External counterpulsation self-adaptive regulation and control system and method based on generative adversarial network and digital twinning

Abstract

The invention relates to the technical field of intelligent control of medical equipment, in particular to an external counterpulsation self-adaptive regulation and control system and method based on a generative adversarial network and digital twinning. The system acquires data through the acquisition module; the multi-modal signal processing module preprocesses the collected signals, and key hemodynamic features are extracted through multi-modal feature fusion; the digital twinborn modeling module constructs a specific blood vessel digital twinborn model, takes key hemodynamic characteristics, multi-modal physiological signals and patient static parameters as core input, and outputs hemodynamic simulation data in real time; the self-adaptive regulation and control module outputs counterpulsation pressure and a time sequence scheme on the basis of hemodynamics simulation data; an execution verification module drives an air bag to sequentially pressurize, a counterpulsation enhancement index is monitored in real time, parameter re-optimization is triggered through curative effect deviation, and meanwhile model re-training is carried out; according to the invention, the defects of weak signal anti-interference capability, lack of individual dynamic adaptation, disjunction of long-term curative effect prediction and the like of traditional external counterpulsation control are overcome.

Description

An Adaptive Control System and Method for External Counterpulsation Based on Generative Adversarial Networks and Digital Twins Technical Field

[0001] This invention relates to the field of intelligent control technology for medical devices, specifically to an adaptive control system and method for external counterpulsation based on generative adversarial networks and digital twins. Background Technology

[0002] Traditional external counterpulsation devices generally rely on the ECG R wave to trigger the inflation and deflation of the cuff. The principle is to determine the end of the systolic phase by detecting the R wave peak, and then drive the cuff to inflate during diastole to enhance coronary blood flow. However, this method has significant limitations in clinical application: ECG / PPG signals are susceptible to electromyographic noise (50-60Hz power frequency interference), motion artifacts (patient positional changes), and the electromagnetic environment of the ward, leading to inaccurate extraction of R wave or ST segment features. Studies have shown that under complex heart rhythms such as atrial fibrillation and frequent premature ventricular contractions, the triggering error of traditional filtering algorithms (such as wavelet threshold denoising) can reach more than 120ms, causing severe asynchrony between the balloon movement and the heart's diastolic phase, and a 30%-40% decrease in blood flow pressurization efficiency. At the same time, existing control models use fixed parameters (such as a uniform inflation pressure of 0.35MPa), which cannot respond to differences in patients' vascular elasticity (the pressure needs to be reduced when the arteriosclerosis coefficient of elderly diabetic patients is >9.0), sudden changes in heart rate (such as a sudden increase in stress heart rate from 60bpm to 120bpm), or dynamic ST segment deviation (a key marker of acute myocardial ischemia). Clinical data show that about 38% of patients with coronary heart disease and diabetes have abnormal vascular compliance, and their ankle-brachial index (ABI) improves by less than 0.1 after receiving standard counterpulsation therapy. Furthermore, the lack of correlation between parameter adjustments and endothelial function biomarkers (such as endothelin-1 (ET-1) and vascular endothelial growth factor (VEGF)) leads to a disconnect between treatment protocols and actual efficacy. A randomized trial involving 200 patients with peripheral artery disease (PAD) confirmed that after 12 weeks of treatment with traditional methods, only 52% of patients experienced a decrease in serum ET-1 concentration >15%, and the final efficacy trend could not be predicted midway through the treatment course, delaying the opportunity to optimize individualized protocols. Although technologies have incorporated machine learning algorithms (such as Kalman filtering to fuse multiple physiological signals), they are still limited to data-level correction: prediction models based on shallow neural networks have a simulation error of >25% for vascular topological variations, the system requires >500ms to resynchronize under atrial fibrillation rhythm, the treatment interruption rate is as high as 35%, and there is a lack of closed-loop verification mechanisms for biomechanical indicators, making it impossible to avoid the risk of local ischemia. In summary, current external counterpulsation technology urgently needs to overcome three major bottlenecks: lack of anatomical-level blood flow simulation, weak dynamic anti-interference, and disconnection from biomarker-driven approaches, in order to achieve precise treatment in complex physiological scenarios. Summary of the Invention

[0003] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide an external counterpulsation adaptive control system and method based on generative adversarial networks and digital twins. This system overcomes the defects of traditional external counterpulsation control, such as weak signal anti-interference ability, lack of individual dynamic adaptation and disconnection from long-term efficacy prediction, and achieves anatomical-level precision treatment and closed-loop optimization of long-term and short-term efficacy.

[0004] This invention is achieved through the following technical solution: an external counterpulsation adaptive control system based on generative adversarial networks and digital twins includes an acquisition module, a multimodal signal processing module, a digital twin modeling module, an adaptive control module, and an execution verification module; the acquisition module is used to acquire electrocardiogram signals, photoplethysmography signals, cardiac impedance signals, and vascular CT image data; the multimodal signal processing module preprocesses the acquired signals based on variational mode decomposition, and fuses the multimodal features of the acquired information through a cross-modal attention fusion mechanism to extract key hemodynamic features; the digital twin modeling module constructs... A specific vascular digital twin model is constructed, using key hemodynamic features, multimodal physiological signals, and patient static parameters as core inputs, and outputting hemodynamic simulation data in real time. The adaptive control module, based on the hemodynamic simulation data, outputs individualized counterpulsation pressure and timing schemes through three stages: initial parameters, dynamic correction, and long-term optimization. The execution verification module drives sequential inflation of the balloon according to the individualized counterpulsation pressure and timing scheme, monitors the counterpulsation enhancement index in real time, and feeds back the efficacy deviation to the adaptive control module to trigger parameter re-optimization, while simultaneously feeding back to the digital twin modeling module for model retraining.

[0005] The multimodal signal processing module includes a variational mode decomposition unit and a cross-modal attention fusion unit; the variational mode decomposition unit is used to separate electromyographic noise and motion artifact interference from photoplethysmography signals in electrocardiogram signals; the cross-modal attention fusion unit is used to extract key hemodynamic features.

