Deep learning-based ventricular assist device rotational speed regulation method and system

Abstract

This invention belongs to the field of medical device technology, specifically relating to a method and system for regulating the rotation speed of a ventricular assist device (LVAD) based on deep learning. The deep learning-based LVAD rotation speed regulation method includes: acquiring current left ventricular pressure and pump speed information; feeding the current pressure and pump speed information into a pre-trained long short-term memory (LSM) network model; obtaining the pump speed information to be regulated through the LSM network model; and sending the pump speed information to be regulated to the LVAD. This invention enables the human LVAD monitoring module to automatically adjust the pump speed according to hemodynamic changes, allowing for adaptive control of the LVAD for individual patients.

Description

Technical Field

[0001] This invention belongs to the field of medical device technology, specifically relating to a method and system for controlling the rotation speed of a ventricular assist device based on deep learning. Background Technology

[0002] Cardiovascular and cerebrovascular diseases remain the leading threat to human health and a primary cause of death. Although heart transplantation remains the most effective treatment for heart failure, donor shortages and clinical complications (including advanced age and chemotherapy-induced heart failure) severely limit its application. Left ventricular assist devices (LVADs), mechanical pumps implanted in heart failure patients, serve as mechanical circulatory support to assist in bridging transplants, bridging recovery, and targeted therapies for the failing left ventricle. The advent of LVADs has overcome the problem of insufficient transplant donors, partially or completely replacing the function and role of the natural heart, and is one of the main measures for treating end-stage heart failure.

[0003] With the successful clinical application of LVADs, how to better assist cardiac blood supply using LVADs has become a topic of increasing interest among researchers. Currently, based on shared clinical experience in LVAD nursing, LVADs are not operated at a constant rate according to the individual's physiological state. Constant-rate operation of LVADs may lead to ventricular aspiration (ventricular collapse due to excessively low intraventricular pressure) or pulmonary congestion (a condition caused by excessive fluid in the lungs due to high ventricular pressure). Ventricular aspiration may lead to hemolysis, damage to cardiac tissue near the pump inlet, right ventricular dysfunction, or the release of ventricular thrombi and subsequent stroke, while pulmonary congestion may lead to pulmonary congestion (pulmonary edema) and shortness of breath.

[0004] Currently, the constant-rate continuous-flow mode of left ventricular assist devices lacks a physiological feedback mechanism, requiring regular monitoring by clinicians to ensure that the assist blood pressure and flow rate are suitable for the patient's physiological state. The human body is a complex time-varying system, and ventricular contractility, heart rate, arterial pressure, and venous return, which are related to ventricular load, also exhibit time-varying properties. However, the responsiveness of the constant-rate continuous-flow mode to ventricular load is limited. If the assist blood pressure or flow rate is not adjusted when the patient's circulatory status changes, it may lead to adverse events.

[0005] Previously, most researchers used fixed-gain proportional-integral-derivative (PID) controllers to implement these physiological control systems. However, optimal performance could only be achieved by tuning to specific patients and conditions. As a consequence, control performance could not be guaranteed under different patients and conditions, potentially leading to more dangerous events such as suction and congestion. In recent years, researchers both domestically and internationally have conducted research on control systems for left ventricular assist devices based on heart rate feedback, flow feedback, and pressure feedback.

[0006] Heart rate feedback adaptive flow control schemes suffer from low feedback accuracy, and the nonlinearity between heart rate and flow models further impacts control precision. While flow feedback can achieve high accuracy, its large sensor size makes implantation in patients difficult. Pressure feedback balances sensor size and feedback accuracy, but demands high real-time computational capabilities from the model. Most adaptive control strategies, such as fuzzy control, require defining rule sets, which is time-consuming and inefficient due to nonlinear behavior. Active disturbance rejection, while capable of responding to physiological needs within a certain range, has limited adaptive capabilities and struggles to handle individual patient differences and dynamic pathological states. Model predictive control, while capable of tracking reference value changes effectively, is computationally expensive and difficult to implement in real-time in practical environments. Summary of the Invention

[0007] The present invention addresses the above-mentioned technical problems by providing a method and system for controlling the rotational speed of a ventricular assist device based on deep learning.

