Artificial intelligence respiratory system based on anaesthesia machine

Abstract

The invention relates to the technical field of intelligent medical treatment, in particular to an artificial intelligence respiratory system based on an anaesthesia machine, which comprises a data acquisition and processing module, a database establishment module, a stress analysis module and an intelligent analysis module, the data acquisition and processing module is mainly used for acquiring and preprocessing data; the database establishing module is used for establishing a historical database according to the ventilation parameters, the physiological parameters and the individual parameters; the stress analysis module is used for establishing a lung geometric model and generating a visual map of stress deformation of a lung injury area through finite element analysis; and the intelligent analysis module is used for performing model training by utilizing a machine learning algorithm and establishing a breathing parameter optimization model and a lung injury risk assessment model. According to the method, the physiological parameters and the breathing characteristics of the patient are individually integrated in real time in the operation, the dynamic multi-dimensional database is constructed, and the breathing parameters are automatically adjusted to balance the pressure and capacity distribution in the lung, so that the incidence rate of lung injury is reduced.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical technology, and in particular to an artificial intelligence breathing system based on an anesthesia machine. Background Art

[0002] In modern clinical anesthesia practice, mechanical ventilation is a technical means to maintain patient oxygenation and carbon dioxide excretion. The use of anesthesia ventilators, especially the setting of respiratory parameters, will be directly related to the safety of patients, the smooth progress of surgery, and the recovery of the disease. Inappropriate respiratory parameter settings may cause excessive expansion of alveoli or repeated opening and closing, causing ventilator-induced lung injury (VILI). Artificial intelligence technology provides a new approach to solving this problem. By analyzing a large amount of respiratory parameters and clinical data through machine learning algorithms, it can monitor key indicators such as airway pressure and lung compliance in real time, dynamically optimize ventilation parameters, and minimize the risk of lung injury while ensuring adequate ventilation and oxygenation and meeting clinical surgical needs.

[0003] Patent publication number CN116492556A discloses an intelligent control method and system for an anesthesia machine. This patent intelligently adjusts the tidal volume operating parameters of the anesthesia machine based on the machine's tidal volume accuracy and the patient's blood gas analysis sample data, ensuring that the machine's tidal volume operating parameters match the patient's blood gas analysis sample data. However, relying on intermittent blood gas analysis has a certain lag and cannot match real-time ventilation conditions. Furthermore, blood gas analysis cannot fully reflect changes in airway stress during mechanical ventilation.

[0004] In summary, how to solve the potential risk of ventilator-related lung injury in current clinical anesthesia surgery based on existing technologies has become a key and difficult issue that needs to be urgently addressed in this field. Therefore, we proposed an artificial intelligence breathing system based on anesthesia machine. Summary of the Invention

[0005] To solve the above problems, the present invention provides an artificial intelligence respiratory system based on an anesthesia machine. By real-time integration of the ventilation parameters of the anesthesia machine, the patient's physiological parameters and individual parameters, a dynamic multidimensional database is constructed to predict the potential risk of lung injury during ventilation, and automatically adjust the respiratory parameters to balance the pressure and volume distribution in the lungs, thereby reducing the incidence of lung injury.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: an artificial intelligence breathing system based on an anesthesia machine, comprising the following modules:

[0007] The data acquisition and processing module is used to collect the ventilation parameters of the anesthesia machine, the physiological parameters and individual parameters of the patient, pre-process the ventilation parameters, physiological parameters and individual parameters; and transmit the pre-processed ventilation parameters, physiological parameters and individual parameters to the database establishment module.

[0008] The database establishment module is used to establish a historical database based on the ventilation parameters, physiological parameters and individual parameters preprocessed by the data acquisition and processing module.

[0009] The stress analysis module is used to establish a lung geometry model based on the individual parameter information stored in the historical database, import ventilation parameters into the lung geometry model, and generate a visual map of the stress deformation in the lung injury area through finite element analysis.

[0010] The intelligent analysis module is used to obtain the ventilation parameters, physiological parameters and individual parameters collected by the data acquisition and processing module, and use machine learning algorithms to train models; establish a respiratory parameter optimization model and a lung injury risk assessment model, and output the training results of the respiratory parameter optimization model and the lung injury risk assessment model.

[0011] The execution control module is used to receive the training results of the respiratory parameter optimization model and the lung injury risk assessment model in the intelligent analysis module and issue adjustment instructions for the ventilation parameters of the anesthesia machine.

[0012] The interactive module is used to display the training results of the respiratory parameter optimization model and lung injury risk assessment model output by the intelligent analysis module and the visualization map of the stress deformation of the lung injury area in the stress analysis module.

[0013] Furthermore, the data acquisition and processing module uses wireless communication technology to transmit data, and the wireless communication technology is selected from one or more of 4G, 5G, BLE, Wi-Fi, ZigBee, LoRa, Sigfox and NB-IoT.

[0014] Furthermore, the stress analysis module includes the following units:

[0015] The model building unit is used to import the patient's lung medical images into the three-dimensional model building software, process the lung medical images, extract the lung parenchymal density change characteristics, and establish a lung geometric model.

[0016] The parameter import unit is used to import ventilation parameters and physiological parameters into the lung geometric model using finite element analysis software to create boundary conditions of the lung geometric model.

[0017] The stress calculation unit is used to calculate the stress distribution of the lung geometric model and generate a visualization map of lung injury stress and deformation.

