False suction identification and nutrition pump closed loop method based on strategy optimization

Abstract

The invention discloses a mistaken suction recognition and nutrition pump closed-loop method based on strategy optimization, and aims to solve the problem of strategy optimization of hierarchical response and action gating based on confidence under security constraints. According to the invention, through multi-modal time sequence signal risk estimation and action conditional conformal prediction calibration, hierarchical gating of a risk budget water tank and a resident clock, combination of a quadratic programming safety shielding layer and a control barrier function, and execution and closed-loop feedback of a normal differential equation, the risk budget water tank and the resident clock are subjected to multi-modal time sequence signal risk estimation and action conditional conformal prediction calibration; the technical effects of safe and feasible control of speed reduction, short stop, long stop and backwashing and pump speed track optimization under the clinical safety boundary are achieved, the mistaken suction recognition reliability is improved, frequent switching and overshoot are restrained, and nutrition target achievement and forward safety invariance are considered.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of clinical enteral nutrition infusion, and particularly relates to an aspiration identification and nutrition pump closed-loop method based on strategy optimization. BACKGROUND

[0002] At present, clinical enteral nutrition infusion is widely carried out in ICUs and general wards. In terms of monitoring and disposal of aspiration risk, the existing technology has developed from manual experience and threshold alarm to intelligent identification of multi-modal signals such as respiratory waveform, blood oxygen saturation, acoustics and body position, and there are explorations of semi-closed loop or prototype closed loop control of the nutrition pump. In terms of safety guarantee, there are also rule bases based on hard threshold such as speed limit, pump stop and backwashing.

[0003] However, these schemes still have the following deficiencies:

[0004] 1. Risk estimation is mostly dependent on single threshold or black box classifier, lacking of evidence weighting and coverable confidence boundary for different signal qualities, and lacking of time alignment and consistency check, which makes it difficult to provide verifiable risk upper bound and coverage width for graded response;

[0005] 2. The strategy layer mostly uses fixed rules to trigger actions, lacks dynamic trade-off mechanism (such as risk budget) for safety and nutrition goals, and lacks resident and sticky constraints, which is easy to cause frequent switching, oscillation and over-conservative / over-aggressive behavior;

[0006] 3. In terms of safe execution and control, there is generally a lack of methods to unify instantaneous and cumulative clinical boundaries into optimization solutions, a lack of safety shields (such as QP combined with CBF) that can guarantee forward invariance, and a lack of consideration of pump-pipeline continuous dynamics and slope limits, which leads to overshoot and unstable execution.

[0007] Therefore, an aspiration identification and nutrition pump closed-loop method capable of solving the above deficiencies of the existing technology is a problem to be solved by those skilled in the art. SUMMARY

[0008] One purpose of the present application is to propose an aspiration recognition and nutritional pump closed-loop method based on strategy optimization, aiming at the problems of lack of calibratable confidence boundary in risk estimation, lack of dynamic trade-off and hysteresis in strategy, and difficulty in unifying instantaneous and cumulative safety constraints in the execution layer, a technical solution is proposed to calculate risk by multi-modal time series features and evidence weighted calculation, and to obtain risk upper bound by action conditioned conformal prediction calibration; to form viscous hysteresis by combining the risk budget tank and the nutritional progress driven hierarchical gate with the resident clock; to introduce control barrier function in the quadratic programming safety shielding layer to unify the maximum pump speed, minimum pump stop time and cumulative constraint; and to match the pump / pipe dynamics and slope limitation by using neural ordinary differential equation. The present application has the technical effects of realizing safe and feasible control of speed reduction, short stop, long stop and back flushing within the clinical safety boundary, improving the reliability of aspiration recognition, inhibiting frequent switching and overshoot, and considering the achievement of nutritional target and safety forward invariance.

[0009] According to an aspiration recognition and nutritional pump closed-loop method based on strategy optimization, the method comprises the following steps:

[0010] S1, bedside multi-modal time series data is collected and preprocessed to obtain unified timestamp respiratory waveform, blood oxygen saturation, acoustic signal, body position signal, pump pipeline pressure and pump pipeline flow;

[0011] S2, the aspiration risk probability and evidence strength at the current time are calculated based on the six types of signals, and the action conditioned conformal prediction calibration is performed on the action level to obtain the risk upper bound and coverage width of each action level;

[0012] S3, according to the aspiration risk probability, evidence strength and risk upper bound and coverage width of each action level, and combined with the execution feedback of the last control cycle, the risk budget tank value and the nutritional target progress value are updated;

[0013] S4, according to the updated risk budget tank value and nutritional target progress value, combined with the risk upper bound and coverage width of each action level, hierarchical gating is implemented, and according to the resident timer in the system state register, action viscosity constraint and switching penalty are applied to generate unfiltered target action level and pump speed reference trajectory;

[0014] S5, based on the unfiltered target action level and pump speed reference trajectory, combined with the resident timer, cumulative pump stop time and cumulative back flushing volume in the system state register, the clinical safety boundary is applied, the safety shielding layer of quadratic programming is adopted and the control barrier function is introduced to solve the safety constraint, to obtain the safe action level and the safe pump speed trajectory, and to update the resident timer, the cumulative pump stop time and the cumulative back flushing volume;

[0015] S6, according to the safety action level and the safety pump speed trajectory, combining the execution dynamics and slope limit of the pump and pipeline, a neural differential equation model is used to generate pump control instructions and output to the pump actuator;

[0016] S7, the pump actuator completes the actual infusion according to the pump control instructions, generates execution feedback, and is used for the next control period to maintain closed-loop operation.

[0017] Optionally, step S1 is specifically:

[0018] The respiratory waveform, blood oxygen saturation, acoustic signal, body position signal, pump pipeline pressure and pump pipeline flow are time-stamped and clock deviation corrected based on the system clock to eliminate the time drift of each channel and realize time alignment;

[0019] The six types of signals are respectively denoised, de-drifted, abnormal segment removed, missing value interpolated and resampled to form a unified timestamp with fixed sampling interval;

[0020] The six types of signals are calibrated in amplitude and unit to ensure the consistency of physical quantities;

[0021] And the six types of signals are respectively generated signal quality marks to indicate the states of probe falling off, saturation, over-range and strong interference;

[0022] The respiratory waveform, blood oxygen saturation, acoustic signal, body position signal, pump pipeline pressure, pump pipeline flow and corresponding signal quality marks with unified timestamp are output.