[0006] The variational mode decomposition unit is set with a decomposition layer number K=8, a penalty factor α=2000, and a center frequency constraint range of 0.5-40 Hz; the cross-modal attention fusion unit uses a weight allocation formula to dynamically enhance the importance of key hemodynamic features. The weight allocation formula normalizes the linear combination of ECG features and photoplethysmography features through the Softmax function.

[0007] The digital twin modeling module is built on a generative adversarial network architecture and includes a generator and a discriminator. The generator takes multimodal physiological signals and patient static parameters as input and outputs real-time hemodynamic simulation data of the cardiovascular system. The discriminator optimizes the model weights by comparing the dynamic differences between the real arterial pressure waveform and the simulation data.

[0008] The generator adopts a 3D-U-Net convolutional neural network structure, and its input layer integrates the temporal features of electrocardiogram signals and the three-dimensional topological data reconstructed from vascular CT images. The discriminator is based on a temporal convolutional network architecture, using the measured radial artery pressure waveform with a sampling frequency of 1000Hz as the authenticity judgment benchmark, and calculates the Wasserstein distance between the generator-generated data and the measured waveform to dynamically update the network weights.

[0009] The adaptive regulation module includes an initial parameter unit, a dynamic correction unit, and a long-term optimization unit. The initial parameter unit generates individualized counterpulsation pressure and timing benchmarks based on the output of a specific vascular digital twin model (vascular elasticity parameters and hemodynamic simulation data). The dynamic correction unit uses an adversarial learning algorithm to respond to ST segment deviation or heart rate mutation events. The long-term optimization unit associates endothelial function biomarker concentrations through a reinforcement learning reward function.

[0010] When the ST segment horizontal depression amplitude is detected to be ≥0.1mV or atrial fibrillation, the dynamic correction unit automatically compresses the inflation delay time to within 5 milliseconds after the ST segment initiation; the deflation timing uses the aortic valve closure time point predicted by a specific vascular digital twin model as the trigger benchmark.

[0011] The reward function in the long-term optimization unit consists of three weighted parts: coronary perfusion improvement rate accounts for 0.6, endothelin-1 concentration prediction error term accounts for 0.3, and vascular endothelial growth factor growth rate accounts for 0.1. When the reward function value calculated for three consecutive days is lower than the 0.7 threshold, the system automatically extends the single treatment time by 10% to 20% and adjusts the regional pressure gradient combination.

[0012] The execution verification module controls the sequential inflation action of the airbag and calculates the counterpulsation enhancement index in real time. If the counterpulsation enhancement index is detected to be lower than the set threshold for three consecutive times (the threshold can be set to 0.15), the parameters are re-optimized. After each treatment course, the system automatically compares the endothelial function improvement rate predicted by the specific vascular digital twin model with the actual detection value. When the absolute error exceeds 15%, the model retraining process is started.

[0013] The adaptive control method for external counterpulsation based on generative adversarial networks (GANs) and digital twins, applied to the aforementioned adaptive control system for external counterpulsation based on GANs and digital twins, includes the following steps: Data acquisition and processing: Real-time acquisition of patient ECG signals, photoplethysmography (PPG) signals, cardiac impedance signals, and vascular CT image data; high-quality key hemodynamic features are extracted through variational mode decomposition and cross-modal attention fusion; Model construction and optimization: Constructing a specific vascular digital twin model, outputting hemodynamic simulation data, and dynamically optimizing network weights based on arterial pressure waveforms; Specifically, the generator outputs a wall shear force simulation sequence, and the discriminator compares the arterial pressure waveforms to dynamically optimize the weights; Adaptive control: including an initial stage, a dynamic correction stage, and a long-term optimization stage; wherein, the initial stage: based on the specific vascular digital twin model... The output sets the counterpulsation pressure and timing parameters; dynamic correction phase: if ST segment deviation ≥0.1mV or atrial fibrillation (RR interval variation >15%) is detected, the inflation and deflation timing is recalculated within 100ms; long-term optimization phase: serum ET-1 and / or VEGF concentrations are collected every 24h, and the counterpulsation pressure and timing scheme for subsequent treatments are adjusted through a reward function; execution verification: the balloon is driven to perform sequential inflation according to the optimized timing scheme based on the counterpulsation pressure and timing scheme, and the counterpulsation enhancement index is monitored in real time. If the counterpulsation enhancement index is detected to be lower than the set threshold for three consecutive times, the process jumps to the dynamic correction phase and re-optimizes the parameters; after each complete treatment course, the deviation between the endothelial function prediction value predicted by the specific vascular digital twin model and the actual detection result is automatically compared. If the absolute error exceeds 15%, the model retraining process is initiated to adapt to the evolution of the patient's vascular status.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This application uses a 3D-U-Net generator to fuse individual vascular CT images and ECG temporal features, dynamically outputting the wall shear force distribution (resolution 0.1 mm) and strain tensor, thereby increasing the coronary artery diastolic pressure boosting efficiency to 35% (compared to an average of 21% for traditional methods). In particular, for patients with complex vascular structures with bifurcation angles >60°, the blood flow prediction error is reduced from 22% to 7%, which helps to improve the effect of subsequent targeted therapy.

[0015] This invention innovatively employs variational mode decomposition (VMD) to constrain the frequency band (0.5-40Hz) for noise separation, combined with a cross-modal attention mechanism to enhance the feature weights of the diastolic notch (>0.8), thereby improving the signal-to-noise ratio by 40%. Furthermore, when ST-segment depression is ≥0.1mV or there is a sudden change in heart rate, the timing sequence is recalculated within 100ms using an adversarial learning algorithm (Dueling DQN), ensuring a treatment success rate of 89.7% for atrial fibrillation patients (compared to only 78.6% with traditional R-wave triggering), and reducing the clinical interruption rate to below 3%. The reinforcement learning reward function is associated with ET-1 / VEGF biomarkers, with an endothelial function prediction error of <15%, allowing for prediction of efficacy trends mid-treatment and achieving a closed loop of long-term and short-term efficacy. Robustness under extreme physiological scenarios is significantly enhanced.

