A ventricular assist device control method based on a kinetic model and deep learning

Abstract

The application discloses a ventricular assist device control method based on a kinetic model and deep learning, and belongs to the technical field of computer systems based on specific calculation models. The method fuses a blood flow dynamics calculation model based on knowledge and deep learning technology, and operates under a double-time-scale loop. In each control cycle, multi-source observation data of the device is acquired, a deep learning model is used to generate a running safety probability and a lower confidence limit thereof, a dynamic safety threshold derived from a blood flow dynamics model is combined, a constraint satisfaction risk is comprehensively evaluated, and a regulation strategy is automatically selected based on a running state: in a normal state, a pump speed is kept stable, in a warning state, a small-step speed regulation is performed, and in an emergency state, a model predictive control is triggered to generate a pump speed adjustment amount. The method realizes a fusion calculation logic of the kinetic model and the deep learning technology, and improves adaptability, safety and multi-objective coordinated control capability of device running.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer systems based on specific calculation models, and relates to a ventricular assist device control method based on a dynamic model and deep learning. BACKGROUND

[0002] A left ventricular assist device (LVAD) is an important device for treating end-stage heart failure, which mechanically assists or replaces the heart pumping function. With the popularization of LVAD application, how to intelligently control the device operation according to the real-time state of the patient and avoid frequent manual intervention has become a research hotspot.

[0003] At present, there are various control methods for LVAD, but there are still problems such as insufficient personalized diagnosis linkage control, limited control target coordination, and improved optimization flexibility. SUMMARY

[0004] To solve the above problems, the application provides a ventricular assist device control method based on a dynamic model and deep learning.

[0005] The technical scheme for achieving the purpose of the application is as follows:

[0006] A ventricular assist device control method based on a dynamic model and deep learning, the method comprising:

[0007] S1, constructing a first deep learning model within a set control period, and obtaining a first running state of the ventricular assist device, wherein the first running state includes normal, warning, and emergency;

[0008] S2, based on the running state obtained in S1, controlling the ventricular assist device: when the running state is normal, maintaining the current pump speed unchanged; when the running state is warning, performing a small step speed control strategy according to a set step increment; when the running state is emergency, performing a model predictive control strategy.

[0009] As a further optimization scheme of the application, the first deep learning model includes two one-dimensional convolution layers in cascade, two bidirectional gated recurrent units, and an output layer.

[0010] As a further optimization scheme of the application, the small step speed control strategy in step S2 specifically includes:

[0011] The pump speed is adjusted according to the following update formula:

[0012] ,

[0013] Wherein, N t represents the pump speed at the t minute, Nt+1 Nt+1 t is the step increment at the tth minute;

[0014] The target pump speed and the step increment both satisfy the set pump speed regulation constraint, and the adjustment direction follows the following principles:

[0015] If or , the pump speed is reduced, that is, ΔNt+1 t is less than 0; wherein, is the minimum left ventricular pressure margin, is the alarm zone suction margin threshold, is the pump power-speed residual, is the alarm zone power-speed standardized residual threshold; otherwise, when and , the pump speed is increased, that is, ΔNt+1 t is greater than 0; wherein, is the alarm zone reference mean pressure down capacity difference, is the alarm zone pulse pressure margin threshold, is the reference mean arterial pressure at the tth minute, is the nominal reference mean arterial pressure at the tth minute, is the reference offset at the tth minute, is the mean arterial pressure safety lower confidence limit at the tth minute, is the pulse pressure margin at the tth minute.

[0016] As a further optimization scheme of the present application, the model predictive control strategy in S3 specifically includes:

[0017] Set the prediction step size L , mark the prediction step index as k = 0, …, L-1, and adjust the pump speed at each prediction step according to the following update formula:

[0018] ,

[0019] wherein, represents the pump speed at the tth minute and the kth prediction step, represents the pump speed at the tth minute and the k+1th prediction step, is the step increment at the tth minute and the kth prediction step;

[0020] The target pump speed and the step increment both satisfy the set pump speed regulation constraint, and is obtained by selecting the sequence that minimizes the objective function from the set feasible set that satisfies the physiological safety constraint, and based on the in the obtained sequence, the pump speed at the tth minute is adjusted; the objective function at the tth minute is:

[0021] ,

[0022] wherein, , is the cost weight of the tth minute, is the mean arterial pressure of the kth prediction step of the tth minute, is the step increment of the k-1th prediction step of the tth minute.

[0023] As a further optimization scheme of the present application, after step S1 and before step S2, the method further comprises:

[0024] constructing a second deep learning model within a set control period to obtain a second operating state of the ventricular assist device, wherein the second operating state includes two kinds of normal and early warning, and the obtaining frequency of the second operating state is less than the obtaining frequency of the first operating state;

[0025] Based on the obtained second operating state, the nominal reference mean arterial pressure, the reference offset, the pump speed control constraint and the cost weight are cooperatively set, and then used for the small step speed control and model predictive control in step S2.

[0026] As a further optimization scheme of the present application, the second deep learning model includes a cascaded down-sampling layer, a one-dimensional convolution layer, two gated recurrent units and an output layer.