[0018] Furthermore, the establishment of the respiratory parameter optimization model includes the following steps:

[0019] S101. Collect ventilation parameters, patient physiological parameters, and individual parameter data collected in a data acquisition and processing module.

[0020] S102. Based on the causes of intraoperative lung injury, a regression model was constructed with lung injury as the dependent variable and ventilation parameters, physiological parameters, and individual parameters as independent variables.

[0021] S103: training the regression model; dynamically updating the training weights of the respiratory parameter optimization model; and generating respiratory parameter optimization adjustment suggestions.

[0022] Furthermore, a stress indicator monitoring module is included, which includes the following units:

[0023] The biomarker collection unit is used to collect stress-related biomarker data in real time.

[0024] The physiological parameter analysis unit is used to integrate the patient's physiological parameters and individual parameters to quantify the intensity of the stress response.

[0025] The data fusion unit is used to fuse the biomarker data with the patient's physiological parameters and individual parameters to generate a comprehensive stress index.

[0026] The stress response unit is used to trigger a graded response based on the comprehensive stress index and identify the type of stressor in combination with the surgical stage and operation events.

[0027] Furthermore, the establishment of the lung injury risk assessment model includes the following training steps:

[0028] S201. Visualize the stress deformation of the lung injury area in the stress analysis module; model different lung injury mechanisms.

[0029] S202. Assign dynamic weights to the biomarker data collected in real time; and adjust the abnormality threshold according to the preoperative biomarker data baseline.

[0030] S203. Output the lung injury risk index and the visualized stress cloud map, stratify the lung injury risk index; and output the lung injury risk grading treatment plan.

[0031] Furthermore, it also includes a preoperative intelligent assessment module, which includes the following units:

[0032] The medical history analysis unit is used to access the patient's electronic medical records, extract risk factors such as height, weight, smoking history, lung disease history, surgical history and trauma history; and output the analysis results of the risk factors to quantify the risk factors.

[0033] The difficulty prediction unit is used to predict the probability of difficult ventilation by combining the visual atlas features of stress and deformation.

[0034] A multidimensional risk reporting unit is used to generate comprehensive reports on stress response prediction, medication alerts, and prevention of postoperative pulmonary complications.

[0035] Furthermore, a cloud hybrid analysis module is included. The establishment of the cloud hybrid analysis module includes the following steps:

[0036] S301. Classify and store real-time data and historical data according to surgery type and patient characteristics.

[0037] S302: training the real-time data and the historical data separately, and fusing the real-time data with the historical data for hybrid training.

[0038] S303: A lightweight model is used for analysis when training real-time data and historical data separately; a deep learning model is used for key analysis when training fused real-time data and historical data.

[0039] Furthermore, the respiratory parameter optimization model uses cross-validation to dynamically adjust the ventilation parameters of the anesthesia machine.

[0040] Furthermore, the lung injury risk assessment model is also used to dynamically detect the changing trends of the patient's physiological parameters through individualized analysis of the triggers of acute risk attacks of potential diseases based on the analysis results of the medical history analysis unit; if the triggers of acute risk attacks of potential diseases appear, the ventilation parameters of the anesthesia machine are adjusted in time, and a complementary plan for the stress response is output.

[0041] The above scheme has the following beneficial effects:

[0042] 1. The present invention integrates the ventilation parameters of the anesthesia machine, the patient's physiological parameters and individual parameters in real time through the data acquisition and processing module to build a dynamic multidimensional database, providing an accurate data basis for subsequent analysis. It also collects the patient's respiratory rate and airway pressure in real time, combines the personalized three-dimensional geometric model generated by preoperative lung medical images, simulates the stress distribution of the lungs during mechanical ventilation through finite element analysis, and generates a stress deformation visualization map of the lung injury area. This biomechanical modeling method can intuitively present high-risk areas of alveolar over-expansion or local collapse, helping doctors to visually identify potential damage. By analyzing the characteristics of uneven lung density distribution, it predicts possible lung injury factors during ventilation, and automatically adjusts respiratory parameters to balance the pressure and volume distribution in the lungs, thereby reducing the incidence of lung injury.

[0043] 2. The present invention uses a machine learning algorithm to train the collected real-time data to establish a respiratory parameter optimization model and a lung injury risk assessment model. The respiratory parameter optimization model dynamically generates ventilation parameter adjustment recommendations based on the patient's individual characteristics and real-time physiological indicators. By real-time monitoring of airway pressure changes and combining stress prediction of the preoperative lung model, respiratory parameters are automatically adjusted to maintain effective ventilation and a good ventilation-to-perfusion ratio while reducing alveolar overexpansion. The lung injury risk assessment model integrates stress analysis results with biomarker data to output risk indices and treatment plans in a stratified manner.

[0044] Secondly, the present invention also utilizes the collaborative operation of lightweight models and deep learning models to achieve efficient integration of real-time and historical data. During individual data analysis, the lightweight model quickly analyzes real-time physiological parameters and, combined with similar case data in the cloud-based historical database, can quickly generate ventilation optimization plans adapted to acute scenarios. By invoking deep learning models to deeply mine large amounts of historical data, potential patterns are discovered and algorithm weights are optimized, thereby improving the generalization capability of model analysis.