[0008] A deep learning-based method for controlling the rotational speed of a ventricular assist device includes:

[0009] The current pressure information and current pump rate information of the left ventricle are collected, and the current pressure information and current pump rate information are sent to a pre-trained long short-term memory network model. The pump rate information to be adjusted is obtained through the long short-term memory network model, and the pump rate information to be adjusted is sent to the ventricular assist device.

[0010] Optionally, the training process of the short-term memory network model is as follows:

[0011] S1, acquire several actual pressure information and actual pump speed information of the left ventricle, send the actual pressure information, the actual pump speed information and the preset reference pressure information into the preset nonlinear model prediction controller to predict future pressure information, and calculate and obtain the pump speed information to be controlled.

[0012] S2, the actual pressure information, the reference pressure information, the actual pump speed information, the future pressure information, and the pump speed information to be adjusted are fed into the Long Short-Term Memory (LSTM) network model for training, and the trained LSM network model is obtained.

[0013] Optionally, before S1, the following also includes:

[0014] S0, Establish a coupling model between the ventricular assist device and the human cardiovascular system, and determine that the input of the coupling model is the pump speed signal of the ventricular assist device and the output is the pressure value of the left ventricle;

[0015] The prediction model of the nonlinear model predictive controller is a continuous-time nonlinear model of the coupling between the ventricular assist device and the human cardiovascular system.

[0016] Optionally, the Long Short-Term Memory network model includes an input layer, an LSTM layer, a fully connected layer, and a regression layer. Each time step of the input layer contains 40 features, and the LSTM layer has 128 hidden units.

[0017] Optionally, the long short-term memory network model uses root mean square error to evaluate the deviation between the predicted value and the reference value.

[0018] Optionally, when training the short-term memory network model, the actual pressure information and actual pump speed information under several different physiological environments are collected for training to obtain long short-term memory network models under different physiological environments.

[0019] Before sending the current pressure information and the current pump speed information into the trained Long Short-Term Memory (LSTM) network model, the LSM network model is switched to the corresponding current physiological environment.

[0020] Optionally, several different physiological conditions include increases and decreases in systemic vascular resistance (SVR), rest to exercise, and passive postural changes.

[0021] Optionally, the predicted pressure information is also obtained through the Long Short-Term Memory Network model, and the predicted pressure information and the current pressure information are used to generate a pressure change curve for real-time display.

[0022] A deep learning-based ventricular assist device speed control system includes:

[0023] A ventricular assist device, the ventricular assist device having a pump head and a pressure sensor, the pressure sensor being used to collect current pressure information of the left ventricle;

[0024] The controller has a speed control module connected to the ventricular assist device. The speed control module is used to acquire current pressure information and current pump speed information, send the current pressure information and current pump speed information into a pre-trained long short-term memory network model, obtain the pump speed information to be controlled through the long short-term memory network model, and send the pump speed information to be controlled to the ventricular assist device.

[0025] Beneficial effects: The present invention has at least one or more of the following advantages:

[0026] 1. This invention integrates a novel control method combining Nonlinear Model Predictive Control (NMPC) and Long Short-Term Memory Network (LSTM) to improve the adaptability and control accuracy of LVAD in practical clinical applications. Specifically, this invention uses a pressure sensor to collect left ventricular pressure as the control target. NMPC enables pressure tracking, and LSTM integrates the control performance of NMPC, solving the problem that NMPC consumes large computational resources and cannot achieve real-time control. Through this invention, the monitoring module of the human LVAD can automatically adjust the pump speed according to hemodynamic changes, enabling adaptive control of the LVAD for individual patients.

[0027] 2. In order to adapt the control scheme to changes within individual patients and improve the adaptability of the model, this invention aims to achieve real-time regulation of multiple dynamic pathological states by considering individual patients under different physiological environments.