[0023] Terminology interpretation:

[0024] The bedside multi-modal time series data is continuous time series data of multiple sensing channels collected around the patient and the infusion pipeline, including at least respiratory waveform, blood oxygen saturation, acoustic signal, body position signal, pump pipeline pressure and pump pipeline flow;

[0025] The respiratory waveform is a time-varying signal representing the patient's respiratory movement or airflow change, which can be obtained by a chest and abdominal belt, airflow / airway pressure flow sensor or bed surface sensor;

[0026] The blood oxygen saturation is a transcutaneous arterial blood oxygen saturation signal (SpO2) provided in time series form, collected by a pulse oximeter;

[0027] The acoustic signal is a time series signal of audio or vibration related to respiration, which can be obtained by a microphone, electronic stethoscope or acceleration / vibration sensor;

[0028] The body position signal is a time series signal representing the patient's posture, body position change or bed body angle, which can be obtained by an inertial measurement unit, bed angle sensor or monitoring system;

[0029] The pump line pressure is the fluid pressure signal in the output side of the pump or in the infusion line, measured by a pressure sensor or estimated by a model;

[0030] The pump line flow is the volumetric flow signal in the infusion line, measured by a flow sensor or estimated by pump operating parameters / model;

[0031] The system clock is the internal clock or synchronized external clock used as the time reference for the whole process of acquisition and control;

[0032] The time stamping is the process of assigning each data sample with a time identifier from the system clock at the time of sampling;

[0033] The clock bias correction is the process of estimating and compensating for the clock offset and frequency drift between different channels to reduce the cross-channel time error;

[0034] The unified time stamp is the common time axis established with the system clock as the reference, so that each channel sample corresponds to the same time reference, preferably a fixed sampling interval time sequence;

[0035] The fixed sampling interval is the constant time interval between adjacent samples of the unified time stamp;

[0036] The denoising is the signal processing step of suppressing random noise and identifiable narrowband / power frequency interference components in each channel to improve the signal-to-noise ratio;

[0037] The de-drifting is the signal processing step of eliminating low-frequency baseline drift or slow trend items to restore a stable baseline;

[0038] The abnormal segment rejection is the process of identifying and removing the untrustworthy signal segment according to the amplitude / variation rate threshold, physical feasibility rule or quality label;

[0039] The missing data interpolation is the process of interpolating the time gap caused by the interruption or sample loss in the sampling to restore the continuous time sequence;

[0040] The resampling is the process of converting channels with different sampling rates or irregular sampling into a unified fixed sampling interval time sequence;

[0041] The amplitude and unit calibration is the process of converting the numerical scale and unit of measurement of each channel to a predetermined physical unit and correcting the gain / bias error;

[0042] The signal quality label is a quality state label assigned to each channel over time, indicating the degree of trustworthiness of the channel at the corresponding time and used for subsequent weight allocation, which can indicate the state of probe falling off, saturation, over-range or strong interference, etc.

[0043] Optionally, step S2 specifically includes:

[0044] Based on respiratory waveform, blood oxygen saturation, acoustic signal, body position signal, pump line pressure and pump line flow, time series features are extracted and consistency is checked so that various signals participate in risk estimation under a unified timestamp;

[0045] The contribution weights of various signals based on signal quality labels are weighted according to the strength of evidence to form the strength of evidence, which is then combined with time-series features to calculate the probability of false absorption risk.

[0046] In the candidate action level set, the action conditional conformal prediction calibration is performed for deceleration, short stop, long stop and backwash respectively. Based on the preset coverage target, the risk upper bound and coverage width are generated for deceleration, short stop and coverage width, long stop and coverage width, and backwash.

[0047] Output the probability of accidental aspiration risk, the strength of evidence, the upper bound and coverage of risk for deceleration, the upper bound and coverage of risk for short stops, the upper bound and coverage of risk for long stops, and the upper bound and coverage of risk for backwashing.

[0048] Terminology definition:

[0049] The aspiration is a clinical adverse event or a high-risk condition in which gastric contents or nutrient solution enter or may enter the respiratory tract during enteral nutrition infusion.

[0050] The time-series features are statistical quantities or transformation quantities extracted from multimodal signals with a unified timestamp that characterize the dynamics of the correlation between the signal and the false absorption, including but not limited to amplitude, slope, rate of change, frequency domain / time frequency features and cross-channel correlation features.

[0051] The consistency check is a process of verifying and correcting the consistency of multi-channel signals in terms of physical feasibility, relative timing and phase relationship, in order to eliminate contradictory information and ensure that features are available under the same time reference.

[0052] The strength of evidence is a scalar measure that reflects the effective amount of information and reliability currently used for risk estimation. It is calculated based on the signal quality label and information contribution of each channel, and the preferred value is [0,1].

[0053] The evidence strength weighting is performed by weighting and aggregating the feature contributions of each channel based on the signal quality label to form the evidence strength and participate in the risk estimation process;

[0054] The aspiration risk probability is an estimated value of the probability of aspiration occurring or being in a high-risk aspiration state at the current moment, and the value is preferably in [0,1].

[0055] The candidate action level set is a set of discrete control actions that the system can execute, including at least deceleration, short stop, long stop and backwash;

[0056] The action level is any specific action category in the candidate action level set, used to refer to different intensities or types of handling strategies;

[0057] The speed reduction is a control action that reduces the instantaneous infusion rate by decreasing the pump speed;

[0058] The short stop is a pump stop control action that keeps the pump speed at zero for a predetermined short dwell time;

[0059] The long stop refers to a pump stop control action that keeps the pump speed at zero for a predetermined long dwell time;

[0060] The backwashing is a control action that involves short-term infusion in the reverse direction within a limited time period to clean the pipeline or reduce the risk of aspiration.

[0061] The action conditionalization is to differentiate the risk estimation or calibration process by using the action level as a condition variable, so that different action levels correspond to their respective risk measures and calibration parameters.

[0062] The conformal prediction calibration is a distribution-independent finite sample statistical calibration method that calibrates the risk estimate under a given coverage target to generate a risk interval or upper bound with guaranteed coverage properties.

[0063] The preset coverage target is the target coverage level or confidence threshold used for conformal prediction calibration, which is usually expressed as a numerical setting of 1-α.

[0064] The risk upper bound is an upper confidence bound or upper side bound of the probability of actual accidental aspiration risk under the preset coverage target, which is used to ensure that the risk is not underestimated under the coverage target.

[0065] The coverage width is a measure of risk uncertainty obtained from conformal prediction calibration, preferably expressed as the width of the risk interval or the margin of the risk upper bound relative to the point estimate, to reflect the conservatism of the current calibration results.

