time-frequency indices of heart rate variability in ECG signals; response frequency and amplitude of skin conductance signals; and the decreasing trend of blood oxygen saturation.

[0057] The sedation-analgesia depth assessment model is a gradient boosting decision tree ensemble learning model.

[0058] In this embodiment, in step S2, the waveform disorder index is obtained by calculating the entropy value of its spectrum after mathematical transformation of the EEG signal; the autonomic nervous system activity index is obtained by analyzing the time series volatility of continuous heartbeat intervals.

[0059] Furthermore, step S3 also includes:

[0060] The patient's basic personal information includes age, weight, height, and history of underlying medical conditions;

[0061] The PK-PD model is a three-compartment model constructed based on population pharmacokinetic principles and using a nonlinear mixed-effects modeling method. Its parameters are learned from historical big data through the expectation-maximization algorithm.

[0062] In this embodiment, in step S3, the pharmacokinetic model is a mathematical model that considers the effects of age, weight, and liver function on the rate of drug metabolism. It can simulate the entire process of a drug from injection into the bloodstream, distribution into body tissues, and breakdown by the liver, and predict its concentration changes at the site of action in the brain.

[0063] Furthermore, step S3 also includes:

[0064] Dynamic prediction is performed on a rolling basis at fixed time periods. Each prediction references the latest physiological signal characteristics and dosing data. The predicted drug concentration-time curve is integrated, and the optimization objective is to maintain the analgesia depth index within the target range. The real-time dosing rate adjustment is calculated through a proportional-integral-derivative controller.

[0065] In this embodiment, the reverse calculation process in step S3 is as follows: the model first predicts the future state trajectory at the current rate. If the predicted trajectory deviates from the target range, the system will simulate the new trajectory after increasing or decreasing the drug administration rate. Through multiple iterative simulations, an optimal rate value that can stably maintain the state value within the target range is found.

[0066] Furthermore, step S4 also includes:

[0067] The safety intervention protocol includes automatically reducing the dosing rate when blood oxygen saturation remains below a preset threshold;

[0068] When the brain burst suppression ratio is detected to exceed the safe limit, it is determined that the sedation is too deep, and the drug administration is immediately suspended and the anesthesiologist is notified.

[0069] In this embodiment, in step S4, the independent safety monitoring thread runs in parallel with the highest priority. Once the abnormal detection threshold it sets (such as blood oxygen saturation below 90% for 10 seconds) is triggered, any ongoing drug administration instructions will be immediately interrupted.

[0070] Furthermore, step S5 also includes:

[0071] The big data platform uses a distributed columnar database to store high-concurrency time-series physiological signal data and utilizes Spark Streaming technology to perform real-time feature calculations on the inbound data, providing data preparation for the immediate optimization of the model.

[0072] In this embodiment, in step S5, the big data platform uses a time-series database to store massive amounts of time-stamped physiological signal data and utilizes a distributed computing framework to efficiently process this data in preparation for model training.

[0073] Furthermore, step S5 also includes:

[0074] During model iteration and optimization, newly added anonymized case data is extracted from the big data platform at fixed intervals to fine-tune the existing assessment model and PK-PD model, continuously absorbing new clinical patterns and achieving continuous improvement in the model's self-evolution and personalized adaptability.

[0075] In this embodiment, step S5, retraining and fine-tuning, refers to: retaining the core structure of the original model and making minor adjustments to some parameters using only new clinical data, so that the model can better adapt to the features reflected in the new data while maintaining the original knowledge. This is a continuous optimization cycle.

[0076] Furthermore, the method also includes:

[0077] S6. At the end of the examination, based on the total amount of medication administered throughout the process, fluctuations in physiological indicators, and individual patient characteristics, the system uses recovery data from similar cases in the big data platform to predict the patient's awakening time and the probability of postoperative nausea and vomiting, and automatically generates an anesthetic effect and safety assessment report.

[0078] The method operates on an architecture that combines edge computing nodes and a cloud-based big data platform. The real-time signal processing and drug delivery control tasks in steps S1 to S4 are completed at the edge, while the data archiving and model optimization tasks in step S5 are completed in the cloud to leverage powerful distributed computing capabilities.