[0028] 3. This invention takes into account three aspects: dynamic adaptability, real-time response capability, and individualized control strategy. Based on NMPC dynamic pressure tracking, LSTM network nonlinear modeling capability, and overall optimization design of integrated control strategy, it not only solves many problems of traditional LVAD constant speed control and meets the individualized needs of patients, but also improves the stability, safety and clinical applicability of the system. It has important theoretical value and engineering significance, and provides a new research direction for the intelligent and adaptive development of future LVAD control systems. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of a coupling model of the present invention;

[0030] Figure 2 This is a block diagram of a speed control strategy according to the present invention;

[0031] Figure 3 This is a system structure diagram of the present invention. Detailed Implementation

[0032] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, so as to better understand the purpose, features and advantages of the present invention. It should be understood that the embodiments shown in the drawings are not intended to limit the scope of the present invention, but are only for illustrating the essential spirit of the technical solution of the present invention.

[0033] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known apparatuses, structures, and techniques associated with this application may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0034] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0035] In the following description, in order to clearly demonstrate the structure and working method of the present invention, a number of directional terms will be used. However, terms such as "front", "back", "left", "right", "outside", "inside", "outward", "inward", "up", and "down" should be understood as convenient terms and not as limiting terms.

[0036] Reference Figure 1 and Figure 2 This invention provides a method and system for controlling the rotational speed of a ventricular assist device (VAD) based on deep learning. This method is used in the rotational speed control strategy of the pump head in a VAD, and specifically includes:

[0037] The system collects current pressure and pump rate information of the left ventricle, feeds the current pressure and pump rate information into a pre-trained long short-term memory network model, obtains the pump rate information to be regulated through the long short-term memory network model, and sends the pump rate information to be regulated to the ventricular assist device.

[0038] In this invention, when making predictions based on current pressure and pump speed information, a pre-trained Long Short-Term Memory (LSTM) network model is used. That is, before actual use, a preset LSTM is trained to be suitable for the speed control strategy of this invention.

[0039] The collected data, such as current pressure and pump speed information, can be preprocessed before being fed into the model, for example, by removing at least one of the following methods: noise removal and outlier removal, to ensure the robustness and accuracy of the subsequent model.

[0040] In one embodiment, the training process of the LSTM of the present invention is as follows:

[0041] S1: Acquire several actual pressure information and actual pump speed information of the left ventricle, and send the actual pressure information, actual pump speed information and preset reference pressure information into a preset nonlinear model predictive controller to predict future pressure information and calculate the pump speed information to be controlled.

[0042] S2, the actual pressure information, reference pressure information, actual pump speed information, future pressure information, and pump speed information to be regulated are fed into the Long Short-Term Memory (LSTM) network model for training, resulting in the trained LSM network model.

[0043] In this embodiment, the reference pressure information is the reference left ventricular pressure (LVP). Specifically, it is the left ventricular pressure (LVP) based on a healthy human cardiovascular system model.

[0044] In this embodiment, after obtaining the actual pressure information and actual pump speed information for training, it is preferable to perform preprocessing, such as removing at least one of noise and outliers.

[0045] This embodiment integrates a nonlinear model predictive controller (NMPC) and LSTM during LSTM training. Dynamic pressure tracking is achieved using NMPC, while LSTM enhances the controller's real-time performance. Ultimately, the trained LSTM replaces the NMPC for real-time speed control, improving the controller's real-time response capability.

[0046] Traditional LVAD control methods often employ fixed-gain PID control or fuzzy control strategies, which suffer from limited response speed and insufficient regulation capability. This invention introduces NMPC (Non-Monitored Pump Control) as a pressure tracker. Using left ventricular pressure (LVP) as the target control variable, it predicts future LVP changes based on real-time hemodynamic parameters. NMPC includes three basic features: a predictive model, finite-time rolling optimization, and feedback correction. The pump speed is dynamically adjusted through rolling optimization. By predicting pressure values ​​N steps ahead, NMPC can achieve accurate tracking of left ventricular pressure in an optimized manner.