[0066] Optionally, step S3 specifically includes:

[0067] The risk intensity value is calculated based on the probability of aspiration risk and the strength of evidence. A conservative risk upper bound is synthesized based on the risk upper bound for deceleration, the risk upper bound for short stops, the risk upper bound for long stops and the risk upper bound for backwashing. The conservative risk upper bound is used to determine the risk consumption rate of the current control cycle by combining the coverage width with a preset weight.

[0068] Based on the actual flow rate, pipeline pressure, and injection volume generated in the previous control cycle, the risk budget tank value is deducted and restored. The deduction amount is calculated as the product of the risk consumption rate and the evidence strength, and additional deductions are applied when rapid fluctuations in pipeline pressure or actual flow rate are detected. The restoration amount takes effect according to a preset time constant when the risk strength value is below the threshold and the signal is stable. The updated risk budget tank value is limited to between zero and the budget upper limit.

[0069] The nutritional target progress value is calculated based on the infusion volume generated in the previous control cycle, the target infusion volume specified in the doctor's order, and the target time window. It does not increase during the pump stoppage period when the actual flow rate is zero, but increases according to the ratio of the cumulative infusion volume to the target infusion volume during effective infusion.

[0070] During the initial control cycle, the risk budget tank value is initialized to the budget limit, the nutrient target progress value is initialized to zero, and the residence timer, cumulative pump stop duration, and cumulative backwash volume are initialized to zero.

[0071] Output the updated risk budget tank value and the updated nutrition target progress value.

[0072] Terminology definition:

[0073] The risk intensity value is a quantitative risk measure obtained by combining the probability of accidental aspiration risk and the strength of evidence, and preferably takes a value in [0,1].

[0074] The conservative risk upper bound is a single upper risk bound synthesized by weighting or taking extreme values ​​of the risk upper bounds for deceleration, short stop, long stop and backwashing according to preset principles, and is used for subsequent budget and gating calculations.

[0075] The preset weights are weighting coefficients used for different action levels or uncertainty factors when synthesizing the upper bound of conservative risk or calculating the risk consumption rate.

[0076] The current control cycle is a fixed-duration window between two consecutive decisions and executions in closed-loop control.

[0077] The previous control cycle is the control cycle immediately preceding the current control cycle;

[0078] The risk consumption rate is the proportion or rate coefficient of the consumption of the risk budget tank value within the current control period, determined based on the conservative risk upper bound and the coverage width.

[0079] The actual flow rate is a representative value obtained by periodically statistically analyzing the pump pipeline flow rate sample in the previous control cycle, preferably a time average or median.

[0080] The infusion volume is the forward infusion volume obtained by time integration of the flow samples with the same timestamp in the previous control cycle;

[0081] The risk budget tank value is a time-varying scalar representing the system's available risk budget margin, limited to between zero and the budget ceiling.

[0082] The deduction amount is the reduction in the risk budget tank value during the current control period, calculated as the product of the risk consumption rate and the strength of evidence, combined with the rapid fluctuation bonus.

[0083] The additional deduction is an additional deduction applied to the risk budget tank value when rapid fluctuations in pipeline pressure or actual flow are detected.

[0084] The recovery amount is the increase in the risk budget tank value when the risk intensity value is below the threshold and the signal is stable, and is controlled by a preset time constant.

[0085] The preset time constant is a time parameter that controls the growth rate of the recovery amount.

[0086] The threshold is a numerical threshold used to determine whether the risk intensity is in the low-risk range;

[0087] The signal stability is defined as a state in which the fluctuation amplitude and rate of change of the key signal do not exceed the corresponding threshold within a preset time window;

[0088] The budget ceiling is the maximum permissible upper limit of the risk budget tank value;

[0089] The nutritional target progress value is a scalar measure representing the completion rate of the medical order target infusion, and is updated according to the ratio of the cumulative infusion volume to the target infusion volume;

[0090] The target infusion volume is the total volume of nutrition that must be infused within the target time window as specified in the doctor's order;

[0091] The target time window is the time range within which the target infusion volume is to be completed as specified in the doctor's order.

[0092] The pump stop period is a continuous period of time during which the actual flow rate is zero;

[0093] The effective infusion is defined as a forward infusion state where the actual flow rate is greater than zero and the system is not in a backflushing state.

[0094] The cumulative infusion volume is the sum of the positive infusion volumes counted from system startup to the current moment;

[0095] The initial control cycle is the first control cycle after the closed-loop system starts up.

[0096] The dwell timer is a system state variable that records the duration of the current action level;

[0097] The cumulative pump stop time is the total time from system startup to the current moment when zero speed is maintained;

[0098] The cumulative backwash volume is the cumulative reverse infusion volume from system startup to the current moment;

[0099] The execution feedback refers to the measurement results obtained at the end of the previous control cycle and used for status and budget updates, including at least the actual flow rate, pipeline pressure, and injection volume.

[0100] Optionally, step S4 specifically includes:

[0101] Based on the updated risk budget tank value and the updated nutrition target progress value, the strategy reward weights for deceleration, short stop, long stop and backwash in the candidate action level set are calculated respectively, and the reward distribution of each action level is formed.

[0102] Based on the risk upper bound and coverage width of deceleration, short stop, long stop and backwashing, a graded gating threshold is set, and graded gating is executed so that the action level obtains effective strategy preference only when the corresponding risk upper bound does not exceed the gating threshold and the coverage width meets the preset coverage target.

[0103] Based on the dwell timer in the system status register, action stickiness constraints and switching penalties are applied. When the dwell timer has not reached the minimum dwell time parameter, the switching of action level is restricted and the switching-related policy preferences are reduced to form endogenous hysteresis.

[0104] Based on the strategy preference after hierarchical gating and viscosity constraint processing, the unfiltered target action level is selected, and a pump speed reference trajectory is generated based on the selected unfiltered target action level and the updated nutrient target progress value. The pump speed reference trajectory satisfies the preset change range and time characteristics, wherein the deceleration corresponds to a downward slope-limited curve, short stop and long stop correspond to zero speed holding curves during their respective residence time, and backwash corresponds to the reverse infusion curve during the backwash period.

[0105] Output unfiltered target motion level and pump speed reference trajectory.

[0106] Terminology definition:

[0107] The hierarchical gating is a decision-making mechanism that allows or blocks actions based on the risk upper bound and coverage width corresponding to each action level relative to the hierarchical gating threshold, so that only actions that meet the safety confidence standard participate in subsequent strategy preferences and action selection.

