Self-adaptive ventilation control edge calculation respirator system and method

Abstract

The invention relates to the technical field of medical equipment, in particular to an edge calculation respirator system and method for adaptive ventilation control, and the system comprises a respirator end which is provided with an edge calculation module and a local control model; the edge calculation module is used for collecting multi-modal physiological parameters of a patient and performing data preprocessing on the multi-modal physiological parameters of the patient to generate a multi-dimensional feature vector; the local control model carries out real-time self-adaptive adjustment on ventilation parameters of the breathing machine based on the multi-modal physiological parameters of the patient and the baseline characteristics of the patient; the cloud server is in communication connection with the respirator end and is configured with a global optimization model; and the cloud server receives historical ventilation data of the respirator end, and the global optimization model performs training based on the historical ventilation data to generate a personalized ventilation strategy, and issues the personalized ventilation strategy to the respirator end for correcting adjustment of the local control model. According to the invention, low latency and high reliability of bedside control are guaranteed through a side cloud collaborative architecture.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and more specifically, to an edge computing ventilator system and method for adaptive ventilation control. Background Technology

[0002] Mechanical ventilation is one of the core life-saving support methods in intensive care medicine (ICU), especially crucial for critically ill patients with acute respiratory distress syndrome (ARDS). Its core objective is to minimize ventilator-associated lung injury (VILI) while ensuring adequate gas exchange. The key to achieving this goal lies in the precise and personalized setting of ventilation parameters. Among these, parameters such as positive end-expiratory pressure (PEEP), pressure support level (PS), and tidal volume (VT) directly affect treatment efficacy and patient prognosis. A patient's lung pathophysiology is highly dynamic, with respiratory mechanics and gas exchange efficiency fluctuating over time, with changes in treatment response and body position. Therefore, an ideal ventilation strategy requires real-time, continuous monitoring of the patient's physiological feedback and dynamic optimization of ventilation parameters to maintain optimal lung-protective ventilation. In recent years, with the development of sensor technology and data processing capabilities, utilizing multimodal physiological data (such as airway pressure, flow rate, blood oxygen saturation, and end-tidal carbon dioxide) to construct closed-loop control systems and achieve automated and intelligent adjustment of ventilation parameters has become a research hotspot and development direction in the field of modern respiratory therapy. However, existing ventilation control technologies still face significant challenges in clinical practice. On the one hand, current parameter adjustments mainly rely on manual settings by medical staff based on clinical experience and intermittent blood gas analysis results. This response is lagging, making it difficult to capture the rapidly changing physiological changes in patients, and it carries a certain degree of subjectivity, failing to guarantee real-time and optimal adjustment. On the other hand, although some automated ventilation modes exist, their built-in algorithm models are mostly relatively simple, mainly based on feedback adjustment using limited real-time parameters, lacking the ability to deeply mine massive amounts of historical patient data and predict long-term trends. Furthermore, while emerging cloud-based intelligent analysis solutions can utilize powerful computing power to process complex models, their inherent network latency and connection instability make it difficult to meet the millisecond-level response requirements for control commands in the ICU environment, and their direct application to life support equipment poses safety risks. Therefore, current research faces an irreconcilable contradiction in balancing the security of real-time response with the forward-looking nature of global optimization. Summary of the Invention

[0003] The purpose of this invention is to provide an edge computing ventilator system and method for adaptive ventilation control. Through an innovative edge-cloud collaborative architecture, it combines real-time rapid response at the edge with deep global optimization in the cloud, which not only ensures low latency and high reliability of bedside control, but also provides forward-looking personalized strategies by utilizing big data analysis in the cloud, thus achieving unprecedentedly accurate, safe and intelligent adaptive ventilation control.

[0004] This invention is achieved through the following technical solution: An edge computing ventilator system with adaptive ventilation control, comprising: The ventilator is equipped with an edge computing module and a local control model. The edge computing module is used to collect the patient's multimodal physiological parameters in real time, including raw airway pressure signals, flow rate, end-expiratory carbon dioxide, and blood oxygen saturation. It also performs data preprocessing on the patient's multimodal physiological parameters to generate multidimensional feature vectors. The local control model is used to adaptively adjust the ventilator's ventilation parameters in real time based on the preprocessed patient multimodal physiological parameters and patient baseline characteristics. The ventilator uploads the multidimensional feature vectors and ventilation parameter sequences as historical ventilation data to the cloud server through the communication unit. The cloud server communicates with the ventilator and is equipped with a global optimization model. The cloud server receives and stores historical ventilation data from the ventilator. The global optimization model is trained based on the historical ventilation data to generate personalized ventilation strategies, which are then sent to the ventilator to correct the local control model's settings.