[0027] As a further optimization scheme of the present application, the cooperative setting includes:

[0028] (1) If the second operating state is normal, the nominal reference mean arterial pressure remains unchanged, otherwise the setting formula of the nominal reference mean arterial pressure is:

[0029] ,

[0030] wherein, is the nominal reference mean arterial pressure, is the average value of the historical measured mean arterial pressure when the first operating state is normal; and are the preset minimum and maximum mean arterial pressures, respectively;

[0031] (2) If the second operating state in two consecutive control periods is normal, the nominal reference mean arterial pressure, the reference offset, the pump speed control constraint and the cost weight are selected from a preset normal parameter table, otherwise the nominal reference mean arterial pressure, the reference offset, the pump speed control constraint and the cost weight are selected from a preset conservative parameter table; wherein the normal parameter table and the conservative parameter table are both preset according to expert knowledge.

[0032] As a further optimization of the present invention, based on the confidence lower limit of the output of the first deep learning model within the control cycle, and combined with the set safety constraints, the first operating state of the ventricular assist device is determined, wherein the safety constraints include mean arterial pressure not lower than the corresponding mean arterial pressure threshold, pulse pressure not lower than the corresponding pulse pressure threshold, and left ventricular pressure not lower than the corresponding left ventricular pressure threshold.

[0033] If any safety constraint exceeds its limit or the minimum confidence lower limit output by the first deep learning model is less than the preset emergency threshold, the first operating state is emergency; if all safety constraints do not exceed their limits, and the minimum confidence lower limit output by the first deep learning model is greater than or equal to the preset emergency threshold and less than the preset normal threshold, the first operating state is warning; if all safety constraints do not exceed their limits, and the minimum confidence lower limit output by the first deep learning model is greater than or equal to the preset normal threshold, the first operating state is normal.

[0034] As a further optimization of the present invention, the mean arterial pressure threshold, pulse pressure threshold, and left ventricular pressure threshold are dynamically obtained by constructing the following individualized hemodynamic model:

[0035] ,

[0036] ,

[0037] ,

[0038] ,

[0039] ,

[0040] in, Let t be the left ventricular pressure at minute t. Let Vt be the left ventricular volume at minute t, and V0 be the stress-free volume. E represents the time-varying elasticity of the left ventricle at minute t. min With E max These are the elastic thresholds at end-diastole and peak systole, respectively. Let be the normalized activation momentum function aligned with the heart rate phase at minute t. Let P be the aortic pressure at minute t. v This is the baseline venous pressure. Let t be the total flow rate entering the aorta at minute t. The native left ventricular ejection flow at minute t. Let be the pump flow rate at minute t, R be the peripheral resistance, C be the arterial compliance, and N(t) be the pump rate at minute t. is the pressure difference across the pump for the tth minute, f(.) is a function that maps pump speed and pressure difference across the pump to flow rate, is a vector of pump parameters that can be calibrated;

[0041] In a set cardiac cycle, simulation is performed by using the individualized hemodynamic model to obtain the minimum simulated mean arterial pressure, the minimum simulated pulse pressure and the minimum simulated left ventricular pressure in the simulation results;

[0042] Finally, the mean arterial pressure threshold, the pulse pressure threshold and the left ventricular pressure threshold are obtained:

[0043] ,

[0044] ,

[0045] ,

[0046] wherein, are the mean arterial pressure threshold, the pulse pressure threshold and the left ventricular pressure threshold for the tth minute respectively, are the lower limit of mean arterial pressure, the lower limit of pulse pressure and the left ventricular pressure threshold for the tth minute respectively, are the minimum simulated mean arterial pressure, the minimum simulated pulse pressure and the minimum simulated left ventricular pressure for the tth minute respectively.

[0047] As a further optimization scheme of the present application, the objective function iteratively updates the parameter vector of the individualized hemodynamic model :

[0048] ,

[0049] wherein, is the mean arterial pressure for the tth minute calculated by the individualized hemodynamic model at θ, is the measured mean arterial pressure for the tth minute, is the nominal reference mean arterial pressure for the tth minute.

[0050] The present application also provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the ventricular assist device control method.

[0051] The present application also provides an electronic device, comprising:

[0052] a memory for storing a computer program;

[0053] a processor for executing the computer program to implement the steps of the ventricular assist device control method.

[0054] The present application has the following advantages compared with the prior art:

[0055] The present application effectively fuses a deep learning model and an individualized hemodynamic model, and constructs a double-time-scale control loop, which can realize adaptive control of ventricular assist device operation, constraint feasible region judgment and multi-objective coordinated optimization, thereby improving responsiveness and energy efficiency under the premise of ensuring operation safety.

[0056] The present application will be further described in detail below in combination with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is the overall flow chart of the embodiment of the present application.

[0058] Figure 2 is a deep learning model structure diagram in the double-time-scale control loop.

[0059] Figure 3 is a deep learning model training data set construction flow chart.

[0060] Figure 4 is a minute-level control loop flow chart.

[0061] Figure 5 is a hour-level control loop flow chart. DETAILED DESCRIPTION

[0062] As shown in Figure 1 , the present application is a ventricular assist device control method based on a dynamic model and deep learning, which includes the following steps:

[0063] S1, double-time-scale control loop construction: in this embodiment, through the synergistic effect of the minute-level control loop and the hour-level control loop, real-time operation safety and long-term control setting of the ventricular assist device are realized.