[0016] This application utilizes a reinforcement learning reward function (Reward = 0.6 × coronary perfusion rate + 0.3 × ET-1 prediction accuracy + 0.1 × VEGF growth) to correlate serum indicators in real time, enabling mid-treatment efficacy prediction (endothelial function prediction error <15%). Treatment duration is dynamically adjusted based on the Reward value (e.g., extended by 15% when <0.7), increasing the ET-1 reduction target achievement rate to 86%. Simultaneously, the execution validation module employs real-time monitoring of the Enhanced Counterpulsation Index (EHI) and GAN model retraining as a double safeguard, effectively mitigating the risk of local ischemia (balloon pressure over-limit alarm response <100ms). It is compatible with complex scenarios such as vascular tortuosity and obesity, and the prediction error for aortic bifurcation variations is reduced from 18% to 7% after retraining, enhancing clinical universality.

[0017] This system supports a wide range of indications, from stable angina pectoris to severe peripheral artery occlusion (ABI<0.4). For high-risk groups such as those with diabetes (CAVI>9.0) and the elderly (>80 years old), the system automatically reduces the baseline pressure by 10%-20%. The synchronous exhaust design controlled by a high-speed solenoid valve (end-diastolic error <8ms) avoids blood backflow, and the pressurization efficiency is 40% higher than that of traditional balloon exhaust, combining therapeutic efficiency and safety. Attached Figure Description

[0018] Figure 1 is a flowchart of the overall system architecture and data closed loop of the present invention; Figure 2 is a flowchart of the model building and adversarial training of the present invention. Detailed Implementation

[0019] 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 a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1, referring to Figures 1-2, describes an external counterpulsation adaptive control system based on generative adversarial networks and digital twins, which includes an acquisition module, a multimodal signal processing module, a digital twin modeling module, an adaptive control module, and an execution verification module.

[0021] The acquisition module is used to acquire multimodal information such as patient electrocardiogram (ECG lead II, sampling rate 1kHz), photoplethysmography (PPG) signal (toe), cardiac impedance (ICG) signal, and vascular CT data (specifically aortic CT image, DICOM format). Specifically, in this embodiment, the acquisition module integrates an ECG sensor, a photoplethysmography sensor, a cardiac impedance sensor, and a vascular image input unit to acquire the above multimodal data information in real time, providing raw data support for subsequent processing and modeling.

[0022] The multimodal signal processing module preprocesses the acquired signals based on variational mode decomposition and extracts key hemodynamic features by fusing multimodal features of the acquired information through a cross-modal attention fusion mechanism.

[0023] The multimodal signal processing module includes a variational mode decomposition unit and a cross-modal attention fusion unit.

[0024] The variational mode decomposition (VMD) unit is used to separate electromyographic noise and motion artifact interference from photoplethysmography (PPG) signals in the electrocardiogram (ECG) signal. Specifically, in this embodiment, the decomposition layer number K=8, the penalty factor α=2000, and the center frequency constraint range is 0.5-40 Hz to effectively separate ECG electromyographic noise (35-45Hz power frequency interference) and PPG motion artifacts (0.1-5Hz low-frequency drift). The cross-modal attention fusion unit is used to extract key hemodynamic features such as the diastolic pulse wave notch. The cross-modal attention fusion unit uses a weight allocation formula to dynamically increase the importance of key diastolic hemodynamic features. The weight allocation formula normalizes the linear combination of ECG features and PPG features using the Softmax function.

[0025] Specifically, this embodiment sets a weight allocation formula. It is expressed as follows: In the formula, This represents the cross-modal attention weight, used to characterize the relative importance of ECG features and photoplethysmography features during fusion.

[0026] The Softmax normalization function maps the linear combination of ECG features and photoplethysmography features to the [0, 1] interval, thus achieving weight normalization.

[0027] This represents the rectified linear unit activation function, used to introduce nonlinearity, enhance the model's ability to express features, and filter negative feature values.

[0028] The weighting coefficient matrix represents the ECG features and is used to adjust the contribution of ECG features in the fusion process. It is obtained through model training and optimization.

[0029] This represents the feature vector extracted from the electrocardiogram, which includes hemodynamically related temporal features such as RR interval variability and QTc dispersion.

[0030] The weighting coefficient matrix represents the photoplethysmography features, which is used to adjust the contribution of the photoplethysmography features in the fusion process. It is obtained through model training and optimization.

[0031] This represents the feature vector extracted from the photoplethysmography, which includes hemodynamic features such as the diastolic pulse wave notch and waveform slope.

[0032] Furthermore, the cross-modal attention fusion unit employs a cross-modal attention fusion mechanism, dynamically selecting diastolic pulse wave notches (characteristic slope > 0.5 mV / s) according to a weighting formula, increasing the weight of key hemodynamic features to above 0.8, and suppressing interference from irrelevant signals. This process ensures that even in obese patients (BMI > 30) where strong noise is generated due to body movement, the system can still extract pure coronary perfusion window features.

[0033] In summary, the multimodal processing module performs eight-level intrinsic mode decomposition on the ECG signal through the variational mode decomposition unit and uses kurtosis-energy dual threshold screening to filter out electromyographic interference. It also implements adaptive bandwidth decomposition on the photoplethysmography signal to suppress motion artifacts. The cross-modal attention unit constructs a spatiotemporal fusion matrix and assigns a temporal weight of 0.7 to the diastolic notch of the pulse wave based on the differential peak value of cardiac impedance. Through the multi-head attention mechanism, the anti-interference capability of key event detection is significantly improved.