[0005] Optionally, the data preprocessing of the edge computing module specifically includes: The adaptive wavelet denoising module is configured to: perform multi-level wavelet decomposition on the acquired raw airway pressure signal, calculate the dynamic threshold based on the statistical characteristics of the airway pressure signal, and reconstruct the decomposed coefficients after threshold processing to generate the denoised airway pressure signal. The multi-parameter feature fusion module is configured to extract inspiratory peak pressure, plateau pressure, inspiratory time percentage, end-expiratory carbon dioxide plateau value, and blood oxygen saturation fluctuation rate features from the noise-reduced airway pressure signal and other multimodal physiological parameters, and construct them into a multidimensional feature vector.

[0006] Optionally, the local control model is a fuzzy neural network model, which is configured as follows: It receives data from multiple input nodes, including patient baseline features and real-time physiological features provided by feature vectors, and organizes the data into input vectors. Parameter mapping is performed through a hidden layer, which sets multiple fuzzy subsets for each input variable. The membership value of each subset is calculated through a trapezoidal membership function, and the rule activation strength is calculated based on the membership value through a product inference engine. The rule activation strength is solved by multiplying multiple membership values. Defuzzification is performed on the output layer using the centroid method, which calculates a weighted average based on the rule activation intensity and the preset output center value. The optimized positive end-expiratory pressure and pressure support level are then output as the optimized ventilation parameters.

[0007] Optionally, the training of the fuzzy neural network model specifically employs an adaptive gradient optimization algorithm, whose objective function is defined as simultaneously maximizing the average value of the fuzzy rule activation intensity and the stability of the output parameters, and its calculation formula is as follows:

[0008] in, Let be the objective function. These are the weighting coefficients. For the total number of rules, For rule indexing, For the number of input variables, For input variable index, For input values, For the first Rule number 1 The trapezoidal membership values ​​of each input. For sequence length, For the first The output of each time step To output the mean, It is a small positive number.

[0009] Optionally, it also includes an extracorporeal membrane oxygenation (ECMO) device that works in conjunction with the ventilator; the ECMO device incorporates a pressure-flow coupling model, the specific configuration of which is as follows: The coupling relationship between ventilator airway pressure and extracorporeal membrane oxygenation (ECMO) equipment is established based on the hemodynamic balance equation. The hemodynamic balance equation includes multiple fitting coefficients and error terms to represent the dynamic balance between pressure and flow. The fitting coefficients were obtained from clinical data, and the coupling error was calculated, whereby the coupling error was controlled by minimizing a combination of pressure deviation and flow deviation. The airway pressure at the ventilator and the blood flow of the extracorporeal membrane oxygenation (ECMO) device are adjusted according to the hemodynamic balance equation. The adjustment process includes error feedback iteration to maintain a balance within a preset range.

[0010] Optionally, the global optimization model of the cloud server is specifically a Transformer-LSTM hybrid model, and its construction process is as follows: Historical ventilation data sequences are processed by a Transformer encoder containing multiple attention heads. The Transformer encoder computes scaled dot product attention, obtains attention weights by performing dot product operations on query, key, and value matrices and applying the softmax function, and captures long-term dependencies in the sequences by concatenating the outputs of multiple heads and performing linear transformations. The output of the Transformer encoder is stabilized by applying residual connections and layer normalization mechanisms. The residual connections add the input to the sub-layer output, and the layer normalization normalizes the addition result. The output of the Transformer encoder is serialized using a decoder containing multiple LSTM layers. Each LSTM layer updates the hidden state and cell state through forget gates, input gates, and output gates to generate a personalized ventilation strategy.

[0011] Optionally, the objective function of the global optimization model is defined as simultaneously maximizing the average information gain of the attention weights and the generation consistency of the decoded sequence, and its calculation formula is as follows:

[0012] in, Let be the objective function. These are the weighting coefficients. For the number of attention heads index, For the number of attention heads, For the first Attention weight distribution based on size, For sequence length, For the first Decoding output at each time step, The mean of the decoded output. This represents the variance of the decoded output.