[0047] While NMPC offers excellent dynamic control performance, its high computational cost in real-time environments makes it difficult to meet the rapid response requirements of LVAD. Therefore, this invention combines LSTM networks with NMPC to design a highly efficient controller to replace the traditional NMPC. Specific advantages include: high efficiency: The LSTM network learns the nonlinear dynamics of LVAD through historical time-series data, enabling prediction of pressure change trends and pump rate regulation with lower computational cost. Nonlinear modeling capability: The LSTM network effectively captures the nonlinear dynamic characteristics in time-series data, compensating for the shortcomings of traditional control methods in nonlinear modeling. Adaptability: By learning from training data of different patients and physiological states, the LSTM controller possesses stronger individualized adaptability, enabling precise control tailored to specific patient pathological conditions.

[0048] Traditional LVAD control methods often use a single indicator as the control target, making it difficult to consider individualized physiological differences and dynamic pathological changes. This invention proposes a fusion control strategy that acquires pressure data and combines the advantages of NMPC and LSTM to achieve individualized adaptive regulation of left ventricular pressure in patients. The strategy is characterized by: individualization, adjusting the LVAD pump rate to match the individual patient's hemodynamic characteristics based on their actual physiological state and dynamic needs; scalability, allowing the addition of other physiological indicators such as aortic pressure and flow rate as auxiliary targets to enhance the LVAD's responsiveness to complex pathological states; and clinical applicability, as the fusion control strategy effectively reduces the computational complexity of NMPC through the LSTM network, meeting real-time requirements while avoiding reliance on implantable flow sensors, thus improving the feasibility of practical clinical applications.

[0049] In one embodiment, prior to S1, the method further includes:

[0050] S0, establish a coupling model between the ventricular assist device and the human cardiovascular system, and determine that the input of the coupling model is the pump speed signal of the ventricular assist device and the output is the pressure value of the left ventricle.

[0051] Therefore, the prediction model of the nonlinear model predictive controller is a continuous-time nonlinear model of the coupling between the ventricular assist device and the human cardiovascular system.

[0052] Reference Figure 1 This is a coupled model simulating the ventricular assist device (LVAD) and the human cardiovascular system. The coupled model is constructed based on a lumped parameter model. Simulations were performed on pressure detection and the implementation of LSTM in Simulink. This coupled model is based on first-principles calculations and includes the left ventricle, systemic circulation, LVAD, relevant parameters, and state variables, as shown in Tables 1 and 2.

[0053] Table 1 Parameter Overview

[0054]

[0055]

[0056] Table 26 represents the meaning of state variables.

[0057]

[0058] According to Kirchhoff's current and voltage laws, the state equations of the coupling model between the human cardiovascular system and LVAD can be written as follows:

[0059]

[0060]

[0061] Where x represents the pressure information at time t, u(t) = ω 2 ω(t) represents pump speed information in rpm.

[0062] C(t) = 1 / E(t)

[0063] E(t) = (Emax - Emin)En(tn) + Emin

[0064]

[0065] Where Emax and Emin are constants, representing the maximum and minimum active elasticity of the left ventricle, respectively, tn = t / Tmax, Tmax = 0.2 + 0.15tc, and tc is the cardiac cycle. E(t) represents the spontaneous elasticity change of the left ventricle during this cardiac cycle. The normal human cardiac cycle is 0.8s, therefore C(t) is a time-varying function, introducing a time-varying term into the system.

[0066] The resistances in the same path as the LVAD can be linearly superimposed, and the inductance also satisfies this condition. The pressure difference ΔP of the blood pump is linearly related to the blood flow rate Q and the derivative of the blood flow rate Q′, so its coefficients β0 and β1 can also be linearly superimposed on the resistance and inductance in this path, respectively.