[0034] The digital twin modeling module constructs a specific vascular digital twin model, using key hemodynamic features, multimodal physiological signals, and patient static parameters as core inputs, and outputs hemodynamic simulation data in real time.

[0035] The digital twin modeling module described in this embodiment is built with a Generative Adversarial Network (GAN) as its core architecture. The digital twin modeling module includes a generator and a discriminator.

[0036] The generator takes multimodal physiological signals and patient static parameters as input and outputs real-time hemodynamic simulation data of the cardiovascular system. In this embodiment, the generator employs a 3D-U-Net convolutional neural network structure. It extracts spatial features and upsamples them using 3×3×3 3D convolutional kernels, combining the spatiotemporal correlations of the multimodal input data to simulate and output hemodynamically relevant data. A discriminator is used for optimization and calibration. Using a measured radial artery pressure waveform with a sampling frequency of 1000Hz as a benchmark, the Wasserstein distance between the generated data and the measured waveform is calculated. The generator weights are updated through gradient penalty backpropagation to ensure that the output data matches the patient's actual cardiovascular state (simulation error ≤ 8%).

[0037] Specifically, the multimodal physiological signals include the temporal characteristics of the electrocardiogram (ECG) signal (RR interval variability, QTc dispersion) and the aortic bifurcation topology reconstructed from vascular CT images (generated using the Marching Cubes algorithm, with bifurcation angle accuracy ±2°); the patient's static parameters include patient age, body surface area (BSA), and vascular elasticity coefficient (CAVI). The generator input layer integrates the temporal characteristics of the ECG signal (RR interval variability, QTc dispersion) and the aortic bifurcation topology reconstructed from vascular CT images (generated using the Marching Cubes algorithm, with bifurcation angle accuracy ±2°). Spatial features are extracted and upsampled using a 3×3×3 3D convolutional kernel. The output hemodynamic simulation data includes a wall shear force distribution cloud map per cardiac cycle (spatial resolution 0.1 mm, pseudo-color marking of high shear force regions >7 Pa) and vascular strain tensors (strain range 0-12%).

[0038] The discriminator optimizes model weights by comparing the dynamic differences between real arterial pressure waveforms and simulated data, enabling the specific vascular digital twin model to accurately match the patient's actual cardiovascular state. In this embodiment, the discriminator employs a temporal convolutional network architecture, specifically a five-layer temporal convolutional network (TCN, kernel width = 7, dilation factor = 1 / 2 / 4 / 8 / 16). It uses 1000 consecutive points of the radial artery pressure waveform (Qianping Medical BP-600, 1000Hz sampling) as the accuracy benchmark. The generator calculates the Wasserstein distance (initial value 0.35±0.05) between the generated data and the measured waveform, and updates the generator weights through backpropagation with gradient penalty (λ=10). A "digital mirror image" of the patient's blood vessels is saved every 10 iterations. For example, when the patient's aorta is tortuous (bifurcation angle > 60°), the generator automatically corrects the shear force simulation model, reducing the prediction error from 25% to 7%.

[0039] In summary, the digital twin construction module in this embodiment integrates a high-precision multi-source signal acquisition unit and a distributed training engine. It constructs a multi-dimensional input matrix by simultaneously acquiring 12-lead ECG signals, dual-channel photoplethysmography, four-electrode impedance data, and coronary artery images. The generator network uses a U-Net++ architecture to calculate in real-time the three-dimensional vascular wall shear force distribution cloud map, circumferential strain field gradient, and local blood flow velocity vector field. The discriminator dynamically compares the waveforms measured by the balloon pressure sensor with Doppler ultrasound blood flow data based on a conditional adversarial framework, continuously optimizing network weights through a gradient penalty mechanism. Ultimately, it establishes a patient-specific digital vascular image with a hemodynamic parameter simulation error ≤8%.

[0040] Finally, the output of the specific vascular digital twin model includes: a wall shear force distribution cloud map with a spatial resolution of 0.1 mm, containing pseudo-color markings of high shear force areas >7 Pa, reflecting the force distribution of blood flow impact on various parts of the vessel wall.

[0041] Vascular strain tensor: with a strain range of 0-12%, it characterizes the degree and direction of deformation of the vascular wall under the action of blood flow.

[0042] Per-cardiac cycle blood flow velocity vector field: describes the direction and magnitude of blood flow within blood vessels during different cardiac cycles.

[0043] Endothelial function prediction values ​​include endothelin-1 (ET-1) concentration prediction values, vascular endothelial growth factor (VEGF) growth rate prediction values, and endothelial function improvement rate prediction values.

[0044] The adaptive control module, based on hemodynamic simulation sequences, outputs individualized counterpulsation pressure and timing protocols through three stages: initial parameters, dynamic correction, and long-term optimization. The adaptive control module comprises an initial parameter unit, a dynamic correction unit, and a long-term optimization unit. Based on these three units, the adaptive control module performs precise decision-making through a three-level linkage of initial parameters, dynamic correction, and long-term optimization.

[0045] The initial parameter unit, combined with the output of a specific vascular digital twin model, determines vascular elasticity parameters, and generates individualized counterpulsation pressure and timing baselines based on these parameters. In the initial parameter stage, the digital twin generates individualized baseline pressure based on vascular elasticity parameters (e.g., CAVI=9.2 for diabetic patients). (For a 60-year-old patient with BSA=1.8m², P=180mmHg), the inflation start point was set at 5ms after the ST segment initiation, and the pressure holding time accounted for 30% of the cardiac cycle.

[0046] Vascular elasticity parameters (such as vascular compliance and arteriosclerosis coefficient CAVI) are mainly obtained based on vascular CT imaging data and hemodynamic simulation data.

[0047] Specifically, this application reconstructs the three-dimensional topology of blood vessels based on vascular CT image data, analyzes the variation patterns of vessel wall thickness and lumen diameter, and preliminarily determines the basic elastic state of blood vessels.