[0013] Optionally, the local control model and the global optimization model work together through the following optimization steps to achieve edge-cloud ventilation parameter adjustment: The local control model generates preliminary ventilation parameters in real time and uploads multi-dimensional feature vectors to the cloud server; The cloud server uses a global optimization model to analyze multidimensional feature vectors and historical ventilation data sequences to generate personalized ventilation strategies, which are then sent to the ventilator as optimization instructions. At the ventilator, a weighted fusion mechanism is used to combine local results and cloud commands; Real-time adaptive adjustment of ventilation parameters is achieved based on fusion parameters.

[0014] Optionally, the combination of local results and cloud instructions at the ventilator end via a weighted fusion mechanism is specifically calculated using the following formula:

[0015]

[0016] in, These are the final ventilation parameters after fusion. For dynamic weighting coefficients, Preliminary ventilation parameters generated for the local control model. Personalized ventilation strategies generated for the global optimization model. It is the sigmoid activation function. For weight parameters, For weight parameters, This is due to physiological feedback error in blood oxygen saturation. This refers to the physiological feedback error of carbon dioxide at the end of expiration. For bias.

[0017] An edge computing method for adaptive ventilation control in a ventilator, the method comprising the following steps: At the ventilator, the edge computing module collects the patient's multimodal physiological parameters in real time and performs data preprocessing on the patient's multimodal physiological parameters to generate multidimensional feature vectors; At the ventilator end, the ventilation parameters of the ventilator are adjusted in real time based on the multidimensional feature vector and the patient's baseline characteristics through the local control model, and the multidimensional feature vector and ventilation parameter sequence are uploaded to the cloud server as historical ventilation data. On the cloud server, historical ventilation data is analyzed through a global optimization model to generate personalized ventilation strategies, which are then sent to the ventilator to correct the local control model's settings.

[0018] The technical solution of the present invention has at least the following advantages and beneficial effects: This invention, on the one hand, deploys an edge computing module and a lightweight local control model at the ventilator end, enabling real-time acquisition and analysis of the patient's multimodal physiological data. This allows for instantaneous, low-latency adaptive adjustment of ventilation parameters, greatly ensuring stable system operation and rapid response to the patient's urgent physiological changes even under extreme conditions such as network fluctuations or interruptions. On the other hand, this invention utilizes a cloud server to deploy a global optimization model. Through deep learning analysis of massive historical data accumulated over a long period, it can uncover the unique pathophysiological patterns of individual patients and generate forward-looking, personalized long-term ventilation strategies. Furthermore, this invention organically combines immediate tactical adjustments at the edge with global strategic guidance from the cloud through a weighted fusion mechanism, achieving fast and accurate collaborative control. This not only ensures the immediacy and safety of control but also endows the system with continuous learning and optimization capabilities through cloud intelligence, thereby providing truly individualized and precise ventilation therapy. Attached Figure Description

[0019] Figure 1 A schematic diagram of the edge computing ventilator system for adaptive ventilation control provided by the present invention; Figure 2 This is a schematic flowchart of the edge computing ventilator method for adaptive ventilation control provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Example 1 like Figure 1 As shown, this embodiment of the invention provides an edge computing ventilator system for adaptive ventilation control. This embodiment aims to address the lack of multimodal real-time monitoring, personalized parameter adjustment, and real-time error feedback mechanisms in existing ventilator systems. The overall architecture of this embodiment includes a ventilator terminal, a cloud server, and can work collaboratively with extracorporeal membrane oxygenation (ECMO) devices.

[0022] Specifically, the ventilator is the core execution unit of the system, equipped with an edge computing module and a local control model. The edge computing module is responsible for processing the patient's physiological data in real time, while the local control model adjusts ventilation parameters in real time based on this data. The cloud server, as the central brain of the system, maintains a communication connection with the ventilator. It is equipped with a global optimization model to receive and analyze the large amount of historical data uploaded from the ventilator, thereby generating a forward-looking personalized ventilation strategy and issuing instructions to correct local adjustments.

[0023] Furthermore, this embodiment also integrates the ability to coordinate with extracorporeal membrane oxygenation (ECMO) equipment. Through a pressure-flow coupling model, it achieves coordinated regulation of the respiratory and circulatory systems to improve the overall treatment effect and safety.

[0024] Example 2 In this embodiment, the edge computing module at the ventilator end is key to achieving real-time response. The edge computing module embeds a high-performance edge computing chip, and its workflow mainly includes data acquisition, adaptive wavelet denoising, and multi-parameter feature fusion.