[0064] S11, minute-level control loop construction:

[0065] S111, minute-level deep learning model construction: as shown in Figure 2 , two layers of one-dimensional convolution (1D-CNN, containing BatchNorm and Dropout) are used to extract high-frequency and transient features, then 2 layers of bidirectional gated recurrent unit (BiGRU) are connected to model short-time dependence, and MC-Dropout sampling is performed with Dropout before reading out to form uncertainty estimation; the output layer maps the time series representation to a multi-time-domain operation safety probability . In this example, the minute-level control loop selects , outputs and the corresponding 5% confidence lower bound , updated every minute.

[0066] The number of layers and dimensions are implementation examples, and equivalent replacements and fine-tuning can be made without changing the technical effects.

[0067] S112, minute-level operation state judgment and strategy triggering: obtain the 5% confidence lower bound of the minute-level operation safety probability every minute , and define the minimum value as the minute-level overall confidence lower bound:

[0068] ,

[0069] and hierarchical judgment is performed in combination with physiological safety constraints, as shown in Figure 2

[0070] Emergency: when any physiological quantity does not meet its threshold value (for example , , ), or , the minute-level operation state is emergency, and the simplified model predictive control MPC-lite strategy is triggered;

[0071] Warning: when and the above physiological quantities all meet the threshold value, the minute-level operation state is warning, and the small-step speed control strategy is triggered to control the pump speed within the amplitude constraint with a fixed step size.

[0072] Normal: when and the physiological quantities all meet the threshold value, the minute-level operation state is normal, and continuous monitoring and recording are maintained in the hour-level adjusted statistics.

[0073] S113, small-step speed control strategy:

[0074] When in the warning state, small-step speed control is performed, and the step size increment ΔN of small-step speed control is a constant value, and the absolute value is in the range of [30, 60]. The pump speed update formula is:

[0075] ,

[0076] wherein, N t represents the pump speed of the tthminute, and the unit is rpm (revolutions per minute), which represents the current actual running pump speed. N t+1 represents the target pump speed of the t+1thminute, and the unit is rpm. After the pump speed is updated, the new pump speed N t+1 must satisfy the speed limit constraint set by the device:

[0077] ,

[0078] ​where N min represents the minimum limit of pump speed, N max represents the maximum limit of pump speed.

[0079] The adjustment direction of the small cloth speed regulation follows the principle of "suction priority, perfusion second":

[0080] Suction indication priority speed reduction: if , or the pump power-speed residual error , then reduce the pump speed. Wherein is the warning zone suction margin threshold, preferably 2 mmHg; is the warning zone power-speed standardized residual error threshold, preferably 2.

[0081] Low perfusion indication and no suction indication speed up: if , and , then increase the pump speed. Wherein is the reference average pressure under the warning zone exploration capacity difference, preferably 5 mmHg; is the warning zone pulse pressure margin threshold, preferably 3 mmHg; the reference mean arterial pressure calculation formula is as follows:

[0082] ,

[0083] wherein is the nominal reference mean arterial pressure, which is set by the hour-level control loop; ∆MAP is the reference offset based on the confidence limit of the minute-level operation safety probability, which is used to adapt the reference value to the working condition under different threshold zones.

[0084] S114, MPC-lite strategy:

[0085] (1) Prediction and constraint:

[0086] Let the prediction step size be L (3-6 steps, 1 minute per step), the prediction step index is denoted as k=0,…,L-1, and the pump speed update formula of the prediction step is:

[0087] .

[0088] The pump speed regulation constraint is:

[0089] ,

[0090] Wherein the pump speed regulation constraint is set by the hour-level control loop.

[0091] For each prediction step k, calculate using the individualized hemodynamic model (PHM). For all k, apply the control safety constraint ( ).

[0092] (2) Construct a candidate step set Φ to form a feasible set satisfying the physiological safety constraints. In this example, the construction of the candidate step set Φ is gated by the emergency type:

[0093] Emergency suction: or pump power-rotation speed residual error then wherein is the power-rotation speed standardized residual emergency threshold, preferably 2.5.

[0094] Emergency low perfusion (and no suction risk): if and then wherein is the mean pressure emergency drop tolerance, preferably 10 mmHg; is the pulse pressure emergency margin threshold, preferably 1 mmHg.

[0095] Others: .

[0096] If "suction" and "low perfusion" are triggered at the same time, handle according to suction priority.

[0097] In this example, within a short prediction window of prediction step length L , a rotation speed change value is selected from the step candidate set for each prediction step, and the selected values for each prediction step are sequentially arranged and combined to form a plurality of candidate minute-by-minute pump speed adjustment sequences, forming a feasible set. For example, L = 3, the candidate step set is {-200, -100, -50, 0}, and the feasible set includes sequences {-200, -200, -200}, {-100, -200, -200}, {-50, -200, -200}, {0, -200, -200}, {-200, -100, -100}, {-200, -50, -200}, {-200, 0, -200}, etc.