[0067] Rd*=R i+Ro+β0

[0068] Ld*=L i+Lo+β1

[0069] The hydrodynamic characteristics of axial flow pumps show that their properties can be expressed using pressure difference H, pump flow rate Q, and rotational speed ω. Steady-state characteristics of axial flow pumps indicate that at a certain rotational speed, the pressure difference between the pump inlet and outlet can be represented as a second- or third-order polynomial related to the pump flow. However, further research reveals a linear relationship between the pump flow rate and the square of the rotational speed, expressed by the following formula:

[0070] H=ΔP=Po-P i=b0Q+b1Q′+b2ω2

[0071] Where Pi and Po represent the inlet and outlet pressures of the pump, respectively, and the three coefficients b0, b1, and b2 are related to the structure of the pump and are preset values.

[0072] In one embodiment, the control method of the nonlinear model predictive controller is as follows:

[0073] S11, Establish a continuous-time nonlinear model of the coupling between the ventricular assist device and the human cardiovascular system, expressed as:

[0074]

[0075] Where x(t) is the state variable of the model, u(t) is the control input vector, and u(t) = ω 2 (t), where ω(t) represents pump speed information, and f is a nonlinear model function;

[0076] S12, Discretize the model as follows:

[0077] x k+1 =F(x) k ,u k )

[0078] Where F is the discrete model function obtained by discretizing f, and k is the actual time S13, finite-time rolling optimization:

[0079] Define the prediction time domain N P Let U be the future N P The control input sequence at each time step:

[0080]

[0081] The predicted sequence of system states is as follows:

[0082]

[0083] U and X are expressed by the formula x k+1 =F(x) k ,u k Perform recursive calculations;

[0084] S14, set constraints including state constraints, control constraints, and terminal constraints:

[0085]

[0086] Where, x min and x max u is the preset minimum and maximum values ​​of the state variables. min and u max y represents the preset minimum and maximum control input values. k The model output vector, y k =x k+1 y min and y max The preset minimum and maximum output values;

[0087] S15, Define the cost function and the optimization problem:

[0088] The cost function, which includes the cost of state x, the cost of control u, and the cost of prediction at the end of the time domain, is expressed as:

[0089]

[0090] Where, x ref For predicting time steps k+i or k+N P The desired reference state vector, Q is the state cost matrix used to measure state deviation, and u ref For predicting time steps k+i or k+N P The desired reference control input vector, R is the control cost matrix used to measure the size of the control input, and P is the weight matrix used to measure the performance of the model at the end of the prediction time domain.

[0091] The optimization problem is represented as:

[0092]

[0093] in, This means finding an optimal control input among all possible control inputs U such that the cost function J reaches its minimum value;

[0094] S16, Scrolling Optimization and Feedback:

[0095] Real-time optimization is achieved through a numerical solver, with control inputs applied:

[0096]

[0097] Optimal control sequence The first input in Applied to the actual model x k+1=F(x) k ,u k )superior:

[0098]

[0099] The model updates the state at real time k:

[0100]

[0101] At the next sampling time k+1, the actual state x of the model is reacquired. k+1 , with x k+1 Given a new initial state, repeat the optimization problem:

[0102]

[0103] The optimal control sequence at time k+1 is obtained:

[0104]

[0105] Optimal control sequence The first input in Applied to the actual model x k+1 =F(x) k ,u k )superior:

[0106]

[0107] The model updates the state at real time k:

[0108]

[0109] Reference Figure 2 In this example, the nonlinear model predictive controller acts as an NMPC pressure tracker. It can pre-set pressure constraints (i.e., state constraints), speed constraints (i.e., control constraints), and prediction intervals (i.e., prediction time domain N) based on the actual control effect. P The system uses the control interval (i.e., the length of the control input sequence), input weights (i.e., the control cost matrix R), and output weights (i.e., the state cost matrix Q) to predict future pressure values ​​based on actual pressure information (i.e., actual LVP) and reference pressure information (i.e., expected LVP).