[0025] The edge computing module collects multimodal physiological parameters of patients in real time at a sampling frequency of 1000Hz, mainly including raw airway pressure signal, flow signal, end-tidal carbon dioxide concentration and blood oxygen saturation.

[0026] An adaptive wavelet denoising module performs high-precision processing on the acquired raw airway pressure signal. This process uses a db4 wavelet basis to perform a 5-level multi-stage wavelet decomposition of the pressure data, obtaining the signal approximation coefficients and detail coefficients at each level. Based on the statistical characteristics of the airway pressure signal—specifically, 1.5 times the average pressure over the previous 10 seconds—the system calculates a dynamic threshold. This dynamic threshold is then used to perform hard thresholding on the high-frequency detail coefficients after decomposition, effectively filtering out noise interference. Finally, wavelet reconstruction generates the denoised airway pressure signal. After this processing, the signal-to-noise ratio of the pressure waveform is significantly improved from the original 25dB to 40dB, providing a high-quality data source for subsequent feature extraction.

[0027] The multi-parameter feature fusion module extracts key information from the denoised multimodal physiological parameters to construct a 12-dimensional multi-dimensional feature vector. This multi-dimensional feature vector contains a comprehensive description of the patient's condition, with specific feature dimensions including: peak inspiratory pressure and plateau pressure extracted from the pressure waveform; inspiratory time percentage extracted from the flow waveform; plateau value extracted from end-expiratory carbon dioxide data; and 5-minute fluctuation rate of blood oxygen saturation reflecting oxygenation stability. The completed feature vector is transmitted via wired network in JSON format to the local control model and the cloud server.

[0028] Example 3 In this embodiment, the local control model adopts a fuzzy neural network model. Its design goal is to adjust the ventilation parameters of the ventilator in real time and adaptively based on the real-time input multidimensional feature vector and the patient's baseline characteristics. Its structure is divided into an input layer, a hidden layer and an output layer.

[0029] The input layer contains six input nodes that receive input vectors that integrate static and dynamic patient information. The input vector data includes patient baseline features, such as patient type (e.g., newborn, adult, or obese patient), weight, and underlying diseases (e.g., whether the patient has ARDS), as well as real-time physiological features provided by multidimensional feature vectors, such as peak inspiratory pressure, plateau pressure, and percentage of inspiratory time.

[0030] The hidden layer is the core inference part of the model, responsible for parameter mapping. For each input variable, the hidden layer sets three fuzzy subsets, such as small, medium, and large. A trapezoidal membership function is used to calculate the membership value of the current input value with respect to each fuzzy subset. Subsequently, the product inference engine calculates the activation strength of each predefined fuzzy rule based on these membership values. The rule activation strength is solved by multiplying the membership values ​​of all preconditions participating in that rule. For example, a rule might be that if the platform pressure is high and the inspiratory time percentage is small, then the pressure support level should be increased.

[0031] The output layer employs the centroid method for defuzzification, transforming fuzzy rule conclusions into precise control commands. The centroid method calculates a weighted average based on the strength of all activated rules and the preset output center value for each rule. Ultimately, the model outputs optimized positive end-expiratory pressure (PEEP) and pressure support level (PS), which are the optimized ventilation parameters directly used to adjust the ventilator.

[0032] To ensure accurate and stable model decisions, an adaptive gradient optimization algorithm, specifically the Adam optimization algorithm, is used during training. The learning rate is set to 0.001, the batch size to 512, and the model converges after 1000 training iterations. The objective function in this embodiment is designed to simultaneously maximize two objectives: the average activation intensity of the fuzzy rules and the stability of the output parameters. The formula for calculating the objective function is:

[0033] in, Let be the objective function. These are the weighting coefficients. For the total number of rules, For rule indexing, For the number of input variables, For input variable index, For input values, For the first Rule number 1 The trapezoidal membership values ​​of each input. For sequence length, For the first The output of each time step To output the mean, It is a small positive number, used to prevent the denominator from being zero.

[0034] Example 4 In this embodiment, the global optimization model of the cloud server adopts a Transformer-LSTM hybrid model to deeply mine massive amounts of historical ventilation data in order to generate highly personalized long-term ventilation strategies.