[0098] (3) Construct the following objective function:

[0099] wherein, , is the cost weight of the tthminute, is the reference mean arterial pressure of the tthminute, is the mean arterial pressure of the kthprediction step of the tthminute, is the step increment of the k-1thprediction step of the tthminute. In this example, is a function of as a variable, and as the risk increases (emergency zone > alert zone > normal zone), respectively, are action energy consumption weight and smoothness weight. The gear and value of the cost weight are obtained by the hour-level control loop tuning. Select the sequence that makes the target function minimum on the feasible set, and based on the ∆N t Pump speed N t is adjusted.

[0100] S12, minute-level control loop construction:

[0101] S121, hour-level deep learning model construction: The hour-level deep learning model faces long-term trends. First, it is down-sampled by 2 times (AveragePool1d) and cooperates with one-dimensional convolution smoothing and noise suppression, and then it is connected with 2 layers of GRU to depict cross-hour dependence. The reading stage also uses MC-Dropout to obtain uncertainty. The output layer maps the time series representation to the multi-time domain running safety probability In this example, the hour-level control loop selects , outputs and the corresponding 5% lower confidence bound , which is updated every 30 minutes.

[0102] The above number of layers and dimensions are implementation examples, and equivalent replacement and fine-tuning are allowed without changing the technical effects.

[0103] S122, hour-level running state determination: Obtain the hour-level running safety probability and its lower confidence bound every 30 minutes , and take the minimum value as the hour-level lower confidence bound:

[0104] ,

[0105] Accordingly, the hierarchical determination is performed:

[0106] When , the hour-level running state is warning;

[0107] When , the hour-level running state is normal.

[0108] Determine the gear and issue parameters to the minute-level control loop:

[0109] When , the hour-level running state is warning, and the conservative gear is switched to;

[0110] When , and stable for two consecutive control periods, the hour-level running state is normal, and the normal gear is switched to.

[0111] S123, tuning strategy of the hour-level control loop:

[0112] For example Figure 3As shown, the control period is 30 minutes in this example, and the operating reference and reference offset, pump speed control constraints, and cost weight are coordinated and set based on the hourly operating state, and are sent to the minute-level control loop. The operating reference and reference offset are , the pump speed control constraints include the minute-level step and the slope upper limit, and the cost weight includes .

[0113] S1231, operating reference setting:

[0114] ,

[0115] wherein is the measured average of 1-minute granularity samples determined as “normal” by the minute-level operating state in the past 2 hours; and are the planned minimum and maximum mean arterial pressure, respectively, which are directly given by experts.

[0116] S1232, reference offset, pump speed control constraint, and cost weight setting:

[0117] To adapt to different working conditions, two sets of parameter tables are defined, namely the normal mode and the conservative mode. The hour-level control loop switches between the two modes according to the operating state, and sends the corresponding parameter set to the minute-level control loop, so that a more conservative control configuration is used when the medium and long-term operating safety probability is reduced.

[0118] Switching rule: if , enter the conservative mode; when and stable for two consecutive periods, switch back to the normal mode.

[0119] Content and value examples of sending:

[0120] 1. Reference offset

[0121] Normal mode: , normal reference offset , unit: mmHg, (such as 0 / 5 / 10 mmHg corresponding to normal / alarm / emergency zone);

[0122] Conservative mode: , conservative reference offset , unit: mmHg, (such as 5 / 10 / 15 mmHg corresponding to normal / alarm / emergency zone).

[0123] Among them, a preferred set of parameters is 0, 5, 10 mmHg in turn; 5, 10, 15, which can be replaced equivalently within the range of 1-10, 1-15, 1-20, according to the patient / model parameterization setting by one skilled in the art.

[0124] 2. Cost weight (for minute-level objective function)

[0125] Normal profile: , , , , respectively, the normal MAP deviation weight, the normal action energy consumption weight, and the normal smoothness weight. The values of the weights can be normalized and calibrated by historical data, so that each term in the equation (1) is of the same order of magnitude. A preferred set of parameters is 1.0 / 1.5 / 2.0 (normal / early warning / emergency), 0.02, 0.05.

[0126] Conservative profile: , , , respectively, the conservative MAP deviation weight, the conservative action energy consumption weight, and the conservative smoothness weight. In the equation (2), the values of the weights are , , . The conservative profile weights are increased relative to the normal profile to increase the degree of punishment for deviation in a risk state.

[0127] 3. Pump speed regulation constraints

[0128] The specific values of the upper limit of pump speed step size and the upper limit of pump speed change rate can be adjusted according to the pump model, tolerance, and device specifications.

[0129] General range: , unit rpm, , unit rpm / min.

[0130] Preferably: normal profile , ;

[0131] Conservative profile , .

[0132] In the equations (3) and (4), respectively, the normal upper limit of pump speed step size, the normal upper limit of pump speed change rate, the conservative upper limit of pump speed step size, and the conservative upper limit of pump speed change rate. S124, setting frequency and triggering:

[0133] The hour-level control loop executes the above setting with a period of 30 minutes. If any of the following events occurs at minute t, the setting program is immediately started:

[0134]

[0135] ​​​1. Low frequency key inputs (e.g. lactate, NYHA, etc. records) refresh;

[0136] 2. Device parameters re-calibration (e.g. N min , N max adjustment);

[0137] 3. Emergency / MPC-lite treatment end (from emergency state to non-emergency state).