[0110] The nonlinear model predictive controller in this embodiment includes three basic features: a predictive model, finite-time rolling optimization, and feedback correction, enabling accurate tracking of left ventricular pressure. Furthermore, it incorporates state and control constraints during the optimization process to prevent safety hazards such as ventricular aspiration or pulmonary congestion caused by excessively high or low pump rates.

[0111] In one embodiment, the Long Short-Term Memory (LSTM) network model includes an input layer, an LSTM layer, a fully connected layer, and a regression layer. Each time step of the input layer contains 40 features, and the LSTM layer has 128 hidden units.

[0112] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network (RNN) that addresses the long-term dependencies and vanishing gradient problems inherent in traditional RNNs by designing input, forget, and output gates. LSTM networks are one of the mainstream deep learning methods for time series data processing. LSTM networks effectively capture the long-term dependencies and nonlinear dynamic characteristics of time series data and are widely used in sequence prediction, classification, and regression tasks.

[0113] This invention uses LSTM to predict LVAD speed and future pressure based on left ventricular pressure (LVP) data. The LSTM is designed with a hierarchical structure consisting of a sequence input layer, an LSTM layer, a fully connected layer, and a regression layer, and is optimized for joint regression tasks of time series data.

[0114] In the input layer of the LSTM, the feature sequence is input in units of time steps, with each time step containing 40 features to ensure that the input signal fully represents the patient's physiological state. Within the LSTM layer, 128 hidden units are designed to selectively remember or forget important information in the time series through a gating mechanism, achieving effective modeling of long-term dependencies. The output mode of the LSTM layer is set to "sequence output" to retain the hidden state features of all time steps and use them for joint prediction of future pressure values ​​and pump rates. Following the LSTM layer, the network employs a fully connected layer to map the time series features extracted by the LSTM to the target predictor variables: a regression prediction of pressure for the next N steps and a univariate prediction of LVAD pump rate. Finally, a regression layer is used to calculate the errors between the predicted and actual values ​​of pressure and pump rate for the next N steps, and to optimize the model weights.

[0115] Compared to traditional recurrent neural networks (RNNs), LSTM networks effectively mitigate the vanishing and exploding gradient problems in time series data by introducing a gating mechanism, while also improving the ability to model nonlinearities in dynamically changing systems. Furthermore, LSTM networks optimize the number of hidden layer units and the selection of output modes, making them more suitable for regression modeling of time series data.

[0116] In one embodiment, the Long Short-Term Memory (LSTM) network model uses the Root Mean Square Error (RMSE) as the core metric to evaluate the deviation between the predicted value and the reference value. Its calculation formula is as follows:

[0117]

[0118] in, y is the model's predicted value. i denoted as the true value, and n as the number of samples. RMSE directly reflects the accuracy of the model's predictive performance; a smaller RMSE value indicates higher predictive accuracy.

[0119] In one embodiment, when training the short-term memory network model, actual pressure information and actual pump rate information under several different physiological environments are collected for training to obtain long short-term memory network models under different physiological environments. Before feeding the current pressure information and current pump rate information into the trained long short-term memory network model, the long short-term memory network model is switched to the corresponding current physiological environment.

[0120] In this embodiment, several training data sets under different physiological environments are collected during LSTM training. Each set of training data is then trained separately to obtain an LSTM controller capable of real-time response to different physiological environments. During actual rotational speed adjustment, the patient's current physiological environment is switched to the corresponding LSTM. This embodiment, by considering different physiological states (such as systemic vascular resistance, changes in body position, etc.), achieves real-time control of multiple dynamic pathological states, overcoming the difficulty in meeting diverse physiological needs in real time when dealing with the dynamic physiological states of individual patients.

[0121] In one embodiment, several different physiological environments include increases and decreases in systemic vascular resistance (SVR), rest to exercise, and passive postural changes.

[0122] In one embodiment, predicted pressure information is also obtained through a long short-term memory network model, and the predicted pressure information and the current pressure information are used to generate a pressure change curve for real-time display.