[0035] The construction process of this embodiment is as follows: Historical ventilation data sequences are processed using a Transformer encoder with eight attention heads. The Transformer encoder captures long-term dependencies between arbitrary time points within the sequence by computing scaled dot product attention. Its core is the dot product operation of the query, key, and value matrices, with attention weights obtained through a softmax function. The multi-head mechanism allows the model to learn information from different perspectives; its output, after concatenation and linear transformation, comprehensively represents the deep features of the sequence. To ensure the stability of deep network training, residual connections and layer normalization mechanisms are applied after each sublayer. Residual connections directly add the input to the sublayer output, while layer normalization normalizes the sum.

[0036] A decoder comprising three LSTM layers serializes the output of the Transformer encoder. Each LSTM layer selectively updates its hidden state and cell state through its ingenious internal forget gate, input gate, and output gate structure, thereby understanding and generating an output sequence with temporal logic. Ultimately, the decoder generates personalized ventilation strategies. For example, for patients with high pulmonary edema-type ARDS, a small tidal volume (6-8 mL / kg) and stepwise PEEP downregulation (reducing PEEP by 2 cmH2O every 2 hours) is recommended; while for patients with low compliance, a higher PEEP level (12-14 cmH2O) is recommended after lung recruitment maneuvers.

[0037] The global optimization model in this embodiment has an objective function designed to simultaneously maximize the average information gain of the attention weights and the generation consistency of the decoded sequence. Its calculation formula is as follows:

[0038] in, Let be the objective function. These are the weighting coefficients. For the number of attention heads index, For the number of attention heads, For the first Attention weight distribution based on size, For sequence length, For the first Decoding output at each time step, The mean of the decoded output. This represents the variance of the decoded output.

[0039] More specifically, to facilitate deployment, the trained cloud model is compressed using knowledge distillation technology, reducing its size to approximately 50MB, and is updated to the edge computing environment weekly. Cloud commands are invoked via a 5G network, achieving inference speeds as fast as 8 milliseconds.

[0040] Example 5 In this embodiment, a major innovation is designed, which is a highly efficient collaborative workflow between the local control model and the global optimization model, achieving a perfect combination of the rapid response of edge computing and the deep intelligence of cloud computing.

[0041] Specifically, this time-saving and labor-saving collaborative process achieves edge-to-cloud ventilation parameter adjustment through the following optimization steps: The local control model on the ventilator generates a preliminary ventilation parameter in milliseconds based on real-time data, and simultaneously uploads the multi-dimensional feature vector on which the parameter is based to the cloud server.

[0042] After receiving the data, the cloud server's global optimization model will perform a comprehensive analysis of all the patient's historical ventilation data sequences to generate a more macroscopic and personalized ventilation strategy, which will then be sent to the corresponding ventilator as an optimization instruction.

[0043] Upon receiving optimization instructions from the cloud, the system does not execute them directly. Instead, it uses a dynamic weighted fusion mechanism to combine the real-time results generated by the local model with the long-term instructions issued from the cloud. The specific calculation formula for this fusion process is as follows:

[0044]

[0045] in, These are the final ventilation parameters after fusion. This is a dynamic weighting coefficient, the value of which is determined by the patient's real-time physiological feedback error. Preliminary ventilation parameters generated for the local control model. Personalized ventilation strategies generated for the global optimization model. It is the sigmoid activation function. For weight parameters, These are weighting parameters, reflecting the importance of different physiological indicators. This is due to physiological feedback error in blood oxygen saturation. This refers to the physiological feedback error of carbon dioxide at the end of expiration. This is a bias. The design based on this embodiment makes the system more inclined to follow the long-term strategy in the cloud when the patient's physiological state is stable. The weight of the local model is relatively small; however, when the patient's condition suddenly deteriorates, leading to an increase in physiological feedback error, the system will dynamically increase the weight of the local model. (larger), to ensure the most timely response.

[0046] Finally, based on the fused parameters, the ventilation parameters are adjusted in real time to form a closed-loop control system that continuously learns and optimizes.

[0047] Example 6 More specifically, this embodiment incorporates several advanced features designed for complex clinical scenarios and system reliability.

[0048] When a patient is simultaneously using an extracorporeal membrane oxygenation (ECMO) device, the system activates its built-in pressure-flow coupling model. This model, based on hemodynamic equilibrium equations, establishes a mathematical coupling relationship between the airway pressure provided by the ventilator and the blood flow provided by the ECMO device. By fitting the coefficients in the equations to clinical data and calculating the coupling error, the system can collaboratively adjust the airway pressure of the ventilator and the blood flow of the ECMO device, maintaining hemodynamic stability in the patient while ensuring gas exchange and avoiding mutual interference between their regulation.