[0108] The system status register is a data structure used to persistently store decision-related status quantities between control cycles, including at least a resident timer and may contain information such as the level of the previous executed action;

[0109] The action stickiness constraint is a constraint on the minimum dwell time and the tendency to maintain action switching based on the dwell timer, in order to reduce frequent switching.

[0110] The switching penalty is a penalty imposed on cross-action level switching during policy evaluation, which suppresses unnecessary switching by reducing the policy weight or preference of the relevant action.

[0111] The unfiltered target action level is the action level to be executed selected after hierarchical gating and sticky constraint processing, and has not yet been constrained by the safety shield layer and clinical boundary.

[0112] The pump speed reference trajectory is a reference curve of the expected pump speed changing over time, generated based on the unfiltered target action level and the current state, for subsequent safety optimization and execution layer tracking;

[0113] The strategy return weight is a weight parameter used to measure the overall return of each action level. It is calculated based on state variables such as risk budget tank value and nutrition target progress value and is used for action priority evaluation.

[0114] The reward distribution is a distribution formed by normalizing or sorting the policy reward weights of candidate action levels, and is used to support action selection.

[0115] The hierarchical gating threshold is a set of thresholds used for hierarchical gating determination, which includes at least the upper limit of the risk upper bound and the minimum requirement for the coverage width.

[0116] The policy preference is a measure of the relative tendency to choose each action level in the current decision, taking into account the policy reward weight, the hierarchical gating result and the switching penalty.

[0117] The minimum dwell time parameter is a parameter that limits the shortest duration that a certain action level must be maintained once adopted;

[0118] The intrinsic hysteresis is the inherent lag characteristic of action switching generated by the action stickiness constraint and the switching penalty, which is used to avoid high-frequency round-trip switching.

[0119] The preset variation amplitude and time characteristics are parameterized restrictions set on dynamic characteristics such as amplitude range, change slope, acceleration or duration when generating the pump speed reference trajectory.

[0120] The slope-limited curve is a reference curve that monotonically decreases or changes at a limited rate under the condition that the rate of change is constrained by an upper limit, and is used to achieve deceleration action;

[0121] The zero-speed holding curve is a reference curve in which the pump speed is always zero during a given residence time, and is used for short-stop or long-stop operations.

[0122] The reverse injection curve is a reference curve in which the pump speed is reversed within a specified time period and is limited by amplitude and slope, and is used for backwashing.

[0123] The backwashing period is a limited time window for performing reverse infusion, determined by the strategy layer or the security layer.

[0124] Optionally, step S5 specifically includes:

[0125] Based on the unfiltered target action level and pump speed reference trajectory, combined with the resident timer, cumulative pump stop duration and cumulative backflushing volume in the system status register, clinical safety boundaries of maximum pump speed, minimum pump stop time, maximum cumulative pump stop duration and maximum cumulative backflushing volume are applied to construct a quadratic programming problem and introduce a control barrier function to ensure the forward invariance of the safety set;

[0126] The goal of the quadratic programming is to minimize the weighted square deviation of the safe pump speed trajectory from the pump speed reference trajectory while satisfying the above constraints, and to suppress the excessively rapid changes in pump speed by limiting the slope.

[0127] The constraints include: the pump speed does not exceed the maximum pump speed; for short and long stops, the pump speed is kept at zero for the respective stop time to meet the minimum stop time; the current action level remains unchanged when the dwell timer has not reached the minimum dwell time; the cumulative stop time does not exceed the maximum cumulative stop time; the cumulative backwash volume does not exceed the maximum cumulative backwash volume; and the control barrier function keeps the system state from leaving the safe set.

[0128] The safe action level and safe pump speed trajectory are obtained by solving the problem, and the state variables are updated: when the safe action level is different from the action level of the previous control cycle and the dwell timer reaches the minimum dwell time, the dwell timer is reset to zero; otherwise, it is accumulated according to the control cycle duration.

[0129] When the safe pump speed trajectory includes a zero-speed segment, the pump stop time is accumulated by adding the duration of the zero-speed segment.

[0130] When the safe pump speed trajectory includes a reverse injection section, the cumulative backwash volume is added to the injection volume of the reverse injection section.

[0131] Outputs the safe action level and safe pump speed trajectory, as well as the updated dwell timer, updated cumulative pump downtime, and updated cumulative backwash volume.

[0132] Terminology definition:

[0133] The clinical safety boundary is a set of clinical and equipment safety restrictions that must be met in closed-loop control, including at least the maximum pump speed, minimum pump stop time, maximum cumulative pump stop time, and maximum cumulative backwash volume.

[0134] The safety shielding layer of the secondary planning is a module that corrects the unfiltered target action level and pump speed reference trajectory online in the form of secondary planning. Through optimization, the reference scheme is projected onto the action domain that meets the constraints of clinical safety boundary and control barrier function.

[0135] The quadratic programming problem is a convex optimization problem that aims to minimize the weighted squared deviation of the safe pump speed trajectory from the pump speed reference trajectory and is solved under linear or affine constraints.

[0136] The control barrier function is a continuous function used to characterize the safe set and to ensure that the system state does not deviate from the safe set by imposing inequality constraints on its derivative;

[0137] The safety set is a subset of the system state space that satisfies the constraints of clinical safety boundaries and control barrier functions.

[0138] The forward invariance is the property that once the system state is within the safe set, it remains within that set as it evolves over time under constraints and control.

[0139] The weighted squared deviation is a target metric that is the squared and accumulated result of the difference between the safe pump speed trajectory and the pump speed reference trajectory weighted by a preset weight.

[0140] The slope limit is a constraint that imposes an upper limit on the rate of change of pump speed over time, used to suppress excessively rapid changes in pump speed.

[0141] The maximum pump speed is the upper limit parameter of the pump speed allowed by the equipment and clinical practice;

[0142] The minimum pump stop time is the shortest duration parameter required for either a short or long stop operation to maintain zero speed.

[0143] The maximum cumulative pump stop time is the upper limit parameter of the total allowed zero-speed maintenance time from system startup to the current moment;

[0144] The maximum cumulative backwash volume is the upper limit parameter of the cumulative reverse infusion volume allowed from system startup to the current moment;

[0145] The solution to the safety constraints is a process of obtaining an execution solution that satisfies the constraints and is optimal in the objective through quadratic programming under the combined effect of the clinical safety boundary and the control barrier function.