[0045] 3. The present invention uses a stress analysis module to deeply integrate clinical imaging data with engineering mechanics methods. By importing the patient's lung medical images into the three-dimensional modeling software, the characteristics of the lung parenchyma density, bronchial branching structure, etc. are extracted to construct a high-precision lung geometric model. By using finite element analysis software to simulate the stress distribution under different ventilation parameters, the system can predict the microscopic damage mechanism of mechanical ventilation to lung tissue, and recommend the optimal PEEP value to balance the improvement of oxygenation and the risk of barotrauma. And by integrating the medical history data in the patient's electronic medical record with the lung imaging characteristics, the probability of intraoperative ventilation difficulty is predicted and the initial respiratory parameter setting range is recommended. In addition, the multidimensional risk reporting unit can generate a comprehensive report covering stress response prediction, drug interaction warnings and postoperative complication prevention, providing data support for multidisciplinary collaboration.

[0046] 4. The present invention outputs dynamic optimization suggestions for respiratory parameters through the intelligent analysis module, and the execution control module directly sends instructions to the anesthesia machine to adjust parameters such as tidal volume, inspiration-expiration ratio, and PEEP value in real time. This automated control can significantly improve the timeliness and accuracy of parameter adjustment. The conversion of complex data into intuitive graphics and charts through a visual interface helps to improve clinical decision-making efficiency. The main interface displays the lung stress cloud map, risk index curve and ventilation parameter change trend in real time. Doctors can retrieve the patient's preoperative assessment report, intraoperative real-time data and historical case comparison with one click. For junior anesthesiologists, the real-time decision support provided by the system can serve as an important reference to help them quickly grasp the management points of complex cases. For highly difficult surgeries, the system can provide precise parameter guidance for lung protective ventilation by combining finite element models with machine learning algorithms, thereby reducing the risk of postoperative lung complications.

[0047] 5. This invention integrates biomarker data with physiological parameters to quantify the intensity of a patient's stress response and identify the type of stressor. Through effective management and timely intervention of stress responses, perioperative complications can be reduced and recovery accelerated.

[0048] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a structural block diagram of the artificial intelligence breathing system based on the anesthesia machine of the present invention.

[0050] Figure 2 This is a structural block diagram of the stress analysis module in the artificial intelligence breathing system based on the anesthesia machine of the present invention.

[0051] Figure 3 This is a structural block diagram of the stress index monitoring module in the artificial intelligence breathing system based on the anesthesia machine of the present invention.

[0052] Figure 4 This is a structural block diagram of the preoperative intelligent evaluation module in the artificial intelligence breathing system based on the anesthesia machine of the present invention. DETAILED DESCRIPTION

[0053] The following is further described in detail through specific implementation methods:

[0054] The embodiment is basically as shown in the attached Figures 1-4 Shown: An artificial intelligence respiratory system based on an anesthesia machine, including a data acquisition and processing module, a database establishment module, a stress analysis module, an intelligent analysis module, a stress index monitoring module, an execution control module, an interaction module, a preoperative intelligent evaluation module and a cloud hybrid analysis module.

[0055] Among them, the data acquisition and processing module is mainly used to collect data and perform preprocessing; the database establishment module is used to establish a historical database; the stress analysis module is used to establish a lung geometric model and generate a visual map of the stress deformation in the lung injury area through finite element analysis; the intelligent analysis module is used to use machine learning algorithms for model training, establish a respiratory parameter optimization model and a lung injury risk assessment model; the stress index monitoring module is used to monitor the patient's stress index data; the execution control module is used to issue adjustment instructions for the ventilation parameters of the anesthesia machine; the interaction module is used to display the training results and visual maps output by the intelligent analysis module; the preoperative intelligent evaluation module is used to obtain the patient's medical history information and output a multi-dimensional comprehensive analysis report; the cloud-based hybrid analysis module is used to train the data separately and hybridly.

[0056] The following is an introduction to the functions of each module:

[0057] The data acquisition and processing module is used to collect the ventilation parameters of the anesthesia machine, the patient's physiological parameters and individual parameters, and pre-process the ventilation parameters, physiological parameters and individual parameters; and transmit the pre-processed ventilation parameters, physiological parameters and individual parameters to the database establishment module; in this embodiment, the anesthesia machine obtains ventilation parameters in real time through RS-485; physiological parameters are collected through a multi-parameter monitor; and after collection, the physiological parameters are transmitted through the Wi-Fi 6 protocol.

[0058] Among them, ventilation parameters include inspired oxygen concentration (FiO2, range 21%-100%), tidal volume (VT, range 50-1500mL), peak airway pressure (Ppeak, range 0-60cmH2O), inspiration to expiration ratio (I:E, range 1:1 to 1:4), positive end-expiratory pressure (PEEP, range 0-20cmH2O) and respiratory rate (RR, range 5-40 times / min).

[0059] Physiological parameters include airway pressure (MPa), end-tidal carbon dioxide (EtCO2, range 0-100 mmHg), oxygen saturation (SpO2, range 70%-100%) and lung compliance index (Cstat, calculation formula: Cstat = lung volume change ΔV / unit pressure change ΔP, calculated and output in real time).