[0138] S2, Individualized hemodynamic model construction: Construct a hemodynamic model for depicting the coupling relationship between the ventricular assist device (VAD) and the circulatory system, which is used for individualized parameter configuration, running safety threshold calculation and control feasible region evaluation, and provides feasibility and constraint boundary for minute / hour level control loop. This embodiment takes left ventricular assist device (LVAD) as an example, and the model adopts the general structure of "left ventricular time-varying elastic cavity + body circulation Windkessel periphery + pump characteristic mapping".

[0139] S21, Hemodynamic model construction:

[0140] The relationship of left ventricular pressure-volume is written as:

[0141] ,

[0142] ,

[0143] Wherein, is the left ventricular pressure, is the left ventricular volume, V0 is the unstressed volume, is the left ventricular time-varying elasticity, E min and E max are the elastic boundary values, is the normalized excitation amplitude function aligned with the heart rate phase.

[0144] Binary Windkessel peripheral capacitive equation:

[0145] ,

[0146] Wherein, is the aortic pressure, P v is the venous reference pressure, is the total flow into the aorta, and satisfies is the native left ventricular ejection flow, is the LVAD pump flow (simulation variable), R is the peripheral resistance, and C is the arterial compliance.

[0147] The pump characteristic adopts a differentiable mapping family calibrated by pump bench data:

[0148] ,

[0149] ,

[0150] where N(t) is the pump speed; is the pressure difference across the pump; the function f(.) maps the pump speed and pressure difference to the flow rate ; is the vector of pump parameters to be calibrated.

[0151] The above equations form a closed loop system, which is solved by numerical integration method. The simulation time is set to 3-5 cardiac cycles, and the model outputs include MAP, PP, and left ventricular pressure curve .

[0152] S22, individualized parameter configuration:

[0153] S221, individualized parameter calculation: To reflect individual differences and ensure the consistency of the model for control calculation, a two-stage parameter configuration method of offline range setting combined with online inversion update is adopted. In the offline stage, the initial feasible interval of the parameters is set according to statistical data: peripheral resistance R ∈ [0.8, 1.6], unit mmHg·s / mL, arterial compliance C ∈ [0.8, 2.0], unit mL / mmHg, upper limit of left ventricular elasticity E max ∈ [0.8, 1.6], unit mmHg / mL, lower limit E min ∈ [0.05, 0.3], unit mmHg / mL. In the deployment stage, the parameters are calibrated by online observation data, and the parameter vector is updated by minimizing the objective function :

[0154] ,

[0155] where, is the mean arterial pressure calculated by the model under the parameter θ, is the measured mean arterial pressure at t minutes, is the nominal reference mean arterial pressure. The optimization can use the least squares method or the gradient method with constraints, and after each update, θ is clipped back to the above offline feasible interval to ensure physical feasibility; when the target function decreases by less than a threshold, the last time parameter is kept to improve stability.

[0156] S222, individualized parameter update frequency:

[0157] By default, online inversion update is performed every 5 minutes. If any of the following control-related events occurs, an additional update is triggered immediately: 1. Pump speed N ​Adjustment; 2. Minute-level operating state changes. 3. Low-frequency key inputs (e.g. lactate, NYHA) refresh. If the period data is insufficient or the optimization does not converge, the parameters of the last moment are temporarily used: .

[0158] S23, Physiological safety constraint configuration:

[0159] S231, Physiological safety constraint calculation:

[0160] To define the feasible region in the control calculation, first give an analytical conservative lower bound from the steady-state relationship, and then correct it in the current pump speed neighborhood by short window simulation to form three types of dynamic control thresholds: mean arterial pressure dynamic threshold , pulse pressure dynamic threshold and left ventricular pressure dynamic threshold .

[0161] Analytical conservative lower bound: the mean arterial pressure lower bound is given by the two-element Windkessel steady-state approximation:

[0162] ,

[0163] where P v is the venous reference pressure, R(t) is the online estimated peripheral resistance, CI min is the minimum cardiac index preset by the schedule, and BSA is the body surface area. The pulse pressure lower bound PP floor and the left ventricular pressure threshold are given by the schedule / configuration item and are considered constant within the configuration interval.

[0164] Short window simulation correction: in the current pump speed N t neighborhood, a finite candidate rotation speed set is constructed, and a conservative parameter combination is used for 3-5 cardiac cycle individualized hemodynamic model (PHM) short window simulation; after eliminating any candidate that violates the constraints, the minimum mean pressure, minimum pulse pressure and minimum left ventricular pressure in the simulation set are recorded to obtain:

[0165] ,

[0166] ,

[0167] ,

[0168] The subscript "sim,min" in the formula represents the minimum value of the above simulation set. The three types of dynamic thresholds obtained are used for the feasible region determination and MPC-lite prediction constraints of the minute-level / hour-level control.

[0169] S232, physiological safety threshold update frequency:

[0170] Updated every 1 hour by default. If any of the following control-related events occurs, an immediate update is triggered: 1. Pump speed adjustment; 2. Minute-level running state switching; 3. Low-frequency key input (e.g. lactate, NYHA) refresh.