[0048] Subsequently, by combining hemodynamic simulation data such as wall shear force distribution and vascular strain tensor, the vascular elastic modulus is inferred to accurately quantify the vascular elasticity level (e.g., when CAVI=9.2 in diabetic patients, it is determined to be arteriosclerosis).

[0049] The calculation of vascular elasticity parameters mainly includes indicators such as vascular elastic modulus and vascular compliance.

[0050] The Coronary Artery Atrophy Index (CAVI) is one of the standard indicators for assessing arteriosclerosis, commonly used to evaluate vascular elasticity, especially in patients with diabetes and hypertension. Calculating CAVI typically requires data such as blood flow velocity waveforms and vessel diameter. Its basic calculation formula is as follows: ;in, , These are systolic blood pressure and diastolic blood pressure, respectively. , These represent the systolic and diastolic values ​​of the vessel diameter, respectively. CAVI is commonly used to assess arteriosclerosis and vascular compliance.

[0051] The elastic modulus of blood vessels is typically calculated using the pressure-volume relationship. It is derived by measuring the stress-strain relationship of the vessel wall, combined with data on the vessel's geometry, thickness, and hemodynamics. Common methods include directly measuring the elastic properties of the vessel wall using techniques such as ultrasound elastography and magnetic resonance elastography.

[0052] The determination of the basic elasticity of blood vessels mainly adopts existing methods, which generally include the following steps: vascular imaging examinations, such as vascular CT scans and ultrasound imaging, can obtain morphological parameters of blood vessels (such as lumen diameter, vessel wall thickness, etc.).

[0053] Hemodynamic analysis, combining data such as blood flow fluctuation characteristics, flow velocity, and pressure, can infer the elastic modulus or vascular compliance of blood vessels.

[0054] The specific process of reverse-engineering the elastic modulus of blood vessels is an existing technology, which involves extracting the geometric parameters of blood vessels from vascular imaging data and then calculating the deformation of the blood vessel wall through hemodynamic simulation, which will not be elaborated here.

[0055] The dynamic correction unit utilizes an adversarial learning algorithm to respond to ST segment deviation or heart rate mutation events. During the dynamic correction phase, the adversarial learning algorithm (Dueling DQN) responds to physiological mutations. The execution logic of the dynamic correction unit includes a dual response mechanism: when a horizontal ST segment depression amplitude ≥0.1mV or atrial fibrillation (RR interval coefficient of variation >15%) is detected, the optimal timing is recalculated within 100ms (the compression inflation delay time is within 5ms after the ST segment initiation, or if the inflation point is advanced to ST+3ms), and the deflation point is synchronized to the aortic valve closure time (error <8ms). This serves as the trigger benchmark, abandoning the traditional dicrotic wave detection method and ensuring that the timing error is controlled within 8ms.

[0056] The prediction of aortic valve closure time is mainly based on the temporal characteristics of electrocardiogram signals and hemodynamic simulation data.

[0057] Based on the temporal characteristics of electrocardiogram signals, by analyzing features such as the ST segment initiation and QT interval, and combining the temporal correlation between cardiac electrical activity and mechanical activity, the time window for aortic valve closure can be preliminarily predicted.

[0058] Based on hemodynamic simulation data, and according to the distribution of wall shear force and the variation of blood flow velocity vector field, the key nodes of blood flow transition from systole to diastole are captured, and the aortic valve closure time is accurately located (prediction error <8ms).

[0059] The aortic valve closure time typically occurs after the ST segment on an electrocardiogram (ECG). By utilizing features such as the ST segment initiation and QT interval, the aortic valve closure time can be estimated. Changes in the ST segment initiation and QT interval provide crucial information about cardiac electrical activity and can indirectly infer the synchronicity of mechanical activity.

[0060] Patients with prolonged QT intervals may experience delayed aortic valve closure; therefore, changes in QT interval are related to the timing of aortic valve closure. ST segment depression or elevation (such as in acute myocardial ischemia) can cause changes in aortic valve closure time, which can be accurately predicted by dynamically monitoring ST segment changes.

[0061] By observing changes in blood flow patterns, especially variations in wall shear stress, the exact timing of aortic valve closure can be estimated. Typically, Doppler ultrasound and cardiac magnetic resonance imaging (MRI) are used to dynamically measure blood flow velocity and vessel wall stress, thereby predicting blood flow status and the aortic valve closure time.

[0062] By simulating the critical points of blood flow transition from systole to diastole and combining information such as blood velocity and pressure, the timing of aortic valve closure can be predicted more accurately. For example, by using data such as wall shear force distribution maps and vascular strain tensors, the transition point between systole and diastole can be dynamically captured, and the timing of aortic valve closure can be accurately predicted. This is a technique well-known to those skilled in the art, and therefore will not be elaborated further.

[0063] By fusing multimodal data such as electrocardiogram (ECG), hemodynamic data, and imaging data (e.g., vascular CT), a neural network model is trained to predict the aortic valve closure time. This method can handle complex time-series data and identify nonlinear features hidden in the signal.

[0064] As the triggering benchmark for the deflation of the dynamic correction unit: when the ST segment horizontal depression amplitude is detected to be ≥0.1mV or atrial fibrillation, the dynamic correction unit uses this predicted time point as a basis to synchronously adjust the timing of the balloon deflation, avoid blood backflow, and ensure precise synchronization between counterpulsation and cardiac diastole.

[0065] The long-term optimization unit associates endothelial function biomarker concentrations with a reinforcement learning reward function. During the long-term optimization phase, the reward function is updated every 24 hours based on serum biomarkers (ET-1, VEGF), and the reward function is learned through reinforcement learning. To predict efficacy, if the reward is <0.7 for 3 consecutive days (e.g., ET-1 measured value 8.2 pg / mL vs predicted value 7.9 pg / mL), the single treatment time will be extended by 15%; conversely, if the reward is >0.85 for 3 consecutive days, the total treatment course will be shortened by 5 times.