[0060] Individual parameters include gender, age, lung medical imaging, BMI index, and hemodynamic data. Hemodynamic data includes blood pressure, heart rate, and cardiac output. Hemodynamic data is collected using a non-invasive cardiac output monitoring device, which collects blood pressure (NIBP, range 30-250 mmHg), heart rate (HR, range 30-200 bpm), and cardiac output (CO, range 1.0-10.0 L / min) in real time. In this embodiment, the patient's gender, age, and BMI are obtained through the HL7 / FHIR protocol. Lung medical imaging is obtained by accessing the hospital's PACS system to retrieve the patient's preoperative lung CT or MRI images, and an image segmentation algorithm (based on the U-Net architecture) is used to extract the lung parenchymal density distribution, bronchial tree structure, and lesion area characteristics.

[0061] The database establishment module is used to establish a historical database based on the ventilation parameters, physiological parameters, and individual parameters preprocessed by the data acquisition and processing module. This historical database is built using a hybrid storage architecture (MySQL, a relational database, and InfluxDB, a time-series database). The relational database is used to store structured data (such as patient information, drug allergy history, and surgical procedure type), while the time-series database is used to store high-frequency dynamic data (ventilation parameters and physiological parameters).

[0062] The stress analysis module is used to establish a lung geometry model based on the individual parameter information stored in the historical database, import ventilation parameters into the lung geometry model, and generate a visual map of the stress deformation in the lung injury area through finite element analysis.

[0063] Specifically, the stress analysis module includes a model building unit, a parameter import unit, and a stress calculation unit. The specific functions of each unit are as follows:

[0064] The model building unit is used to import the patient's lung medical images into the 3D model building software, process the lung medical images, extract the lung parenchyma density change characteristics, and build a lung geometric model. In this embodiment, the patient's preoperative lung CT images (DICOM format, slice thickness ≤ 1mm) are imported and an adaptive threshold segmentation algorithm (based on ITK-SNAP software) is used to separate the lung parenchyma, bronchial and vascular structures, and eliminate artifacts and noise.

[0065] Mimics 22.0 software was used to convert the segmented two-dimensional slices into three-dimensional point cloud models, and the surface mesh was generated using the MarchingCubes algorithm; ANSYS ICEM CFD was used for volume meshing.

[0066] Based on the CT value-density relationship formula (ρ=0.001*CT value+1.0g / cm 3 ) The elastic modulus of the lung parenchyma was calculated, where the modulus of normal lung tissue (CT value = -800HU) was E = 1.2 kPa, and the modulus of the emphysema area (CT value = -950HU) was E = 0.8 kPa; the bronchial wall adopted the hyperelastic Ogden model (μ = 15 kPa, α = 2.0), and the vascular wall was defined as a linear elastic material (E = 50 kPa, ν = 0.45).

[0067] The parameter import unit is used to import ventilation parameters and physiological parameters into the lung geometric model using finite element analysis software to create boundary conditions of the lung geometric model.

[0068] In this example, ventilation parameters (e.g., Ppeak = 28 cmH2O, PEEP = 8 cmH2O, and I / E ratio = 1:2) were acquired in real time from a historical database and converted into mechanical boundary conditions. A cyclic pressure load (inspiratory pressure = Ppeak, expiratory pressure = PEEP) was applied, with the waveform fitted as a sinusoidal curve. Displacement constraints (fixed displacement of the thoracic wall nodes in the X, Y, and Z directions ≤ 0.1 mm) were set to simulate rigid thoracic support. Real-time lung compliance (Cstat = 35 mL / cmH2O) was dynamically linked to the model's elastic modulus. If Cstat decreased by 10%, the elastic modulus was automatically corrected to E = 1.5 kPa to reflect the tendency of lung tissue to sclerosis.

[0069] The stress calculation unit is used to calculate the stress distribution of the lung geometry model and generate a visualization of the stress and deformation of lung injury. In this example, the ANSYS Mechanical solver was used for transient nonlinear analysis (time step = 0.1 seconds, total duration = 3 respiratory cycles); the convergence criteria were set as follows: node residual error < 1e-4N, energy error < 1e-6J, and maximum number of iterations = 50.

[0070] Calculate the maximum principal stress (σ_max), shear stress (τ_max) and strain energy density (U) of the alveolar wall and identify areas exceeding the safety threshold: barotrauma risk area (σ_max>10kPa red mark); collapse risk area (U<0.05J / m 3 and local strain>30% (blue mark).

[0071] ParaView 5.10 is used to render three-dimensional stress cloud maps, supporting simultaneous display of multiple views (coronal, sagittal, or transverse). The interactive module provides a dynamic probe function, allowing doctors to click on any area to obtain detailed mechanical parameters.

[0072] The intelligent analysis module is used to obtain the ventilation parameters, physiological parameters and individual parameters collected by the data acquisition and processing module, and use machine learning algorithms to train models; establish a respiratory parameter optimization model and a lung injury risk assessment model, and output the training results of the respiratory parameter optimization model and the lung injury risk assessment model.

[0073] Specifically, the establishment of the respiratory parameter optimization model includes the following steps:

[0074] S101. Collect ventilation parameters, patient physiological parameters, and individual parameter data collected in the data acquisition and processing module. In this embodiment, 100,000 surgical records were extracted from a historical database, including ventilation parameters (such as VT, Ppeak, and FiO2), physiological parameters (SpO2, EtCO2, and lung compliance), and individual parameters (BMI, lung CT features, and preoperative pulmonary function). Missing data were filled using the multiple interpolation method (MICE), and continuous variables were standardized to the [0, 1] interval.