[0171] S3, Deep learning module training dataset construction: As shown in Figure 5 , multi-source data is collected from the device side and the patient side, time synchronization, quality control, and multi-scale statistical feature calculation are completed, and the input tensor for the model is formed.

[0172] S31, Data acquisition and preprocessing: Collecting high-frequency time series (device / physiological running data) and low-frequency asynchronous (running context / configuration data) two types of data.

[0173] S311, High-frequency data acquisition and feature engineering:

[0174] Read heart rate (HR, unit: beats / minute), mean arterial pressure (MAP, unit: mmHg), oxygen saturation (SpO2, unit: %), and pump speed (N, unit: rpm), pump power (Power, unit: W) five types of high-frequency data with 1 minute as the step, build minute-level feature flow for control minute-level deep learning model. And, to reduce the information loss brought by 1 minute aggregation, calculate the statistical features of 3 minutes and 5 minutes rolling window on each 1 minute time step in parallel, including rolling mean, variance, slope, extreme value (minimum / maximum), and generate 2 items of derived features HRV and 1 item of pump power-speed residual. Thus, the high-frequency part contains 9 items (original value + two windows x 4 statistics) x 5 types = 45 items per type of signal; plus HRV 2 items, power-speed residual 1 item, and 5 missing indicators, a total of 53 feature channels. These features will be used as input quantities and evidence for constraint evaluation to form the running safety probability and its lower confidence limit.

[0175] S312, Low-frequency data acquisition and preprocessing:

[0176] The low-frequency structured information read includes lactate (Lactate, unit: mmol / L), cardiac function classification (NYHA), drug compliance, postoperative days, etc. All low-frequency data are aligned to the minute-level time axis according to the timestamp, synchronized with high-frequency signals using step-hold, and generate time Δt (days) and missing flag features since the last update. Specifically, step-hold uses a zero-order hold / forward filling strategy, on the minute-level time grid, using the latest measurement value no later than the time step as the current value, without extrapolation or backfilling.

[0177] For lactate data, the current lactate value, time since last update (days), lactate data staleness flag (1 if not updated for >7 days), lactate elevation flag (1 if >2.0 mmol / L), and lactate absence flag are generated.

[0178] For NYHA, NYHA four-level one-hot encoding (I / II / III / IV), time since last assessment (days), NYHA data staleness flag (1 if not updated for >7 days), worsening since last flag (1 if worse than last classification), improvement since last flag (1 if better), and absence flag are generated.

[0179] For medication adherence data, 1-day adherence (0-1), 7-day average adherence (0-1), time since last adherence record (days), adherence data staleness flag (1 if not updated for >7 days), and adherence absence flag are recorded.

[0180] For postoperative day data, a three-level one-hot encoding of postoperative time period (0-7 days / 8-30 days / >30 days) is provided. After low-frequency expansion, there are 5 lactate-related features, 9 NYHA-related features, 5 medication adherence features, 3 postoperative day features, and a total of 22 feature channels.

[0181] S32, Model input specification:

[0182] The multi-source operational data is used as the model input, including the high-frequency continuous signals and low-frequency asynchronous signals after the above preprocessing, totaling 53 (high-frequency) + 22 (low-frequency) = 75 feature channels. The minute-level deep learning model inputs the last 30 minutes of data, with a data shape of [30, 75] and an update step of 1 minute; the hour-level deep learning model inputs the last 240 minutes of data, and takes the mean value every 5 minutes to obtain a shape of [48, 75], with an update step of 30 minutes.

[0183] S4, Dual-time-scale collaborative optimization:

[0184] S41, Minute-level control loop collaborative strategy:

[0185] The minute-level control loop has a control period of 1 minute, and the minute-level lower confidence bound and the dynamic control threshold are used to regulate the VAD device. If the operational state is normal, the system maintains the current pump speed unchanged, and only monitors the core physiological quantities (MAP, PP, and ) and its dynamic control threshold, while monitoring the VAD device running state (pump speed N, power Power); if the running state is pre-warning, execute small step speed regulation; if the running state is emergency, start MPC-lite rolling optimization.

[0186] The system refreshes every minute and the observation quantities used for decision-making include: core physiological data MAP, PP, and , and device running state pump speed N, power Power, to ensure that the device works within the predetermined range.

[0187] Engineering approximation of lower confidence bound (LCB) of core physiological quantity and one-sided smoothing:

[0188] For easy online implementation, the "point value-safety margin" method is used to approximate the lower confidence bound when only point prediction is used, and one-sided exponential smoothing is used to suppress jitter alignment. Let t be the current minute, and the point prediction of the kth minute in the future , define the original lower confidence bound as:

[0189] ,

[0190] , wherein, is the predicted value of the kth minute in the future, m x is a fixed safety margin (MAP: 3 mmHg; PP: 2 mmHg; : 1 mmHg). Then the safety lower confidence bound after one-sided exponential smoothing is:

[0191] ,

[0192] , is the smoothing coefficient (default α = 0.3); the initial condition is taken as . If the value of the previous minute does not exist, let .

[0193] Core physiological quantity margin:

[0194] The calculation formulas of left ventricular minimum pressure margin, mean arterial pressure margin and pulse pressure margin are as follows:

[0195] ,

[0196] ,

[0197] .