[0066] function The formula for expressing this is as follows: .

[0067] in, This represents the measured value of endothelin-1 concentration. This represents the predicted value of endothelin-1 concentration. This indicates the growth rate of vascular endothelial growth factor.

[0068] Coronary perfusion-related features (such as peak diastolic blood flow and blood flow duration) were extracted using cardiac impedance signals and photoplethysmography signals, and the values ​​before and after counterpulsation treatment were recorded.

[0069] The coronary perfusion improvement rate is calculated as follows: (mean value of coronary perfusion-related indicators after counterpulsation - mean value of coronary perfusion-related indicators before counterpulsation) / mean value of coronary perfusion-related indicators before counterpulsation × 100%.

[0070] The coronary perfusion improvement rate, as a core weight term (accounting for 0.6) in the reinforcement learning reward function, directly affects the decision-making of the long-term optimization unit: if the coronary perfusion improvement rate is high (e.g., ≥30%), the reward function value increases, and the system maintains or fine-tunes the current counterpulsation parameters.

[0071] If the coronary perfusion improvement rate is low (e.g., <15%), the reward function value is reduced, the system extends the single treatment time or adjusts the regional pressure gradient combination to optimize the coronary perfusion effect.

[0072] The specific vascular digital twin model outputs a predicted value for endothelin-1 (ET-1) concentration. The actual ET-1 concentration is then obtained via ELISA (detection limit 0.1 pg / mL) through a validation module. The ratio of the absolute value of the difference between the two values ​​to the predicted value is the prediction error. This error is used as a weighted term in the reward function (accounting for 0.3%). The smaller the error (e.g., <10%), the higher the reward function value, indicating higher accuracy in predicting endothelial function and a more suitable treatment plan for the patient's actual condition.

[0073] The validation module collects serum VEGF concentration data from patients every 24 hours and calculates the growth rate between two consecutive test results. = (Later detection value - Previous detection value) / Previous detection value × 100%.

[0074] The growth rate of vascular endothelial growth factor is used as a weighted term in the reward function (accounting for 0.1). The higher the growth rate (e.g., ≥20%), the better the endothelial repair effect, and the higher the reward function value, allowing the system to maintain the current optimization strategy. If the growth rate is too low (e.g., <5%), parameter adjustments are triggered to enhance the treatment effect.

[0075] During the initialization phase, the basic pressure parameters of the airbag and the inflation / deflation timing benchmark are automatically generated based on the vascular compliance heatmap output by the digital twin. During the mid-term phase, the adversarial regulation unit monitors pathological features such as ST segment deviation and T wave inversion in real time. Through a generator-discriminator dynamic game mechanism, a timing compensation algorithm is triggered when a sudden change in ST slope is detected to ensure that the inflation / deflation timing error is strictly controlled within 100ms. During the long-term phase, the optimization unit collects serum endothelin-1 and vascular endothelial growth factor concentration data daily and inputs them into a deep Q-network reinforcement learning agent. The treatment parameters are iteratively optimized using the improvement rate of vascular endothelial function as the reward function.

[0076] The execution verification module drives the airbag to be sequentially inflated according to the individualized counterpulsation pressure and timing scheme, monitors the counterpulsation enhancement index in real time, and feeds back the efficacy deviation to the adaptive control module to trigger parameter re-optimization. At the same time, it feeds back to the digital twin modeling module for model retraining.

[0077] The execution verification module drives the multi-channel balloon to perform sequential inflation according to the corrected timing, while simultaneously calculating the counterpulsation enhancement index in real time. If EHI < 0.15 for three consecutive times, parameter re-optimization is triggered. After each treatment course, the endothelial function improvement rate predicted by the specific vascular digital twin model is compared with the actual detection value. If the error is ≥ 15%, the specific vascular digital twin model is retrained.

[0078] The counterpulsation enhancement index, denoted as EHI, is calculated using the following formula: .

[0079] The execution verification module drives the multi-channel airbag to perform sequential inflation according to the corrected timing, and calculates the counterpulsation enhancement index EHI in real time (target value ≥0.25). If EHI <0.15 for 3 consecutive times (indicating insufficient inflation), the dynamic correction unit is immediately triggered to re-optimize the parameters. After each treatment course (36 treatments), the improvement rate of endothelial function predicted by the digital twin is compared by ELISA (detection limit 0.1 pg / mL) (e.g., predicted ET-1 decline rate 20% vs. measured 18%). When the absolute error is ≥15%, the GAN model is retrained (learning rate 0.0001, 50 iterations) to reduce the error of the updated model to below 6%.

[0080] The execution verification module drives the three-stage airbag group in a time-sharing manner through an embedded controller: initiating wave-sequence inflation and configuring regional pressure gradients in the early diastolic phase, implementing dynamic pressure compensation based on the wall shear force distribution fed back in real time by a specific vascular digital twin model during the pressure holding phase, and strictly synchronizing the aortic valve closure event at the end of diastole to achieve high-speed degassing.

[0081] Each module forms a real-time closed-loop system via gigabit industrial Ethernet, which can effectively adapt to complex cases such as coronary heart disease with atrial fibrillation and peripheral artery disease with frequent premature ventricular contractions.