[0075] Feature extraction mainly includes static features, dynamic features and imaging features. Static features mainly include patient age, gender and BMI; dynamic features mainly include intraoperative coefficient of variation (CV_VT = σ(VT) / μ(VT)) and Ppeak trend slope (ΔPpeak / Δt); imaging features mainly include the proportion of emphysema area (the proportion of voxels with CT value ≤ -950HU) and bronchial bifurcation angle (calculated by three-dimensional model).

[0076] S102. Based on the causes of intraoperative lung injury, a regression model is constructed with lung injury as the dependent variable and ventilation parameters, physiological parameters, and individual parameters as independent variables. In this embodiment, a multi-task regression model is constructed with optimal oxygenation index (PaO2 / FiO1) and minimum lung stress (σ_max) as dual objectives.

[0077] Specifically, the algorithm selection and formula derivation adopt the XGBoost regression model, and its prediction function formula is:

[0078]

[0079] Among them, fk is the predicted value of the kth tree, and F is the tree space;

[0080] The regularization objective function formula is:

[0081]

[0082] Where T is the number of leaf nodes, w is the leaf weight, γ = 0.1, λ = 1.0; l is the Huber loss function, which enhances the robustness of the model to outliers.

[0083] S103: training the regression model; dynamically updating the training weights of the respiratory parameter optimization model; and generating respiratory parameter optimization adjustment suggestions.

[0084] Specifically, every 100 new surgical data triggers incremental model training, and the update formula is as follows:

[0085] ω t+1 =ω t -η.▽L(ω t ) (3);

[0086] Among them, η = 0.01 is the learning rate, and the Adam optimizer is used for adaptive adjustment; the most recent 10,000 data are retained with high weight (exponential decay factor = 0.95).

[0087] The respiratory parameter optimization model uses cross-validation to dynamically adjust the ventilation parameters of the anesthesia machine.

[0088] In this embodiment, k-fold cross validation is used to divide the data set into 5 subsets, one subset is used as the validation set in turn, and the rest are used as the training set; the average performance index is calculated.

[0089] If the model predicted VT error was greater than 10% for three consecutive times, the training frequency was automatically increased (from every 100 cases to every 50 cases). The importance of features was analyzed using SHAP values, and features with a contribution of less than 1% (such as patient gender) were dynamically eliminated.

[0090] The stress index monitoring module is used to monitor the patient's stress index data. Specifically, the stress index monitoring module includes a biomarker acquisition unit, a physiological parameter analysis unit, a data fusion unit, and a stress response unit. The functions of each unit are as follows:

[0091] The biomarker collection unit is used to collect stress-related biomarker data in real time. In this embodiment, a sweat sensor is used to detect lactate and cortisol concentrations (accuracy ±0.1 mmol / L); a microneedle patch is used to continuously monitor epinephrine and norepinephrine levels in subcutaneous tissue fluid; and a blood gas analyzer is used to obtain intraoperative arterial blood pH, lactate (Lac), and blood glucose (GLU).

[0092] The physiological parameter analysis unit integrates patient physiological and individual parameters to quantify the intensity of the stress response. Specifically, heart rate variability (HRV) reflects sympathetic nerve activity through the low-frequency / high-frequency power ratio (LF / HF), while blood pressure fluctuation indicates a high stress state through a systolic blood pressure standard deviation (SD) greater than 15 mmHg.

[0093] The data fusion unit is used to fuse biomarker data with patient physiological parameters and individual parameters to generate a comprehensive stress index (0-100%). The formula is as follows:

[0094] Comprehensive stress index = 0.4 × cortisol score + 0.3 × HRV score + 0.3 × Lac score (4);

[0095] The stress response unit is used to trigger a graded response based on the comprehensive stress index, identifying the type of stressor by combining the surgical stage and procedural events. Specifically, the graded response standard is: low stress (0-30%) is monitored routinely without intervention; medium stress (31-70%) is recommended by push notification optimization suggestions (such as "deepen the depth of anesthesia"); high stress (>70%) is linked to the drug infusion system (such as adding 0.1μg / kg / min of remifentanil).

[0096] In this embodiment, the stressor type is realized through positioning logic: mechanical stress is a sudden change in ventilation parameters (such as a sudden increase in driving pressure ΔP) leading to an increase in cortisol; painful stress is an increase in HRV accompanied by an increase in remifentanil demand; metabolic stress is a continuous increase in lactate and a pH <7.25.

[0097] The establishment of the lung injury risk assessment model includes the following training steps:

[0098] S201. Visualize the stress deformation of the lung injury area in the stress analysis module; model different lung injury mechanisms.

[0099] Specifically, based on the results of finite element stress analysis, the lung injury mechanism was defined as follows: barotrauma refers to a main stress σ_max greater than 10 kPa and a duration greater than 5 minutes; volutrauma refers to an alveolar strain ε greater than 25% and a tidal volume VT greater than 8 mL / kg ideal body weight; and biotrauma refers to an interleukin IL-6 greater than 50 pg / mL or a verification factor TNF-α greater than 20 pg / mL, accompanied by an increase in inflammatory factors in the stress concentration area.