[0198] S42, hour-level control loop coordination strategy:

[0199] The hour-level loop takes 30 minutes as the control period, and according to the lower confidence bound of the hour-level running safety margin The operating reference and reference bias, pump speed control constraints and cost weight are coordinated and set, and are issued to the minute-level loop.

[0200] The hour-level loop performs the above setting with a period of 30 minutes. When any of the following events occurs at minute t, the setting program is immediately started:

[0201] 1. Low-frequency key input (such as lactate, NYHA, etc. record) refresh;

[0202] 2. Device parameter re-calibration (such as N min , N max adjustment);

[0203] 3. Emergency / MPC-lite treatment ends (from emergency zone to non-emergency zone).

[0204] S43, closed-loop feedback intelligent control execution:

[0205] The sensor acquisition and device telemetry enter the system with a 1-minute period, and the lower confidence limit of the minute-level / hour-level operation safety probability output by the deep learning model; S22 online inversion updates θ and keeps it within the offline feasible range, and calculates the dynamic control threshold based on PHM. On this basis, the following closed-loop process is executed:

[0206] (1) The minute-level loop performs operation state determination (normal / early warning / emergency) on the current working condition with a 1-minute period, and executes "maintain current pump speed / small step speed adjustment / trigger MPC-lite" respectively;

[0207] (2) The hour-level loop sets the  control period of 30 minutes, and switches between the normal file / conservative file two sets of parameter tables according to the operation state, and issues the corresponding parameter set to the minute-level loop.

[0208] (3) Decision synthesis and issuance: the two-level decisions are synthesized into a single pump speed command after feasibility screening and sent to the device;

[0209] (4) Feedback loop and frozen strategy: the physiological response (physiological indicators, PHM calculated physiological parameters) and device telemetry return flow (such as pump speed, power) after execution will be fed back to the control system in real time, used to refresh , update θ(t) and recalculate the dynamic control threshold, forming a closed-loop adaptive control. During the emergency / MPC-lite treatment, is temporarily frozen to avoid frequent adjustment of the reference benchmark in emergency situations. At this time, the system only adjusts the equipment operation based on real-time feedback, and after the emergency state is lifted, the setting is restarted.

[0210] The above examples only illustrate the technical idea of the present application, and cannot be used to limit the protection scope of the present application. Any modification made according to the technical idea of the present application on the basis of the technical scheme falls within the protection scope of the present application.

[0211] The present application also provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the ventricular assist device control method.

[0212] The present application also provides an electronic device, comprising:

[0213] a memory for storing a computer program;

[0214] a processor for executing the computer program to implement the steps of the ventricular assist device control method.

[0215] Those skilled in the art should understand that embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer-usable program code.

[0216] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce the functions specified in the flowcharts and / or block diagrams for implementing the flows and / or blocks in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified in the flow or flows and / or block or blocks.

[0217] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including instruction means, which implement the flows and / or blocks in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 means for performing the function specified in the flow or flows and / or block or blocks.

[0218] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are generated to realize the computer-implemented processes in the computer or other programmable devices, and the instructions executed in the computer or other programmable devices provide steps for implementing the functions specified in the flowchart Figure 1 or multiple flows and / or blocks Figure 1 or multiple blocks.

Claims

1. A control method for a ventricular assist device based on a dynamic model and deep learning, characterized in that, The method includes: S1. Construct a first deep learning model within a set control cycle to obtain the first operating state of the ventricular assist device, wherein the first operating state includes three types: normal, warning, and emergency. Based on the confidence lower bound of the first deep learning model output within the control cycle, and combined with the set safety constraints, the first operating state of the ventricular assist device is determined. The safety constraints include mean arterial pressure not lower than the corresponding mean arterial pressure threshold, pulse pressure not lower than the corresponding pulse pressure threshold, and left ventricular pressure not lower than the corresponding left ventricular pressure threshold. If any safety constraint exceeds the limit or the minimum confidence lower limit output by the first deep learning model is less than the preset emergency threshold, the first operating state is emergency; if all safety constraints do not exceed the limit, and the minimum confidence lower limit output by the first deep learning model is greater than or equal to the preset emergency threshold and less than the preset normal threshold, the first operating state is warning; if all safety constraints do not exceed the limit, and the minimum confidence lower limit output by the first deep learning model is greater than or equal to the preset normal threshold, the first operating state is normal. S2. Based on the operating status obtained from S1, control the ventricular assist device: when the operating status is normal, maintain the current pump speed; when the operating status is warning, execute a small-step speed control strategy according to the set step increment; when the operating status is emergency, execute a model predictive control strategy.

2. The method according to claim 1, characterized in that, The first deep learning model consists of two cascaded one-dimensional convolutional layers, two bidirectional gated recurrent units, and an output layer.