[0079] In this embodiment, S6, Personalized Recovery Prediction and Report Generation: At the end of the surgery, the system comprehensively considers the total amount of medication used throughout the process, the fluctuation curve of physiological indicators, and the individual characteristics of the patient. It compares these with the postoperative recovery records of a large number of similar cases in the database to predict the time required for the patient to fully wake up and the probability of postoperative complications such as nausea and vomiting. The system then automatically generates an anesthesia summary report containing the above information for medical staff to refer to.

[0080] The computational tasks of the method are distributed between the edge computing device and the cloud platform: the local computer (edge ​​device) in the inspection room is responsible for executing all signal processing and closed-loop control tasks with extremely high real-time requirements in steps S1 to S4 to ensure low latency; while the data archiving and model optimization tasks in step S5 are sent to the cloud server with powerful storage and computing capabilities for asynchronous execution.

[0081] According to a second embodiment of the present invention, the present invention claims protection for a painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion, referring to... Figure 2 ,include:

[0082] One or more processors;

[0083] A memory storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the aforementioned painless gastrointestinal endoscopy adaptive drug delivery method based on multi-source physiological signal fusion.

[0084] This embodiment aims to describe in detail the practical application of a painless gastrointestinal endoscopy adaptive drug delivery method based on multi-source physiological signal fusion. By integrating multimodal physiological monitoring data, artificial intelligence-based deep state assessment, big data-driven pharmacokinetic prediction, and closed-loop automatic control, safe and precise individualized anesthesia management is achieved. This embodiment will gradually demonstrate the specific implementation of the method and provide clinical trial test data to verify its effectiveness and safety.

[0085] Subjects and Design: 120 patients undergoing painless gastroscopy and colonoscopy were randomly divided into two groups: the experimental group (n=60) received anesthesia management using the adaptive drug delivery system, while the control group (n=60) received medication manually controlled by an experienced anesthesiologist based on clinical routines and subjective judgment. There were no statistically significant differences between the two groups in baseline characteristics such as age, weight, and ASA classification.

[0086] Physiological signal monitoring: SedLine brain function monitor (acquires 4-channel EEG signals), Philips IntelliVue MP50 monitor (acquires ECG and pulse oximetry), Empatica E4 wristband (acquires skin conductance activity).

[0087] Drug delivery device: Alaris GH Syringe Pump intelligent infusion pump, which communicates with the main control computer via RS-485 serial protocol.

[0088] Main control computer: A workstation equipped with an Intel Core i7-11800H processor, responsible for real-time data processing and algorithm decision-making.

[0089] Signal processing: Filtering and artifact removal using the FDA toolbox based on MATLAB.

[0090] Machine learning model: The sedation-analgesia depth assessment model is a gradient boosting decision tree model trained using the XGBoost framework, with 5000 historical anesthesia data.

[0091] Pharmacokinetic model: A three-compartment PK-PD model of propofol was constructed based on NONMEM software, and the parameters were derived from published population pharmacokinetic studies.

[0092] Big Data Platform: Built on Hadoop 3.3.4 and HBase 2.4.9, used for data storage and model iteration.

[0093] Step S1: Real-time acquisition and preprocessing of multimodal physiological signals;

[0094] Implementation Procedure: All monitoring devices were connected to each patient in the experimental group. SedLine electrodes were attached to the forehead at a sampling rate of 128Hz; ECG was performed using lead II at a sampling rate of 500Hz; the pulse oximeter was clipped to the index finger; and an E4 wristband was worn on the wrist on the non-infusion side. After receiving the raw signals in real time, the system first applied a 50Hz notch filter to eliminate power frequency interference. Then, a wavelet transform-based algorithm was used to identify and remove motion artifacts caused by patient coughing and body movement. Finally, the system resampled all signals at a uniform frequency of 100Hz and aligned the timestamps to generate a synchronized multimodal signal stream.

[0095] Test results: After signal preprocessing, the motion artifact removal success rate reached 92.3%, and the multimodal signal synchronization error was less than 10 milliseconds.

[0096] Step S2: Deep feature extraction and analgesia depth status assessment;

[0097] Implementation process: The system calculates the following features from the synchronous data stream in 5-second time windows: 1) Electroencephalogram (EEG) signals: spectral entropy (SE), 95% spectral edge frequency (SEF95), and delta wave relative power (1-4Hz); 2) Electrocardiogram (ECG): heart rate (HR), heart rate variability (RMSSD); 3) Electrodermal conductance (EDC) signals: peak frequency and average amplitude of skin conductance response (SCR); 4) Blood oxygen saturation (… The features and their trends are combined into a feature vector and input into a pre-trained XGBoost model. The model outputs a quantitative value called the "Patient State Index" (PSI), ranging from 0 to 100 (0: burst suppression, 100: full consciousness).