[0146] The safe and feasible action level is the category of executable actions that satisfy all constraint conditions after the safety constraint solution is obtained;

[0147] The safe pump speed trajectory is a pump speed reference curve over time obtained by solving for safety constraints and satisfying the clinical safety boundary and control barrier functions;

[0148] The execution action level is the type of action actually performed by the pump actuator in the previous control cycle;

[0149] The control cycle duration is a fixed time length parameter between two adjacent decisions and executions;

[0150] The zero-speed segment is the continuous time period in the safe pump speed trajectory where the pump speed is equal to zero.

[0151] The reverse infusion segment is a continuous time period in the safe pump speed trajectory during which the pump speed is negative and backwashing is performed;

[0152] The status update is a process of resetting or accumulating the dwell timer, cumulative pump stop time, and cumulative backwash volume based on the safety action level and safe pump speed trajectory.

[0153] Optionally, step S6 specifically includes:

[0154] The control mode is selected based on the level of safe and feasible action, so that the deceleration corresponds to the downward tracking reference curve, the short stop and long stop correspond to the zero speed maintenance during their respective residence time, and the backwash corresponds to the reverse injection curve during the backwash period.

[0155] Based on the safe pump speed trajectory as a reference signal, the execution dynamics of the pump and pipeline are continuously predicted using a neural network constant differential equation model. The execution dynamics cover the dynamic characteristics of pump rotational inertia, pipeline compliance and frictional hysteresis. In the prediction process, a slope limit is applied to constrain the rate of change of pump speed and suppress overshoot, and to ensure that the instantaneous target pump speed does not exceed the envelope of the safe pump speed trajectory.

[0156] During the zero-speed maintenance phase of short and long stops, maintain zero speed until the end of the corresponding residence time. During the backwashing phase, generate the target pump speed curve for reverse infusion and maintain it during the backwashing period.

[0157] The instantaneous target pump speed, forward or reverse indication, and holding time are combined into a pump control command and output to the pump actuator.

[0158] Terminology definition:

[0159] The control mode is a control strategy type selected by the execution layer based on the level of safe and feasible action, including at least deceleration tracking, short stop zero speed maintenance, long stop zero speed maintenance and backwash reverse infusion.

[0160] The downward-tracking reference curve is a monotonically decreasing pump speed reference curve that needs to be tracked in the deceleration control mode and is limited by the slope.

[0161] The zero-speed maintenance refers to the execution state of maintaining the pump speed at zero for a set time.

[0162] The dwell time is a parameter set for the duration during which the action needs to be maintained, which is set for short or long stop actions. It can be set by the strategy layer or constrained by the safety layer.

[0163] The aforementioned neural network-parameterized ordinary differential equation model is used to model and numerically solve the continuous-time dynamics of the pump-pipeline system.

[0164] The pump and pipeline execution dynamics refer to the continuous time response behavior of the pump actuator and the delivery pipeline under control, including the evolution of speed, pressure and flow rate over time.

[0165] The pump rotational inertia is the inertial impedance characteristic of the rotating mass of the pump drive component in response to changes in rotational speed.

[0166] The pipeline compliance refers to the pressure-volume-flow buffering characteristics of the infusion pipeline due to its elasticity and compressibility.

[0167] The frictional hysteresis refers to the lag and history dependence characteristics between input and output in the system caused by friction and viscosity effects.

[0168] The continuous-time prediction is a process of forward evolution calculation of pump speed in the continuous-time domain based on the current state and the model.

[0169] The reference signal is the desired input used by the execution layer for tracking, specifically the safe pump speed trajectory;

[0170] The safety envelope is the upper limit of the allowable range set around the safety pump speed trajectory, used to ensure that the instantaneous target pump speed does not exceed the safety allowable value;

[0171] The instantaneous target pump speed is the target pump speed value generated by the execution layer at the current moment, which is subject to slope limitation and safety envelope constraint.

[0172] The target pump speed curve is a curve showing the change of the target pump speed over time, generated and maintained during a specified period (including the backwashing period).

[0173] The forward or reverse indicator is a control indicator that indicates the direction of infusion, representing forward infusion or reverse cleaning towards the patient, respectively.

[0174] The holding time is a duration parameter set for the zero-velocity holding or reverse infusion phase;

[0175] The pump control command is a set of control messages sent to the pump actuator, which includes at least the instantaneous target pump speed, forward or reverse identifier, and duration of hold.

[0176] The pump actuator is a device unit that converts pump control commands into pump speed and injection direction and generates actual injection.

[0177] Optionally, step S7 specifically includes:

[0178] Based on the pump control command, execute the control mode of speed reduction, short stop, long stop or backwashing within the current control cycle;

[0179] During execution, the system clock is used for time stamping, and the readings of the flow sensor and pressure sensor are continuously collected to form flow samples and pressure samples with the same timestamp.

[0180] At the end of the control cycle, the flow samples with the same timestamp are integrated over time according to the sampling interval to generate the infusion volume, and the infusion volume is not included in the backwash period;

[0181] At the end of the control cycle, periodic statistics are performed on flow samples and pressure samples with the same timestamp to generate actual flow and pipeline pressure.

[0182] The system outputs actual flow rate, pipeline pressure, and injection volume, which are used to update the risk budget tank value and nutrient target progress value in the next control cycle and maintain closed-loop operation.

[0183] Terminology definition:

[0184] The time stamping is the process of recording the time identifier of each sampling moment according to the system clock, which is used to form a sample sequence with a unified timestamp in the future;

[0185] The flow sensor reading is the instantaneous flow measurement value output by the flow sensor installed in the infusion pipeline at the sampling time;

[0186] The pressure sensor reading is the instantaneous pressure measurement value output by the pressure sensor installed in the infusion pipeline at the sampling time;

[0187] The flow sample is a discrete time series data point composed of flow sensor readings after time stamping;

[0188] The pressure sample is a discrete time series data point composed of pressure sensor readings after time stamping;

[0189] The sampling interval is the time interval between adjacent samples with a unified timestamp, and is preferably a fixed value for numerical integration and statistical calculation.

[0190] The time integration is a process of numerically integrating the discrete time series according to the sampling interval, used to accumulate the flow samples into the infusion volume of the corresponding time period;

[0191] The periodic statistics refer to the process of statistically summarizing samples with the same timestamp within a single control period to form a representative value, preferably including the mean, median, or weighted average.