3. The method according to claim 1, characterized in that, The small-step speed control strategy in step S2 specifically includes: The pump speed is adjusted according to the following update formula: ， Where, N t N represents the pump speed in minute t. t+1 ∆N represents the target pump speed at minute t+1. t This represents the stride increment at minute t. Both the target pump speed and the step increment satisfy the set pump speed control constraints, and the adjustment direction follows the principles below: like or If so, reduce the pump speed, i.e., ∆N t Less than 0; where, This represents the minimum pressure margin of the left ventricle. The threshold for suction margin in the warning zone. This is the pump power-speed residual. Set the power-speed standardized residual threshold for the warning zone; otherwise, in and Increase pump speed, i.e., ∆N t Greater than 0; where, The average pressure tolerance for the warning zone is used as a reference. MAP is the pulse pressure margin threshold for the warning zone. ref (t) represents the reference mean arterial pressure at minute t. , Let t be the nominal reference mean arterial pressure. Let be the reference offset at minute t. Let t be the safe confidence limit for mean arterial pressure at minute t. Let be the pulse pressure margin at minute t.

4. The method according to claim 1, characterized in that, The model predictive control strategy in S2 specifically includes: Set prediction step size L The prediction step index is denoted as k=0,…,L-1, and the pump speed is adjusted according to the following update formula for each prediction step: ， in, This represents the pump speed at the k-th predicted step in minute t. This represents the pump speed at the (k+1)th prediction step in minute t. This represents the step increment for the k-th prediction step at minute t. Both the target pump speed and the step increment meet the set pump speed control constraints, and The objective function is obtained by selecting the sequence that minimizes the objective function from the feasible set that satisfies physiological safety constraints, and based on the obtained sequence... pump speed at minute t Adjustments are made. Let be the stride increment at minute t; the objective function at minute t is: ， in, , MAP represents the cost weight at minute t. ref (t) represents the reference mean arterial pressure at minute t. , Let t be the nominal reference mean arterial pressure. Let be the reference offset at minute t. Let be the mean arterial pressure at the k-th prediction step in minute t. This represents the step increment at the (k-1)th prediction step in the t-th minute.

5. The method according to claim 4, characterized in that, After step S1 and before step S2, the method further includes: Within a set control cycle, a second deep learning model is constructed to obtain the second operating state of the ventricular assist device. The second operating state includes two types: normal and warning. The acquisition frequency of the second operating state is less than that of the first operating state. Based on the acquired second operating state, the nominal reference mean arterial pressure, reference offset, pump speed regulation constraint and cost weight are co-tuned and then used for small-step speed control and model predictive control in step S2.

6. The method according to claim 5, characterized in that, The second deep learning model consists of a cascaded downsampling layer, a one-dimensional convolutional layer, two gated recurrent units, and an output layer.

7. The method according to claim 5, characterized in that, Coordinated tuning includes: (1) If the second operating state is normal, the nominal reference mean arterial pressure remains unchanged; otherwise, the tuning formula for the nominal reference mean arterial pressure is: ， in, The nominal reference mean arterial pressure, The average of historically measured mean arterial pressures during the first normal operating state; and These are the preset minimum and maximum mean arterial pressures, respectively; (2) If the second operating state is normal in two consecutive control cycles, the nominal reference mean arterial pressure, reference offset, pump speed control constraint and cost weight are selected from the preset normal mode parameter table; otherwise, the nominal reference mean arterial pressure, reference offset, pump speed control constraint and cost weight are selected from the preset conservative mode parameter table. The normal mode parameter table and the conservative mode parameter table are both preset based on expert knowledge.

8. The method according to claim 1, characterized in that, By constructing the following individualized hemodynamic model, the mean arterial pressure threshold, pulse pressure threshold, and left ventricular pressure threshold are dynamically obtained: ， ， ， ， ， in, Let t be the left ventricular pressure at minute t. Let Vt be the left ventricular volume at minute t, and V0 be the stress-free volume. E represents the time-varying elasticity of the left ventricle at minute t. min With E max These are the elastic thresholds at end-diastole and peak systole, respectively. Let be the normalized activation momentum function aligned with the heart rate phase at minute t. Let P be the aortic pressure at minute t. v Q is the reference venous pressure. in (t) represents the total flow rate entering the aorta at minute t. , The original left ventricular ejection flow rate at minute t. Let be the pump flow rate at minute t, R be the peripheral resistance, C be the arterial compliance, and N(t) be the pump rate at minute t. Let be the pressure difference across the pump at minute t, and f(.) be a function that maps the pump speed and the pressure difference across the pump to flow rate. This is a calibrable vector of pump parameters; Within a set cardiac cycle, an individualized hemodynamic model was used for simulation to obtain the minimum simulated mean arterial pressure, minimum simulated pulse pressure, and minimum simulated left ventricular pressure in the simulation results. Finally, the mean arterial pressure threshold, pulse pressure threshold, and left ventricular pressure threshold are obtained: ， ， ， in, These are the mean arterial pressure threshold, pulse pressure threshold, and left ventricular pressure threshold at minute t, respectively. These are the lower limit of mean arterial pressure, the lower limit of pulse pressure, and the left ventricular pressure threshold at minute t, respectively. These represent the minimum simulated mean arterial pressure, minimum simulated pulse pressure, and minimum simulated left ventricular pressure at minute t, respectively.

9. The method according to claim 3, characterized in that, By minimizing the objective function Iteratively update the parameter vector of the individualized hemodynamic model : ， in, To calculate the mean arterial pressure at minute t using an individualized hemodynamic model at θ, Let be the measured mean arterial pressure at minute t. Let be the nominal reference mean arterial pressure at minute t.

Images