[0098] Test results: The XGBoost model achieved a correlation coefficient of 0.91 between the predicted PSI value and the expert-assessed sedation depth (based on OAA / S score) on the test set. The average inference time of the model was 15 milliseconds.

[0099] Step S3: Dynamic prediction based on big data-driven pharmacokinetic model;

[0100] Implementation Process: The system executes a prediction cycle every 10 seconds. It takes the current PSI value, patient weight (kg), age (years), and propofol dosing history within the past 20 minutes as input, and calls upon a propofol three-compartment PK-PD model. This model predicts the PSI trajectory over the next 5 minutes while maintaining the current dosing rate. The system's core optimization objective is to maintain the PSI within the target range (65-75). By simulating the predicted trajectory under different dosing rates (e.g., increasing by 0.5 mg / kg / h or decreasing by 0.3 mg / kg / h), the system uses a PID control algorithm to calculate the optimal dosing rate adjustment and converts it into control commands for the infusion pump.

[0101] Test results: The mean absolute error between the PSI trajectory predicted by the PK-PD model and the actual PSI value was 3.2. The experimental group's PSI value spent 85.7% of its time within the target interval, which was significantly higher than the control group's 63.4% (p<0.01).

[0102] Step S4: Closed-loop adaptive drug delivery control and safety monitoring;

[0103] Implementation process: The main control computer sends the calculated dosing rate command to the Alaris infusion pump via serial port. Simultaneously, a separate high-priority monitoring thread runs continuously, with its set safety thresholds including: <90% duration for 10 seconds, HR <45 beats / min, or burst suppression ratio (BSR) >40%. Once any indicator triggers the threshold, the system immediately performs a hard interruption, stops drug administration, and activates the bedside alarm and flashes a warning on the screen.

[0104] Test results: In the experimental group, the system successfully maintained stable drug administration control. Three hypoxemia safety interventions were triggered, all of which successfully prevented serious adverse events, and no false positive alarms were detected.

[0105] Step S5: Data archiving and model iterative optimization;

[0106] Implementation Process: After the inspection, all time-series data (raw signals, features, PSI values, dosing instructions, alarm events) were encrypted and uploaded to the Hadoop big data platform, stored in Parquet columnar format. Each week, a Spark job extracts new data, incrementally trains the XGBoost model (learning rate 0.01, 100 iterations), and fine-tunes the inter-individual variation parameters of the PK-PD model using the Bayesian method.

[0107] Test results: After 4 weeks of data accumulation and iterative optimization, the correlation coefficient of the PSI prediction model increased to 0.93, and the prediction error of the PK-PD model decreased to 2.9.

[0108] Signal preprocessing details: To address power frequency interference in the ECG signal, an adaptive notch filter was used, with its center frequency automatically tracking the mains frequency (50 or 60 Hz). For motion artifacts, a soft thresholding denoising algorithm based on Discrete Wavelet Transform (DWT) was employed, effectively distinguishing high-amplitude abrupt changes in the ECG R wave and motion artifacts.

[0109] Test results: The signal-to-noise ratio (SNR) of the ECG signal improved from the original 15.2 dB to 28.5 dB.

[0110] Feature extraction process: Spectral entropy (SE) is calculated by performing a Fast Fourier Transform (FFT) on 5-second EEG window data to obtain the power spectrum, and then calculating its normalized Shannon entropy. Heart rate variability (RMSSD) is calculated by detecting the peak value of the R wave and calculating the root mean square of the difference between consecutive normal sinus beats (NN interval).

[0111] Test results: SE was negatively correlated with sedation depth (r = -0.85), and RMSSD was significantly reduced under stress.

[0112] Detailed implementation process of the PK-PD model: The propofol three-compartment model used has established mathematical relationships between parameters such as the central compartment distribution volume (V1) and clearance rate (CL) and age and weight (e.g., CL decreases linearly with age). The system calculates the personalized model parameters in real time based on the patient's individual information.

[0113] Test results: The prediction accuracy of the model with personalized parameters is 28% higher than that of the model with fixed parameters.