[0192] The beneficial effects of this invention are:

[0193] 1. By using multimodal signal quality labeling and evidence strength weighting, combined with action conditional conformal prediction calibration, the risk upper bound and coverage width of each action level are obtained, realizing confidence-based hierarchical gating, improving the reliability and interpretability of aspiration risk identification, and reducing false alarms and false negatives.

[0194] 2. Introduce a dual-objective trade-off between risk budget water tank and nutrient target progress, combined with the sticky gate and switching penalty of the dwell timer, and the secondary planning safety shield layer combined with the control obstacle function to unify the maximum pump speed, minimum pump stop time and cumulative constraints, to ensure that the safety set remains unchanged forward, suppress frequent switching and overshoot, and obtain safe and feasible actions and smooth pump speed trajectory.

[0195] 3. The pump-pipeline continuous dynamics are characterized by the use of the constant differential equation of the psychiatric system and the application of slope constraints and safety envelopes to improve the consistency and controllability of execution and reference. Combined with closed-loop feedback, the infusion efficiency is improved under safety constraints, which promotes the timely achievement of nutritional targets. Attached Figure Description

[0196] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0197] Figure 1 This is a flowchart of a strategy-optimized method for identifying aspiration and a closed-loop nutrient pump proposed in this invention. Detailed Implementation

[0198] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0199] refer to Figure 1 A method for identifying aspiration and implementing a closed-loop nutrient pump based on strategy optimization, characterized by comprising:

[0200] S1. Collect and preprocess bedside multimodal time-series data to obtain respiratory waveforms, blood oxygen saturation, acoustic signals, body position signals, pump line pressure and pump line flow rate with a unified timestamp;

[0201] S2. Calculate the probability of accidental aspiration and the strength of evidence at the current moment based on the six types of signals, and perform conformal prediction calibration of action conditionalization at the action level to obtain the risk upper bound and coverage width of each action level;

[0202] S3. Based on the probability of aspiration risk, the strength of evidence, and the risk upper bound and coverage of each action level, and combined with the execution feedback of the previous control cycle, update the risk budget tank value and the nutritional target progress value.

[0203] S4. Based on the updated risk budget tank value and nutrient target progress value, and combined with the risk upper bound and coverage width of each action level, hierarchical gating is implemented, and action stickiness constraints and switching penalties are applied according to the dwell timer in the system status register to generate unfiltered target action level and pump speed reference trajectory.

[0204] S5. Based on the unfiltered target action level and pump speed reference trajectory, combined with the dwell timer, cumulative pump stop duration and cumulative backwash volume in the system status register, apply the clinical safety boundary, use the safety shield layer of quadratic programming and introduce the control barrier function to solve the safety constraints, obtain the safe action level and safe pump speed trajectory, and update the dwell timer, cumulative pump stop duration and cumulative backwash volume.

[0205] S6. Based on the safe action level and safe pump speed trajectory, combined with the execution dynamics and slope limits of the pump and pipeline, the pump control command is generated using a normal differential equation model and output to the pump actuator.

[0206] S7. The pump actuator completes the actual injection according to the pump control command, generates execution feedback, and uses it in the next control cycle to maintain closed-loop operation.

[0207] In this specific embodiment, S1 specifically refers to:

[0208] Using the system clock as a unified reference, time alignment, signal shaping, and quality marker generation are performed on six channels: respiratory waveform, blood oxygen saturation, acoustic signal, body position signal, pump line pressure, and pump line flow. First, the original timestamps of each channel are linearly corrected to eliminate clock offset and drift, using the following formula:

[0209] ;

[0210] in Indicates channel The Each original timestamp is in seconds. This indicates the corrected timestamp, in seconds. For channel The clock offset, in seconds. For channel Clock drift ratio, channel index Corresponding to respiratory waveform, blood oxygen saturation, acoustic signal, body position signal, pump line pressure and pump line flow rate, respectively, sample index. This is the sampling sequence number for this channel, starting from 1;

[0211] After cross-channel time correction, a unified time axis is established to achieve time alignment at fixed sampling intervals, using the following formula:

[0212] Generate a unified timestamp sequence;

[0213] in For the first A unified sampling time, in seconds. The system clock start time, in seconds. To standardize and fix the sampling interval, with the unit being seconds, To unify the sampling sequence numbers on the time axis, counting starts from 1;

[0214] Subsequently, the non-uniform samples corrected for each channel were resampled onto a unified time axis and missing measurements were compensated using an interpolation operator formula:

[0215] Construct a unified timestamp sample;

[0216] in For channel unified timestamp The interpolated sample at that location, in units consistent with the physical quantity of that channel. For channel Interpolation and missing measurement recovery operators, capable of piecewise linear or spline interpolation, and dimensionless operator symbols. For channel Original timestamp The original amplitude sample at that location, and the unit consistent with the physical quantity of that channel;

[0217] After resampling, denoising and de-drifting are performed on each channel, and amplitude and unit calibration are conducted to eliminate random noise and low-frequency drift, and to unify physical scale. Filtering and calibration formulas are used as follows:

[0218] and Obtain standardized samples;

[0219] in The unit is the same as the physical quantity of the channel for the filtered and de-drifted sample. For channel Filtering operators, used for noise and drift suppression, dimensionless operator notation, The sample is calibrated for amplitude and unit, and the unit is a predetermined standard physical unit. For channel Amplitude scaling calibration coefficient, Convert to standard units, dimensionless. For channel The amplitude offset correction term is consistent with standard physical units;

[0220] To ensure the reliability and interpretability of subsequent risk estimates, a quality marker vector is generated for each uniform timestamp sample, and the probe status and interference are recorded. The quality marker is defined as follows:

[0221] ;

[0222] in For channel At any moment A four-dimensional mass vector, dimensionless. This indicates a probe detachment or low signal-to-noise ratio detection result; 1 indicates abnormality, 0 indicates normality; dimensionless. This indicates the amplitude saturation detection result; 1 indicates saturation, 0 indicates unsaturation; it is dimensionless. This indicates an over-range test result; 1 indicates over-range, 0 indicates not over-range, and it is dimensionless. This indicates the result of strong interference detection; 1 indicates the presence of strong interference, and 0 indicates no obvious interference; it is dimensionless.

[0223] The final output is a unified timestamp sequence. calibrated sample sequences and quality marker sequence For use in subsequent steps.