[0049] The system integrates an intelligent weaning prediction function. This function is based on a logistic regression model and comprehensively assesses factors such as the decreasing trend of pressure support level (PS), respiratory rate, and blood gas parameters (e.g., pH). For example, when a patient's PS level is monitored to be below 5 cmH2O, respiratory rate below 30 breaths / minute, and arterial blood pH above 7.35, the model predicts a weaning success rate of ≥85%, providing data support for healthcare professionals' decision-making.

[0050] To ensure maximum system reliability, a dual-chip hot backup mechanism is employed. The system continuously monitors the response time of the primary chip, and if the response time exceeds a preset threshold (e.g., 5 milliseconds), a hot-swap switching mechanism is triggered, seamlessly switching to the backup chip within less than 10 milliseconds. This effectively avoids the risk of ventilation interruption due to a single point of hardware failure.

[0051] In addition, the system includes a data anonymization module that encrypts and anonymizes all patient identity information before uploading data to the cloud, ensuring patient privacy and security. Simultaneously, the system supports remote monitoring and diagnostic functions, allowing authorized medical staff to view detailed patient data and system operating status in real time via mobile devices, and to remotely adjust parameters and provide guidance.

[0052] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An edge computing ventilator system for adaptive ventilation control, characterized in that, include: The ventilator is equipped with an edge computing module and a local control model; The edge computing module is used to collect patients' multimodal physiological parameters in real time, including raw airway pressure signals, flow rate, end-tidal carbon dioxide and blood oxygen saturation. The module also performs data preprocessing on the patients' multimodal physiological parameters to generate multidimensional feature vectors. The local control model is used to adaptively adjust the ventilation parameters of the ventilator in real time based on the preprocessed multimodal physiological parameters and baseline characteristics of the patient. The ventilator uploads the multidimensional feature vector and ventilation parameter sequence as historical ventilation data to the cloud server through the communication unit. The cloud server communicates with the ventilator and is configured with a global optimization model. The cloud server is used to receive and store historical ventilation data from the ventilator. The global optimization model is trained based on the historical ventilation data to generate personalized ventilation strategies, which are then sent to the ventilator to correct the local control model's settings.

2. The edge computing ventilator system for adaptive ventilation control according to claim 1, characterized in that, The data preprocessing of the edge computing module specifically includes: The adaptive wavelet denoising module is configured to: perform multi-level wavelet decomposition on the acquired raw airway pressure signal, calculate the dynamic threshold based on the statistical characteristics of the airway pressure signal, and reconstruct the decomposed coefficients after threshold processing to generate the denoised airway pressure signal. The multi-parameter feature fusion module is configured to extract inspiratory peak pressure, plateau pressure, inspiratory time percentage, end-expiratory carbon dioxide plateau value, and blood oxygen saturation fluctuation rate features from the noise-reduced airway pressure signal and other multimodal physiological parameters, and construct them into a multidimensional feature vector.

3. The edge computing ventilator system for adaptive ventilation control according to claim 1 or 2, characterized in that, The local control model is a fuzzy neural network model, which is configured as follows: It receives data from multiple input nodes, including patient baseline features and real-time physiological features provided by feature vectors, and organizes the data into input vectors. Parameter mapping is performed through a hidden layer, which sets multiple fuzzy subsets for each input variable. The membership value of each subset is calculated through a trapezoidal membership function, and the rule activation strength is calculated based on the membership value through a product inference engine. The rule activation strength is solved by multiplying multiple membership values. Defuzzification is performed on the output layer using the centroid method, which calculates a weighted average based on the rule activation intensity and the preset output center value. The optimized positive end-expiratory pressure and pressure support level are then output as the optimized ventilation parameters.

4. The edge computing ventilator system for adaptive ventilation control according to claim 3, characterized in that, The training of the fuzzy neural network model specifically employs an adaptive gradient optimization algorithm. Its objective function is defined as simultaneously maximizing the average value of the fuzzy rule activation intensity and the stability of the output parameters. The calculation formula is as follows: in, Let be the objective function. These are the weighting coefficients. For the total number of rules, For rule indexing, For the number of input variables, For input variable index, For input values, For the first Rule number 1 The trapezoidal membership values ​​of each input. For sequence length, For the first The output of each time step To output the mean, It is a small positive number.