[0114] Dynamic prediction and control logic implementation process: The PID controller takes the deviation between PSI and the target median (70) as input, and its proportional, integral, and derivative coefficients are tuned through prior simulation. The controller outputs an increment (or decrement) of the dosing rate, which is then superimposed with the current rate to form a new setpoint sent to the pump.

[0115] Test results: PID control resulted in PSI overshoot of less than 5%, and the average system settling time (from the start of induction to the first entry of PSI into the target range) was 145 seconds.

[0116] Safety monitoring protocol implementation process: The safety thread operates independently of the main control loop, scanning and monitoring data at 100ms intervals. For The judgment was made by calculating the average value of a sliding window lasting 10 seconds to avoid false alarms due to instantaneous fluctuations.

[0117] Test results: The response latency of the safety monitoring thread was less than 50ms, and all three real hypoxia events were correctly captured and processed.

[0118] Big data platform architecture implementation process: The Hadoop cluster consists of 1 master node and 3 data nodes. HBase is used to store unstructured waveform data fragments, and Parquet is used to store structured feature data. Spark jobs read data from HBase and perform feature calculations.

[0119] Test results: After compression, the data of a single patient occupies approximately 15MB of storage space. The cluster can support the parallel storage and processing of thousands of surgical data per day.

[0120] Model iteration and optimization process: During incremental training, newly added data from the past month is first extracted from HBase and mixed with the original training set. When training a new model, most of the tree structure of the original model is retained, and only a small number of trees are added to learn patterns in the new data, avoiding catastrophic forgetting.

[0121] Test results: Incremental training takes about 20% of the time of full training, and the model performance is significantly improved.

[0122] Personalized recovery prediction implementation process: At the end of the examination, the system extracts the total propofol dosage, PSI fluctuation curve, patient age and weight as input features. By querying the big data platform, it finds the 100 most similar historical cases, calculates their average awakening time and PONV incidence rate, and outputs them as predicted values.

[0123] Test results: The average deviation between the predicted awakening time and the actual awakening time was ±2.1 minutes, and the PONV prediction accuracy reached 82%.

[0124] Edge-cloud collaborative architecture implementation process: In the inspection room, all real-time tasks (S1-S4) are deployed on local workstations (edge). Data upload and model optimization (S5) tasks are encapsulated as jobs and sent to the cloud big data platform for queued execution via the network. Edge devices only interact with the cloud in large amounts of data after get off work when the network is idle or when they receive updated models from the cloud.

[0125] Test results: The edge control loop latency is stable within 150ms, and the cloud model is updated weekly and pushed to each edge device via the network.

[0126] The main outcome indicators of the comprehensive test results in this embodiment include:

[0127] Total propofol dosage: experimental group (125.3 ± 35.6 mg) vs. control group (142.8 ± 40.1 mg), p<0.05, saving 12.2%.

[0128] Percentage of time spent in the target interval during the examination: experimental group (85.7%) vs. control group (63.4%), p<0.01.

[0129] hypoxemia <90%) Event: Experimental group (3 times) vs. control group (11 times), p<0.05.

[0130] Awakening time: experimental group (8.5 ± 2.3 min) vs. control group (10.2 ± 3.1 min), p<0.05.

[0131] Doctor and patient satisfaction: Anesthesiologists in the experimental group reported a significant reduction in workload, and patients had higher satisfaction scores.

[0132] This embodiment fully verifies the effectiveness, safety, and superiority of the adaptive drug delivery method. Through multimodal physiological signal fusion, AI real-time status assessment, big data-driven accurate prediction, and closed-loop control, the drug delivery process is automated and individualized, significantly improving the quality and safety of painless gastrointestinal endoscopy anesthesia, and possessing important clinical value and promotional significance.