[0224] In this specific embodiment, S2 specifically refers to:

[0225] The system extracts temporal features from respiratory waveforms, blood oxygen saturation, acoustic signals, body position signals, pump line pressure, and pump line flow under a unified timestamp and performs cross-channel consistency verification. This eliminates suspicious and contradictory segments before they are included in risk estimation. Subsequently, it performs evidence strength weighting based on signal quality labels and obtains conformal prediction calibration results with action conditions to support hierarchical gating and subsequent strategy optimization. To quantify the amount of currently available information and reliability, the evidence strength is defined as a weighted average.

[0226] ;

[0227] in Indicates that the unified timestamp index is The range of the strength of evidence at any given time is: And it is dimensionless. The channel indexes correspond sequentially to respiratory waveform, blood oxygen saturation, acoustic signal, body position signal, pump line pressure, and pump line flow rate. For channel The prior reliability weight is a non-negative dimensionless constant. For channel At any moment The range of instantaneous quality scores obtained based on quality label mapping is as follows: And it is dimensionless;

[0228] After completing the consistency check and evidence weighting, the fused feature vector is used as the input for risk estimation, and a concise logistic regression method is used to calculate the current probability of aspiration risk.

[0229] ;

[0230] in For a moment The range of the probability of aspiration risk is as follows: And it is dimensionless. The sigmoid activation function is dimensionless. For at any time Extracted from six types of signals and after consistency verification The units of each component in the dimensional fusion feature vector depend on the source physical quantity, but they have been standardized before being input into the model. The model weight vector is dimensionless. The bias term is dimensionless;

[0231] To provide a confidence bound with target coverage at the action level, for the candidate action set Perform conformal prediction calibration with action conditions separately and generate risk upper bound and coverage width, using a compact calibration expression:

[0232] ;

[0233] in These represent deceleration ds, short stop sp, long stop lp, and backwash rf, respectively. For at any time action The upper bound of risk ranges from 1 to 1. And it is dimensionless. For action The coverage width range is And it is dimensionless. To be the operator that takes the smaller of the two values, For action The conformal calibration coefficients are derived from the set of inconsistencies in the calibration set. of The quantiles are given and are dimensionless. For action The significance level of its coverage target is And it is dimensionless. For quantile operators, For action The set of calibration inconsistencies can be composed of residuals or conformal scores and is dimensionless. The range of values ​​for the dynamic uncertainty scaling factor, which complements the strength of evidence, is as follows: Furthermore, it is dimensionless, which allows for wider coverage even with weaker evidence to meet the coverage objective;

[0234] Finally, the output at each control moment and And the four action levels corresponding to For use in subsequent steps.

[0235] In this specific embodiment, S3 specifically refers to:

[0236] The system takes the probability and strength of evidence of aspiration risk, as well as the risk upper bound and coverage width of each action level, as inputs. Combined with the execution feedback of actual flow rate, pipeline pressure, and infusion volume from the previous control cycle, it updates the risk intensity, conservative risk upper bound, risk consumption rate, risk budget tank value, and nutrient target progress value under a unified timestamp to drive subsequent graded gating and strategy optimization. First, the current risk intensity is quantified as follows:

[0237] ;

[0238] in Indicates that the unified timestamp index is The risk intensity value and its range at any given time are: And it is dimensionless. Indicates time The probability of aspiration risk and its range are as follows: And it is dimensionless. Indicates time The strength of evidence and the range of values ​​are as follows: And it is dimensionless;

[0239] To provide conservative guidance for the strategy layer, a set of candidate actions is defined. These correspond to deceleration, short stop, long stop, and backwashing, respectively, and are combined to form a conservative upper bound for risk. ;

[0240] in Indicates time The upper bound of conservative risk and the range of values ​​are: And it is dimensionless. Indicates time action The risk upper bound and the range of values ​​are: And dimensionless, operators For finding the maximum value operator;

[0241] To incorporate uncertainty into consumption decisions, a coverage width-weighted mean is defined:

[0242] Risk consumption rate ;

[0243] in Indicates time The weighted average coverage width and its range are: And it is dimensionless. Indicates time action The coverage width and value range are: And it is dimensionless. For action Normalized weights, satisfying And it is dimensionless. For a moment The risk consumption rate and its range are as follows: And it is dimensionless. and These are non-negative weighting coefficients, dimensionless;

[0244] When incorporating execution feedback, the system monitors rapid fluctuations in pipeline pressure and actual flow rate and applies additional deductions, defining a fluctuation judgment indicator:

[0245] ;

[0246] in Indicates whether a rapid fluctuation is currently occurring; 1 represents yes and 0 represents no, and it is dimensionless. The change in pipeline pressure between two adjacent control cycles is expressed in units consistent with the pressure (e.g., millimeters of mercury or pascals). and These are the periodic statistical pressure representative values ​​for the previous and the control cycles prior, respectively, with units consistent with the pressure. This represents the actual flow rate change between two adjacent control cycles, with the unit consistent with the flow rate (e.g., milliliters per second). and These are the periodic statistical flow representative values ​​for the previous and the control cycles before that, respectively; the units are consistent with the flow rate. and These are the threshold values ​​for judging rapid fluctuations in pressure and flow rate, respectively; the units are consistent with the corresponding physical quantities, and the symbols are... For indicator functions, the value is 1 if the condition is true and 0 otherwise. (Logical symbol) Indicates "or";

[0247] Update the risk budget tank value accordingly:

[0248] ;

[0249] in Indicates time The risk budget water tank value and its range are as follows: And it is dimensionless. The risk budget tank value and its range for the previous period are as follows: And it is dimensionless. Additional deduction factor for rapid fluctuations, dimensionless. The budget recovery coefficient is dimensionless. The low-risk threshold for risk intensity is dimensionless. For stable signal indication, a value of 1 indicates that pressure and flow fluctuations are below the corresponding threshold within a preset time window, while a value of 0 indicates instability and dimensionlessness. (Function) Indicates the real number Limited to Saturation operators within the interval Risk budget upper limit for water tank, dimensionless;

[0250] Nutritional target progress is updated proportionally to the ordered target infusion volume and the actual infusion volume, and does not increase when the pump is stopped or backflushed.

[0251] ;

[0252] in Indicates time The progress value and range of the nutrition target are as follows: And it is dimensionless. The progress value of the previous cycle, with a range of values. And it is dimensionless. This represents the forward infusion volume of the previous cycle (excluding backflushing volume), in units of volume (e.g., milliliters). The target infusion volume specified in the doctor's order, in units of volume (e.g., milliliters), symbol To select the operator with the smaller value, the condition is... This indicates the existence of a valid positive infusion;

[0253] Initialization is performed during the first control cycle. The dwell timer, cumulative pump stop duration, and cumulative backwash volume were set to zero to ensure that the closed-loop start state was consistent with the budget.