[0082] In summary, the multimodal signal processing module provides high-quality input for digital twins through dynamic noise separation and feature fusion. After real-time acquisition of ECG signals (ADS1299 chip, sampling rate 1kHz) and PPG signals (MAX30102 sensor), the signals are first filtered out for EMG interference and motion artifacts through VMD decomposition (layer number K=8, penalty factor α=2000), and then diastolic key feature points are weighted and fused through a cross-modal attention mechanism. The digital twin modeling module uses a generator to combine the preprocessed signals with vascular CT images, and outputs a spatiotemporal distribution map of wall shear force based on 3D-U-Net; the discriminator updates the weights every 10 iterations based on the radial artery pressure waveform to ensure consistency between the simulation sequence and real hemodynamics. The adaptive control module generates individualized parameters based on digital twins in the initial stage (e.g., P=180mmHg for a 60-year-old patient with BSA=1.8m²). In the dynamic correction stage, it responds to sudden abnormalities such as ST-segment depression of 0.12mV, compressing the inflation delay to 5ms after the ST initiation point within 100ms. In the long-term optimization stage, it adjusts the reward function based on changes in ET-1 concentration (e.g., reward value = 0.82 when it decreases by 15%), dynamically extending or shortening the treatment course. The execution verification module verifies the inflation effect in real time through the EHI index during sequential balloon inflation (target EHI≥0.25). If the target is not met, the dynamic correction unit is triggered to recalculate the parameters. After the treatment course, the predicted endothelial function value is compared with the ELISA test results (e.g., VEGF growth rate deviation >15%), and GAN retraining is automatically initiated to adapt to the patient's vascular status evolution.

[0083] Example 2, based on Example 1, proposes an adaptive control method for external counterpulsation based on generative adversarial networks and digital twins, including the following steps: Data acquisition and processing: Real-time acquisition of patient electrocardiogram (ECG) signals, photoplethysmography (PPG) signals, cardiac impedance signals, and vascular CT image data; high-quality key hemodynamic features are extracted through variational mode decomposition and cross-modal attention fusion; Based on multimodal signal collaborative processing, the effective components of the ECG signals are reconstructed after variational mode decomposition, and morphological filtering is used to eliminate residual distortion; the PPG signals are used to construct a prediction model based on accelerometers to suppress motion artifacts; in the cross-modal fusion stage, a feature weight matrix is ​​constructed using the differential peak value of cardiac impedance as the time reference to achieve spatiotemporal alignment of key parameters.

[0084] Model Construction and Optimization: A specific vascular digital twin model is constructed to generate hemodynamic simulation data. The network weights are dynamically optimized based on the arterial pressure waveform. Specifically, in the specific vascular digital twin model construction stage, this application first simultaneously collects the patient's twelve-lead electrocardiogram, dual-source photoplethysmography, cardiac impedance, and coronary CT image data. The generator network processes the vascular images through three-dimensional convolutional layers and fuses time-domain physiological signals. The discriminator introduces the Wasserstein distance loss function to dynamically verify the matching degree between the balloon pressure waveform and the Doppler blood flow velocity data. When the discrimination accuracy drops below 55% and the wall shear force simulation error meets the standard, the specific vascular digital twin model verification is completed.

[0085] Adaptive regulation includes an initial stage, a dynamic correction stage, and a long-term optimization stage.

[0086] The initial stage involves setting counterpulsation pressure and timing parameters based on the output of a specific vascular digital twin model; the dynamic correction stage involves recalculating the inflation / deflation timing within 100ms if ST segment deviation ≥0.1mV or atrial fibrillation (RR interval variation >15%) is detected; and the long-term optimization stage involves collecting serum ET-1 and / or VEGF concentrations every 24 hours and adjusting the counterpulsation pressure and timing scheme for subsequent treatments using a reward function.

[0087] The three-stage closed-loop regulation process comprises three core components: the initial stage automatically sets differentiated pressurization strategies for calcified and healthy segments based on the vascular elastic modulus distribution map; the dynamic correction stage continuously monitors the dynamic changes of the ST segment, and triggers a high-frequency iterative game between the generator and discriminator when an abnormal slope in consecutive cycles is detected, generating a proportional compensation amount and verifying the correlation between the pressure waveform and the diastolic wave amplification; and the long-term optimization stage analyzes the changes in serum biomarker concentration gradients daily, and updates the treatment strategy table through a Q-learning algorithm to achieve the self-evolution of treatment parameters.

[0088] Execution verification: Based on the counterpulsation pressure and timing plan, the balloon is driven to perform sequential pressure according to the optimized timing, and the counterpulsation enhancement index is monitored in real time. If the counterpulsation enhancement index is detected to be lower than the set threshold for three consecutive times, the process jumps to the dynamic correction stage and re-optimizes the parameters. After each complete treatment course, the deviation between the endothelial function prediction value predicted by the specific vascular digital twin model and the actual test results is automatically compared. If the absolute error exceeds 15%, the model retraining process is initiated to adapt to the evolution of the patient's vascular status.

[0089] During the verification phase, the three-stage airbag inflation sequence is precisely initiated and the regional pressure ratio is configured before the pulse wave notch. During the pressure holding period, dynamic pressurization and duration compensation are implemented based on the wall shear force fed back in real time by the digital twin. The venting action is strictly synchronized with the prediction results of the aortic valve closure characteristic point. The embedded controller implements safety monitoring in sync, and immediately triggers emergency venting and audible and visual alarms when the regional pressure exceeds the limit.

[0090] This approach achieves dynamic mapping of vascular physiological state through a specific digital twin model of blood vessels. The three-stage regulation mechanism deeply couples instantaneous physiological response with long-term endothelial repair. Multimodal fusion technology significantly improves signal reliability in complex noise environments. Clinical validation shows that the system can increase coronary flow reserve by 0.12±0.03 and reduce the treatment interruption rate of atrial fibrillation patients to below 3.2%.

[0091] The above description is merely an optional embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the content of the present invention under the concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. An adaptive control system for external counterpulsation based on generative adversarial networks and digital twins, characterized in that, It includes an acquisition module, a multimodal signal processing module, a digital twin modeling module, an adaptive control module, and an execution verification module; the acquisition module is used to acquire electrocardiogram signals, photoplethysmography signals, cardiac impedance signals, and vascular CT image data; the multimodal signal processing module preprocesses the acquired signals based on variational mode decomposition, and fuses the multimodal features of the acquired information through a cross-modal attention fusion mechanism to extract key hemodynamic features; The digital twin modeling module constructs a specific vascular digital twin model, using key hemodynamic features, multimodal physiological signals, and patient static parameters as core inputs, and outputs hemodynamic simulation data in real time. The adaptive control module, based on hemodynamic simulation data, outputs counterpulsation pressure and timing scheme through three stages: initial parameters, dynamic correction, and long-term optimization. The execution verification module drives the balloon to be sequentially inflated according to the counterpulsation pressure and timing scheme, monitors the counterpulsation enhancement index in real time, and feeds back the efficacy deviation to the adaptive control module to trigger parameter re-optimization. At the same time, it feeds back to the digital twin modeling module for model retraining.