[0133] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0134] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0135] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion, characterized in that, include: One or more processors; A memory having stored one or more programs, which, when executed by one or more processors, cause the one or more processors to perform the following steps: S1, real-time acquisition of physiological signals from patients undergoing gastrointestinal endoscopy through multiple sensors, including at least electroencephalogram (EEG) signals, electrocardiogram (ECG) signals, skin conductance signals, and pulse oximetry signals; noise reduction, filtering, and motion artifact removal preprocessing of the acquired raw signals; and unified sampling frequency and timestamp to generate a synchronized multimodal signal stream. S2 extracts time-domain, frequency-domain, and nonlinear features characterizing the patient's pain stress and sedation depth from the preprocessed multimodal signal stream and inputs them into a pre-trained sedation-analgesia depth assessment model, outputting a comprehensive quantitative analgesia depth index for real-time assessment of the patient's comfort status. S3 calls upon a pharmacokinetic-pharmacodynamic model trained on massive historical anesthesia data. Using the patient's basic information, current analgesia depth index, and history of drug administration as input, it predicts the changes in drug concentration in the patient's body and the corresponding expected analgesia depth within a preset time period in the future, and calculates the optimal dosing rate required to achieve and maintain the target analgesia depth range. S4 sends the optimal dosing rate command calculated in S3 to the drug infusion pump to achieve real-time and precise control of the anesthetic drug, continuously monitor the physiological signals for abnormal fluctuations, and immediately trigger the safety intervention protocol if a predefined risk event such as respiratory depression or bradycardia is detected, automatically suspending the dosing and issuing an alarm. S5 archives all multimodal physiological signals, drug administration records, model decision logs, and postoperative recovery scores from the entire examination process to the big data platform, and periodically uses the accumulated dataset to incrementally learn and optimize the evaluation model in S2 and the PK-PD model in S3.

2. The painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion according to claim 1, characterized in that, Step S1 also includes: The sensors include a forehead-mounted EEG monitor, ECG leads, a finger-clip pulse oximeter, and a wristband-type skin conductance sensor; the preprocessing includes using an independent component analysis algorithm to separate electrooculography artifacts in the EEG signal and using an adaptive filter to eliminate power frequency interference in the ECG signal.

3. The painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion according to claim 1, characterized in that, Step S2 also includes: Deep features include, but are not limited to: spectral entropy, burst suppression ratio, and relative power of delta waves in EEG signals; time-frequency indices of heart rate variability in ECG signals; response frequency and amplitude of skin conductance signals; and the decreasing trend of blood oxygen saturation. The sedation-analgesia depth assessment model is a gradient boosting decision tree ensemble learning model.

4. The painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion according to claim 1, characterized in that, Step S3 also includes: The patient's basic personal information includes age, weight, height, and history of underlying medical conditions; The PK-PD model is a three-compartment model constructed based on population pharmacokinetic principles and using a nonlinear mixed-effects modeling method. Its parameters are learned from historical big data through the expectation-maximization algorithm.

5. The painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion according to claim 1, characterized in that, Step S3 also includes: Dynamic prediction is performed on a rolling basis at fixed time periods. Each prediction references the latest physiological signal characteristics and dosing data. The predicted drug concentration-time curve is integrated, and the optimization objective is to maintain the analgesia depth index within the target range. The real-time dosing rate adjustment is calculated through a proportional-integral-derivative controller.

6. The painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion according to claim 1, characterized in that, Step S4 also includes: The safety intervention protocol includes automatically reducing the dosing rate when blood oxygen saturation remains below a preset threshold; When the brain burst suppression ratio is detected to exceed the safe limit, it is determined that the sedation is too deep, and the drug administration is immediately suspended and the anesthesiologist is notified.

7. The painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion according to claim 1, characterized in that, Step S5 also includes: The big data platform uses a distributed columnar database to store high-concurrency time-series physiological signal data and uses Spark Streaming technology to perform real-time feature calculations on the inbound data, providing data preparation for the immediate optimization of the model.

8. The painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion according to claim 7, characterized in that, Step S5 also includes: During model iteration and optimization, newly added anonymized case data is extracted from the big data platform at fixed intervals to fine-tune the existing assessment model and PK-PD model, continuously absorbing new clinical patterns and achieving continuous improvement in the model's self-evolution and personalized adaptability.

9. The painless gastrointestinal endoscopy adaptive drug delivery system based on multi-source physiological signal fusion according to claim 1, characterized in that, The system execution steps also include: S6. At the end of the examination, based on the total amount of medication administered throughout the process, fluctuations in physiological indicators, and individual patient characteristics, the system uses recovery data from similar cases in the big data platform to predict the patient's awakening time and the probability of postoperative nausea and vomiting, and automatically generates an anesthetic effect and safety assessment report. The system operates on an architecture that combines edge computing nodes and a cloud-based big data platform. The real-time signal processing and drug delivery control tasks in steps S1 to S4 are completed at the edge, while the data archiving and model optimization tasks in step S5 are completed in the cloud to leverage powerful distributed computing capabilities.

Images