[0100] In this embodiment, a weighted fusion algorithm is used to calculate the comprehensive risk index (LIRI):

[0101] LIRI=0.5Barotrauma Score+0.3Volutrauma Score+0.2Biotrauma Score(5);

[0102] S202. Dynamic weights are assigned to the biomarker data collected in real time; abnormal thresholds are adjusted based on the preoperative biomarker data baseline. Baseline threshold calculation is performed by collecting serum biomarkers (IL-6, CRP, KL-6) from patients before surgery to calculate an individualized baseline using the following formula:

[0103] Baseline threshold = μ preoperative + 2σ preoperative (6);

[0104] Here, μ represents the mean and σ is the standard deviation.

[0105] For example, if a patient's preoperative IL-6 mean is 15 pg / mL and the standard deviation is 3 pg / mL, then the intraoperative abnormal threshold is 21 pg / mL.

[0106] Dynamic weight allocation is activated when real-time biomarker data (e.g., IL-6 = 45 pg / mL) exceeds a baseline threshold, using the following formula:

[0107] ω bio =min(1.0, 0.5+0.1×(IL-6-baseline threshold)) (7);

[0108] If IL-6 = 45 pg / mL and baseline threshold = 21 pg / mL, then ω bio =0.5+0.1×(45-21)=2.9, truncated to 1.0; Update LIRI formula:

[0109] LIRI=0.4Barotrauma Score+0.3Volutrauma Score+0.3Biotrauma Score(8);

[0110] S203: Output the lung injury risk index and visualized stress cloud map, stratify the lung injury risk index, and output the lung injury risk grading treatment plan. Specifically, the risk index stratification standard is low risk (LIRI ≤ 2.0) with a green mark, recommending maintenance of current parameters; medium risk (2.0 < LIRI ≤ 4.0) with a yellow mark, recommending a 10% reduction in VT and an increase in PEEP of 2 cmH2O; high risk (LIRI > 4.0) with a red mark, recommending immediate discontinuation of pneumoperitoneum (laparoscopic surgery) or switching to pressure control mode.

[0111] For example, a three-color risk heat map is superimposed on the interactive interface, with red areas corresponding to σ_max > 10 kPa or IL-6 > threshold; clicking on a high-risk area allows users to view detailed treatment plans (such as "Risk of right middle lobe barotrauma: recommended PEEP + 2 cmH2O, FiO2 + 5%").

[0112] The preoperative intelligent assessment module is used to obtain the patient's medical history information and output a multi-dimensional comprehensive analysis report. Specifically, the preoperative intelligent assessment module includes a medical history analysis unit, a difficulty prediction unit, and a multi-dimensional risk reporting unit. The specific functions of each unit are as follows:

[0113] The medical history analysis unit is used to access the patient's electronic medical records, extract risk factors such as height, weight, smoking history, history of lung diseases (such as chronic bronchitis, asthma, emphysema, pulmonary hypertension, cor pulmonale, etc.), surgical history and trauma history; and output the analysis results of risk factors to quantify the risk factors.

[0114] Specifically, a natural language processing (NLP) engine (based on a BERT pre-trained model) was used to parse unstructured text (e.g., "20-year smoking history") and extract key risk factors: Lung disease history was matched with ICD-10 codes (J44.9) and pulmonary function reports (FEV1 / FVC < 0.7); the smoking history was quantified using the following formula:

[0115] Smoking index = number of packs smoked per day × number of years of smoking (9);

[0116] The classification is mild (<10), moderate (10-30), and severe (>30). Non-standard descriptions (such as "long-term smoking") are converted using a rule engine (the default is 10 pack-years per day for 10 years).

[0117] The difficulty prediction unit is used to predict the probability of difficult ventilation by combining visual atlas features of stress and deformation. Specifically, finite element feature extraction obtains the mechanical characteristics of the preoperative lung geometry model from the stress analysis module. Principal component analysis (PCA) is used for dimensionality reduction to extract the first three principal components (cumulative variance contribution ≥ 85%).

[0118] In this embodiment, the prediction model uses a random forest classifier to predict the probability of difficult ventilation (defined as: intubation attempts ≥ 3 or need for fiberoptic bronchoscope), and its formula is as follows:

[0119]

[0120] Among them, the weight parameters (ω1=0.52, ω2=0.33, ω3=0.15, b=-1.2) are fitted through 5000 training sets; if P>0.65, it is judged as high risk and an early warning is triggered.

[0121] Visual atlas fusion helps doctors intuitively assess anatomical abnormalities by superimposing three-dimensional airway stenosis areas (such as red marks) and stress concentration areas (such as orange marks) on the interactive interface.

[0122] The multidimensional risk reporting unit is used to generate comprehensive reports on stress response prediction, medication warnings, and prevention of postoperative pulmonary complications. Specifically, stress response prediction is based on preoperative heart rate variability (HRV) and cortisol levels (>20μg / dL). The LSTM time series model is used to predict intraoperative stress intensity. The formula is as follows:

[0123] Stress intensity = LSTM(HRV_SDNN,Cortisol,BMI)(11);

[0124] The output of stress intensity is graded as low (<3), medium (3-5), and high (>5), and corresponding intervention measures (such as additional analgesia or α2 receptor agonists) are used.

[0125] Drug interaction warnings are linked to drug databases (such as ), construct a knowledge graph to analyze the interaction risk between the proposed drug (such as rocuronium bromide) and the patient's current medication (such as beta-blockers).