2. The external counterpulsation adaptive control system based on generative adversarial networks and digital twins according to claim 1, characterized in that, The multimodal signal processing module includes a variational mode decomposition unit and a cross-modal attention fusion unit; the variational mode decomposition unit is used to separate electromyographic noise and motion artifact interference from photoplethysmography signals in electrocardiogram signals; the cross-modal attention fusion unit is used to extract key hemodynamic features.

3. The external counterpulsation adaptive control system based on generative adversarial networks and digital twins according to claim 2, characterized in that, The variational mode decomposition unit is set with a decomposition layer number K=8, a penalty factor α=2000, and a center frequency constraint range of 0.5-40 Hz; The cross-modal attention fusion unit uses a weight allocation formula to dynamically enhance the importance of key hemodynamic features. The weight allocation formula normalizes the linear combination of ECG features and photoplethysmography features using the Softmax function.

4. The external counterpulsation adaptive control system based on generative adversarial networks and digital twins according to claim 1, characterized in that, The digital twin modeling module is built on a generative adversarial network architecture and includes a generator and a discriminator. The generator takes multimodal physiological signals and patient static parameters as input and outputs real-time hemodynamic simulation data of the cardiovascular system. The discriminator optimizes the model weights by comparing the dynamic differences between the real arterial pressure waveform and the simulation data.

5. The external counterpulsation adaptive control system based on generative adversarial networks and digital twins according to claim 4, characterized in that, The generator adopts a 3D-U-Net convolutional neural network structure, and its input layer integrates the temporal features of electrocardiogram signals and the three-dimensional topological data reconstructed from vascular CT images. The discriminator is based on a temporal convolutional network architecture, using the measured radial artery pressure waveform with a sampling frequency of 1000Hz as the authenticity judgment benchmark, and calculates the Wasserstein distance between the generator-generated data and the measured waveform to dynamically update the network weights.

6. The external counterpulsation adaptive control system based on generative adversarial networks and digital twins according to claim 1, characterized in that, The adaptive control module includes an initial parameter unit, a dynamic correction unit, and a long-term optimization unit; The initial parameter unit generates individualized counterpulsation pressure and timing benchmarks based on the output of a specific vascular digital twin model; the dynamic correction unit uses an adversarial learning algorithm to respond to ST segment deviation or heart rate mutation events. The long-term optimization unit associates the concentration of endothelial functional biomarkers through a reinforcement learning reward function.

7. The external counterpulsation adaptive control system based on generative adversarial networks and digital twins according to claim 6, characterized in that, When the ST segment horizontal depression amplitude is detected to be ≥0.1mV or atrial fibrillation, the dynamic correction unit automatically compresses the inflation delay time to within 5 milliseconds after the ST segment initiation; the deflation timing uses the aortic valve closure time point predicted by a specific vascular digital twin model as the trigger benchmark.

8. The external counterpulsation adaptive control system based on generative adversarial networks and digital twins according to claim 6, characterized in that, The reward function in the long-term optimization unit consists of three weighted parts: coronary perfusion improvement rate accounts for 0.6, endothelin-1 concentration prediction error term accounts for 0.3, and vascular endothelial growth factor growth rate accounts for 0.

1.

9. The external counterpulsation adaptive control system based on generative adversarial networks and digital twins according to claim 1, characterized in that, The execution verification module controls the sequential inflation action of the airbag and calculates the counterpulsation enhancement index in real time. If the counterpulsation enhancement index is detected to be lower than the set threshold for three consecutive times, the parameters are re-optimized. After each treatment course, the endothelial function improvement rate predicted by the specific vascular digital twin model is compared with the actual detection value. When the absolute error exceeds 15%, the model retraining process is started.

10. An adaptive control method for external counterpulsation based on generative adversarial networks and digital twins, characterized in that, The system applied to the external counterpulsation adaptive control system based on generative adversarial networks and digital twins as described in any one of claims 1-9 includes the following steps: data acquisition and processing: real-time acquisition of patient electrocardiogram signals, photoplethysmography signals, cardiac impedance signals and vascular CT image data, and extraction of key hemodynamic features through variational mode decomposition and cross-modal attention fusion; Model construction and optimization: Construct a blood vessel-specific digital twin model, output hemodynamic simulation data, and dynamically optimize network weights based on arterial pressure waveform; Adaptive regulation includes an initial phase, a dynamic correction phase, and a long-term optimization phase. The initial phase involves setting counterpulsation pressure and timing parameters based on the output of a specific vascular digital twin model. The dynamic correction phase involves recalculating the inflation / deflation timing within 100ms if ST segment deviation ≥0.1mV or atrial fibrillation is detected. The long-term optimization phase involves collecting serum ET-1 and / or VEGF concentrations every 24 hours and adjusting the counterpulsation pressure and timing scheme for subsequent treatments using a reward function. Execution verification involves sequentially pressurizing the balloon according to the optimized timing scheme, monitoring the counterpulsation enhancement index in real time, and if the counterpulsation enhancement index is detected to be below a set threshold three times consecutively, the system jumps to the dynamic correction phase to re-optimize the parameters. After each complete treatment cycle, the system automatically compares the deviation between the endothelial function prediction value predicted by the specific vascular digital twin model and the actual detection results. If the absolute error exceeds 15%, the model retraining process is initiated.

Images