5. The edge computing ventilator system for adaptive ventilation control according to claim 1, characterized in that, It also includes an extracorporeal membrane oxygenation (ECMO) device that works in conjunction with the ventilator; the ECMO device has a built-in pressure-flow coupling model, the specific configuration of which is as follows: The coupling relationship between ventilator airway pressure and extracorporeal membrane oxygenation (ECMO) equipment is established based on the hemodynamic balance equation. The hemodynamic balance equation includes multiple fitting coefficients and error terms to represent the dynamic balance between pressure and flow. The fitting coefficients were obtained from clinical data, and the coupling error was calculated, whereby the coupling error was controlled by minimizing a combination of pressure deviation and flow deviation. The airway pressure at the ventilator and the blood flow of the extracorporeal membrane oxygenation (ECMO) device are adjusted according to the hemodynamic balance equation. The adjustment process includes error feedback iteration to maintain a balance within a preset range.

6. The edge computing ventilator system for adaptive ventilation control according to claim 1, characterized in that, The global optimization model of the cloud server is specifically a Transformer-LSTM hybrid model, and its construction process is as follows: Historical ventilation data sequences are processed by a Transformer encoder containing multiple attention heads. The Transformer encoder computes scaled dot product attention, obtains attention weights by performing dot product operations on query, key, and value matrices and applying the softmax function, and captures long-term dependencies in the sequences by concatenating the outputs of multiple heads and performing linear transformations. The output of the Transformer encoder is stabilized by applying residual connections and layer normalization mechanisms. The residual connections add the input to the sub-layer output, and the layer normalization normalizes the addition result. The output of the Transformer encoder is serialized using a decoder containing multiple LSTM layers. Each LSTM layer updates the hidden state and cell state through forget gates, input gates, and output gates to generate a personalized ventilation strategy.

7. The edge computing ventilator system for adaptive ventilation control according to claim 6, characterized in that, The objective function of the global optimization model is defined as simultaneously maximizing the average information gain of the attention weights and the generation consistency of the decoded sequence, and its calculation formula is as follows: in, Let be the objective function. These are the weighting coefficients. For the number of attention heads index, For the number of attention heads, For the first Attention weight distribution based on size, For sequence length, For the first Decoding output at each time step, The mean of the decoded output. This represents the variance of the decoded output.

8. The edge computing ventilator system for adaptive ventilation control according to claim 1, characterized in that, The local control model and the global optimization model work together through the following optimization steps to achieve edge-cloud ventilation parameter adjustment: The local control model generates preliminary ventilation parameters in real time and uploads multi-dimensional feature vectors to the cloud server; The cloud server uses a global optimization model to analyze multidimensional feature vectors and historical ventilation data sequences to generate personalized ventilation strategies, which are then sent to the ventilator as optimization instructions. At the ventilator, a weighted fusion mechanism is used to combine local results and cloud commands; Real-time adaptive adjustment of ventilation parameters is achieved based on fusion parameters.

9. The edge computing ventilator system for adaptive ventilation control according to claim 8, characterized in that, The specific calculation formula for combining local results and cloud instructions at the ventilator end using a weighted fusion mechanism is as follows: in, These are the final ventilation parameters after fusion. For dynamic weighting coefficients, Preliminary ventilation parameters generated for the local control model. Personalized ventilation strategies generated for the global optimization model. It is the sigmoid activation function. For weight parameters, For weight parameters, This is due to physiological feedback error in blood oxygen saturation. This refers to the physiological feedback error of carbon dioxide at the end of expiration. For bias.

10. An edge computing ventilator method for adaptive ventilation control, characterized in that, The steps of this method include: At the ventilator, the edge computing module collects the patient's multimodal physiological parameters in real time and performs data preprocessing on the patient's multimodal physiological parameters to generate multidimensional feature vectors; At the ventilator end, the ventilation parameters of the ventilator are adjusted in real time based on the multidimensional feature vector and the patient's baseline characteristics through the local control model, and the multidimensional feature vector and ventilation parameter sequence are uploaded to the cloud server as historical ventilation data. On the cloud server, historical ventilation data is analyzed through a global optimization model to generate personalized ventilation strategies, which are then sent to the ventilator to correct the local control model's settings.

Images