[0126] The postoperative complication prevention program combines the patient's PRI index with the type of surgery and calls on the evidence-based medicine rule library to generate personalized recommendations. For example, for the prevention of lung infection, when the PRI is greater than 3, nebulized inhalation of budesonide (0.5 mg) is recommended 2 hours before surgery; for the prevention of deep vein thrombosis, when the BMI is greater than 30 and the operation duration is greater than 3 hours, intraoperative intermittent pneumatic compression (IPC) treatment is recommended; for the optimization of analgesia strategies, for patients with a smoking index greater than 20 pack-years, avoid the use of opioids as monotherapy for analgesia, and recommend multimodal analgesia (such as NSAIDs + regional blockade).

[0127] The report output is generated in PDF / HTML format, which includes a risk overview using a visual instrument to display the PRI index, probability of difficult ventilation and stress intensity level; the intervention list includes medication adjustments, ventilation strategies and complication prevention measures sorted by priority.

[0128] The lung injury risk assessment model is also used to dynamically detect the changing trends of the patient's physiological parameters through individualized analysis of the triggers of acute risk attacks of potential diseases based on the analysis results of the medical history analysis unit. If triggers of acute risk attacks of potential diseases appear, the ventilation parameters of the anesthesia machine are adjusted in a timely manner, and a complementary plan for the stress response is output.

[0129] Specifically, the medical history association rules are constructed by constructing a knowledge graph to associate the patient's medical history (such as radiation lung injury and pneumoconiosis) with the susceptible injury type: if the patient has a history of pneumoconiosis, the volume injury weight is increased by 20%; if the preoperative CT shows pulmonary fibrosis, the barotrauma safety threshold is lowered to σ_max≤8kPa.

[0130] The trend prediction algorithm uses the LSTM network to predict the trend of physiological parameters (such as PaO2 / FiO2 in the next 5 minutes). The formula is as follows:

[0131] h t =LSTM(x t , h t-1 ) (12),

[0132] Among them, the input feature x t Including VT, PEEP and SpO2, etc., output predicted values If PaO2 / FiO2 is predicted to be less than 200, alveolar recruitment maneuvers should be triggered in advance.

[0133] The execution control module is used to receive the training results of the respiratory parameter optimization model and the lung injury risk assessment model in the intelligent analysis module, and issue adjustment instructions for the ventilation parameters of the anesthesia machine.

[0134] The interactive module is used to display the training results of the respiratory parameter optimization model and lung injury risk assessment model output by the intelligent analysis module and the visualization map of the stress deformation of the lung injury area in the stress analysis module.

[0135] The cloud-based hybrid analysis module is used to perform separate and hybrid training on data. The establishment of the cloud-based hybrid analysis module includes the following steps:

[0136] S301. Categorize and store real-time and historical data by surgical type and patient characteristics. Specifically, this is done by establishing data labeling and hierarchical storage. For example, the surgical type label defines the surgical category based on ICD-9-CM codes (e.g., laparoscopic cholecystectomy = 51.23, lobectomy = 32.49); the patient characteristic label identifies patient attributes based on preoperative assessment results.

[0137] The data association index is constructed using Elasticsearch to establish an inverted index, supporting multi-condition compound queries (such as "retrieve cases with Ppeak>30cmH2O in laparoscopic surgery with BMI>30 in 2023"); extract key frame features (based on ResNet-50) from unstructured data (such as surgical videos), generate feature vectors and store them in the vector database (Milvus), supporting similar case retrieval.

[0138] S302: Train real-time data and historical data separately, and then fuse the real-time data with historical data for hybrid training. Specifically, lightweight training uses lightweight random forest (tree depth = 5, number of trees = 50) or XGBoost (depth = 3, learning rate = 0.1); the training process extracts the latest 1000 data points from the hot data layer every 5 minutes and performs online learning.

[0139] The hybrid training strategy mixes real-time data (hot data layer) with similar historical cases (warm data layer) in a ratio of 6:4 to enhance model generalization; Adversarial Training: Adds Gaussian noise (σ = 0.1) and adversarial samples to the training set to improve model robustness; Federated Learning (FL): Shares model parameters rather than raw data among multiple centers, and updates the global model through weighted aggregation (such as the FedAvg algorithm).

[0140] S303: A lightweight model is used for analysis when training real-time data and historical data separately; a deep learning model is used for key analysis when training fused real-time data and historical data.

[0141] Specifically, the random forest model is converted into TensorFlow Lite format and deployed to the edge computing node in the operating room (NVIDIA Jetson AGX Xavier), supporting offline inference (latency <50ms); the dynamic memory allocation strategy automatically switches the model accuracy based on the remaining memory to ensure real-time performance.

[0142] Cloud-based inference of deep learning models uses a Kubernetes cluster to schedule GPU resources (NVIDIA A100), launching the MT-Transformer model for complex queries (such as "predicting the risk of intraoperative lung injury in asthma patients"); result caching and updating: For high-frequency queries (such as laparoscopic surgery ventilation strategies), prediction results are cached and incrementally updated every 15 minutes.

[0143] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. For those skilled in the art, other variations or modifications can be made based on the above description. It is not necessary and impossible to list all implementation methods here.