[0123] Reference Figure 3 This invention also provides a deep learning-based ventricular assist device speed control system, comprising:

[0124] The ventricular assist device has a pump head 10 and a pressure sensor 20, which is used to collect the current pressure information of the left ventricle.

[0125] The controller 30 has a speed control module connected to the ventricular assist device. The speed control module is used to acquire current pressure information and current pump speed information, and sends the current pressure information and current pump speed information into a pre-trained long short-term memory network model. The pump speed information to be controlled is obtained through the long short-term memory network model and then sent to the ventricular assist device.

[0126] The pressure monitoring end 21 of the pressure sensor 20 is located on the side near the pump head 10 to facilitate real-time monitoring of blood flow pressure data within the patient's body. Figure 3 As shown, the pressure monitoring end 21 is located near the pump head 10.

[0127] The pump head 10, pressure sensor 20, and other structures of the ventricular assist device are existing technologies and will not be described in detail here.

[0128] The preferred embodiments of the present invention have been described in detail above. However, it should be understood that after reading the above teachings, those skilled in the art can make various alterations or modifications to the present invention. These equivalent forms also fall within the scope defined by the appended claims.

Claims

1. A deep learning-based ventricular assist device speed control system, characterized in that, include: A ventricular assist device, the ventricular assist device having a pump head and a pressure sensor, the pressure sensor being used to collect current pressure information of the left ventricle; The controller has a speed control module connected to the ventricular assist device. The speed control module is used to acquire current pressure information and current pump speed information, send the current pressure information and the current pump speed information into a pre-trained long short-term memory network model, obtain the pump speed information to be controlled through the long short-term memory network model, and send the pump speed information to be controlled to the ventricular assist device. The training process of the short-time memory network model is as follows: S0, Establish a coupling model between the ventricular assist device and the human cardiovascular system, and determine that the input of the coupling model is the pump speed signal of the ventricular assist device and the output is the pressure value of the left ventricle; S1, acquire several actual pressure information and actual pump speed information of the left ventricle, send the actual pressure information, the actual pump speed information and the preset reference pressure information into the preset nonlinear model prediction controller to predict future pressure information, and calculate and obtain the pump speed information to be controlled. S2, the actual pressure information, the reference pressure information, the actual pump speed information, the future pressure information, and the pump speed information to be adjusted are fed into the Long Short-Term Memory (LSTM) network model for training, and the trained LSM network model is obtained.

2. The ventricular assist device speed control system based on deep learning as described in claim 1, characterized in that, The prediction model of the nonlinear model prediction controller is a continuous-time nonlinear model that couples the ventricular assist device with the human cardiovascular system after discretization.

3. The ventricular assist device speed control system based on deep learning as described in claim 1, characterized in that, The Long Short-Term Memory (LSTM) network model includes an input layer, an LSTM layer, a fully connected layer, and a regression layer. Each time step of the input layer contains 40 features, and the LSTM layer has 128 hidden units.

4. The ventricular assist device speed control system based on deep learning as described in claim 1, characterized in that, The Long Short-Term Memory (LSTM) network model uses root mean square error to evaluate the deviation between the predicted value and the reference value.

5. The ventricular assist device speed control system based on deep learning as described in claim 1, characterized in that, When training the short-term memory network model, the actual pressure information and actual pump speed information under several different physiological environments are collected for training to obtain long short-term memory network models under different physiological environments. Before sending the current pressure information and the current pump speed information into the trained Long Short-Term Memory (LSTM) network model, the LSM network model is switched to the corresponding current physiological environment.

6. The ventricular assist device speed control system based on deep learning as described in claim 5, characterized in that, Several different physiological environments include increases and decreases in systemic vascular resistance, the transition from rest to exercise, and passive postural changes.

7. The ventricular assist device speed control system based on deep learning as described in claim 1, characterized in that, The Long Short-Term Memory Network model also obtains predicted pressure information, and the predicted pressure information and the current pressure information are used to generate a pressure change curve for real-time display.

Images