Claims

1. An artificial intelligence breathing system based on an anesthesia machine, characterized in that: Includes the following modules: The data acquisition and processing module is used to collect the ventilation parameters of the anesthesia machine, the physiological parameters and individual parameters of the patient, and pre-process the ventilation parameters, physiological parameters and individual parameters; and transmitting the pre-processed ventilation parameters, physiological parameters and individual parameters to the database establishment module; A database establishment module is used to establish a historical database based on the ventilation parameters, physiological parameters and individual parameters pre-processed by the data acquisition and processing module; The stress analysis module is used to establish a lung geometry model based on the individual parameter information stored in the historical database, import ventilation parameters into the lung geometry model, and generate a visualization map of the stress and deformation of the lung injury area through finite element analysis; The intelligent analysis module is used to obtain the ventilation parameters, physiological parameters, and individual parameters collected by the data acquisition and processing module, and use machine learning algorithms to train models; establish a respiratory parameter optimization model and a lung injury risk assessment model, and output the training results of the respiratory parameter optimization model and the lung injury risk assessment model; An execution control module is used to receive the training results of the respiratory parameter optimization model and the lung injury risk assessment model in the intelligent analysis module and issue an adjustment instruction for the ventilation parameters of the anesthesia machine; The interactive module is used to display the training results of the respiratory parameter optimization model and lung injury risk assessment model output by the intelligent analysis module and the visualization map of the stress deformation of the lung injury area in the stress analysis module.

2. The artificial intelligence breathing system based on anesthesia machine according to claim 1, characterized in that: The data acquisition and processing module uses wireless communication technology to transmit data. The wireless communication technology can be one or more of 4G, 5G, BLE, Wi-Fi, ZigBee, LoRa, Sigfox and NB-IoT.

3. The artificial intelligence breathing system based on anesthesia machine according to claim 2, characterized in that: The stress analysis module includes the following units: A model building unit is used to import the patient's lung medical images into the 3D model building software, process the lung medical images, extract the lung parenchyma density change characteristics, and build a lung geometric model; A parameter import unit is used to import ventilation parameters and physiological parameters into the lung geometric model using finite element analysis software to create boundary conditions of the lung geometric model; The stress calculation unit is used to calculate the stress distribution of the lung geometric model and generate a visualization map of lung injury stress and deformation.

4. The artificial intelligence breathing system based on anesthesia machine according to claim 3, characterized in that: The establishment of the respiratory parameter optimization model includes the following steps: S101, collecting ventilation parameters, patient physiological parameters and individual parameter data collected in the data acquisition and processing module; S102. Based on the causes of intraoperative lung injury, a regression model was constructed with lung injury as the dependent variable and ventilation parameters, physiological parameters, and individual parameters as independent variables; S103: training the regression model; dynamically updating the training weights of the respiratory parameter optimization model; and generating respiratory parameter optimization adjustment suggestions.

5. The artificial intelligence breathing system based on anesthesia machine according to claim 4, characterized in that: It also includes a stress indicator monitoring module, which includes the following units: A biomarker collection unit, used to collect stress-related biomarker data in real time; A physiological parameter analysis unit is used to integrate the patient's physiological parameters and individual parameters to quantify the intensity of the stress response; a data fusion unit for fusing biomarker data with patient physiological parameters and individual parameters to generate a comprehensive stress index; The stress response unit is used to trigger a graded response based on the comprehensive stress index and identify the type of stressor in combination with the surgical stage and operation events.

6. The artificial intelligence breathing system based on anesthesia machine according to claim 5, characterized in that: The establishment of the lung injury risk assessment model includes the following training steps: S201. Modeling different lung injury mechanisms based on the visualization of stress and deformation in the lung injury region in the stress analysis module. S202, assigning dynamic weights to the biomarker data collected in real time; Abnormal thresholds were adjusted based on baseline preoperative biomarker data; S203, outputting a lung injury risk index and a visual stress cloud map, and stratifying the lung injury risk index; And output lung injury risk grading treatment plan.

7. The artificial intelligence breathing system based on anesthesia machine according to claim 6, characterized in that: It also includes a preoperative intelligent assessment module, which includes the following units: Medical history analysis unit, used to access patients' electronic medical records and extract risk factors such as height, weight, smoking history, history of lung disease, surgical history, and history of trauma; And output the analysis results of risk factors and quantify the risk factors; Difficulty prediction unit, used to predict the probability of difficult ventilation by combining the visual map features of stress and deformation; A multidimensional risk reporting unit is used to generate comprehensive reports on stress response prediction, medication alerts, and prevention of postoperative pulmonary complications.

8. The artificial intelligence breathing system based on anesthesia machine according to claim 7, characterized in that: It also includes a cloud hybrid analysis module. The establishment of the cloud hybrid analysis module includes the following steps: S301. Classify and store real-time data and historical data according to surgery type and patient characteristics; S302, training the real-time data and the historical data separately, and fusing the real-time data with the historical data for hybrid training; S303: A lightweight model is used for analysis when training real-time data and historical data separately; a deep learning model is used for key analysis when training fused real-time data and historical data.

9. The artificial intelligence breathing system based on anesthesia machine according to claim 8, characterized in that: The respiratory parameter optimization model uses cross-validation to dynamically adjust the ventilation parameters of the anesthesia machine.

10. The artificial intelligence breathing system based on anesthesia machine according to claim 9, characterized in that: The lung injury risk assessment model is also used to dynamically detect the changing trends of the patient's physiological parameters through individualized analysis of the triggers of acute risk attacks of potential diseases based on the analysis results of the medical history analysis unit. If triggers of acute risk attacks of potential diseases appear, the ventilation parameters of the anesthesia machine are adjusted in a timely manner, and a complementary plan for the stress response is output.

Images