[0254] In this specific embodiment, S4 specifically refers to:

[0255] The system applies graded gating and viscosity constraints to candidate action levels based on the updated risk budget tank value and nutrient target progress value, and generates a pump speed reference trajectory, where the candidate action set is defined as follows: These correspond to deceleration, short stop, long stop, and backwashing, respectively, and are indexed using a unified timestamp. At the current control moment, the policy reward weight is first calculated for each action to reflect the trade-off between safety and nutrition goals, using a compact linear combination:

[0256] Perform measurement;

[0257] in For a moment action The strategy reward weights are dimensionless real numbers. For action At any moment The risk upper bound and the range of values ​​are: And it is dimensionless. For a moment The risk budget water tank value and its range are as follows: And it is dimensionless. For risk budget ceiling, dimensionless constant, For a moment The progress value and range of the nutrition target are as follows: And it is dimensionless. For action Non-negative weighting coefficients, dimensionless;

[0258] Then, based on the confidence boundaries of each action, hierarchical gating is implemented, and gating instructions are defined:

[0259] Actions that only meet the criteria of "risk upper bound not exceeding the threshold and coverage width not exceeding the upper limit" are given effective preferences;

[0260] in For action At any moment Gating results, dimensionless, For action At any moment The coverage width and value range are: And it is dimensionless. For action The risk upper bound gating threshold and dimensionless constant. For action Coverage width gating threshold, dimensionless constant, sign For indicator functions, the value is 1 if the condition is true and 0 otherwise. (Logical symbol) Indicates "and";

[0261] To suppress frequent handovers and the resulting intrinsic hysteresis, a stickiness and penalty are applied to handovers based on the resident timer in the system status register. First, a resident state indicator is defined:

[0262] And construct effective policy preferences ;

[0263] in For a moment The "Minimum stay time not met" indication is dimensionless. This is the dwell timer reading at the end of the previous control cycle, in seconds. This is the minimum dwell time parameter, in seconds. For a moment action Effective strategy preference, dimensionless real number, To switch between penalty coefficients and dimensionless constants, A penalty coefficient and a dimensionless constant are added to the viscosity. The type of action executed in the previous control cycle, with values ​​taken from a set. ,symbol It means "not equal to";

[0264] Based on this, select the unfiltered target action level. And prioritize keeping the maximum value when there are ties. To further reduce jitter, among which For a moment Unfiltered target action, value in set Operator Returns the argument that makes the input reach its maximum value;

[0265] When generating the pump speed reference trajectory, the current pump speed is used as the starting point, and the amplitude and time characteristics are satisfied, defining the relative time within the control cycle. With the current starting pump speed and maximum pump speed and minimum pump speed coupled with progress Segmentation construction reference:

[0266] ;

[0267] in For a moment The pump speed reference trajectory is in volumetric flow rate or corresponding pump speed units (such as milliliters per hour or rotational speed). For saturation operators, real numbers Limited to the range Inside, This represents the upper limit of the deceleration slope, measured in pump speed per second. The duration of a single control cycle, in seconds. These are the reference residence / duration parameters for short stops, long stops, and backwashing, respectively, in seconds. The reference amplitude for backwashing, with a range of values ​​as follows: And the unit is pump speed. This is the parameter for the maximum backwash amplitude, in units of pump speed;

[0268] The final output is the unfiltered target action level. pump speed reference trajectory Provided for use by the security shielding layer and the execution layer.

[0269] In this specific embodiment, S5 specifically includes:

[0270] The system uses unfiltered target actions pump speed reference trajectory As input, and in conjunction with the resident timer reading in the system status register Cumulative pump downtime With cumulative backwash volume Apply clinical safety boundaries, which include the maximum pump rate. Upper limit constraints, minimum pump stop time Detention constraints, maximum cumulative pump shutdown duration Cumulative constraints and maximum cumulative backwash volume The cumulative constraints are then used to construct a quadratic programming safety shield layer within the discretized time window of the current control cycle and introduce a control barrier function to ensure the forward invariance of the safety set, where the control cycle duration is expressed as... (Unit: seconds), the discretization step size is expressed as... (Unit: seconds), and let the number of discrete grid points be... (dimensionless), the first Each discrete time point is represented as (Unit: seconds, index) The pump speed sequence to be optimized is represented as follows: (Units are pump speed or volumetric flow rate, such as milliliters per hour), the reference sequence is represented as follows (Unit same) The objective function is the weighted squared deviation:

[0271] Optimize;

[0272] in The objective function value (dimensionless). For the first Non-negative weights at discrete time points (dimensionless);

[0273] To suppress excessively rapid changes in pump rate and meet clinical safety boundaries, linear inequalities and equality constraints are set, including a maximum pump rate constraint. (Unit same) ) and slope constraint (Unit: pump speed difference, upper limit of slope) The unit is pump speed per second (for all). This takes effect, simultaneously forcing both short and long stops to maintain a zero speed during their respective dwell times, thus minimizing the number of dwell steps. (dimensionless) and in Apply at time The equality constraints apply to all Effective, where the symbols sp and lp represent the short stop and long stop action categories respectively (dimensionless category symbols).

[0274] If the target action is backflushing, then apply direction and amplitude constraints during the backflushing period. and (Unit same) ),in The parameter for the maximum amplitude of backwashing (units are the same). To maintain motion stickiness and satisfy the condition that "the current motion remains unchanged until the dwell timer reaches the minimum dwell time", a dwell time not reached indicator is defined. (dimensionless, indicator function) (If the condition is true, return 1; otherwise, return 0). Only actions performed at this time are allowed. Consistent trajectory patterns enter the feasible region of optimization, where The type of action executed in the previous control cycle (dimensionless category symbol);

[0275] To unify the instantaneous and cumulative clinical boundaries and ensure the forward invariance of the safety set, a discrete form of the control barrier function is introduced, defining the safety potential function:

[0276] (Units correspond to safety margins for pump speed, time, and volume, respectively), where discrete cumulative quantities are... (Unit: seconds) and (Unit is volume, such as milliliters), and apply discrete barrier inequalities. (Dimensionless inequality form, For class The gain parameter (dimensionless) guarantees the forward invariance of the safe set for all... Effective, under the above constraints, for the objective function The safe pump speed sequence is obtained by solving the problem. (Unit same) ), and interpolate it into a safe pump speed trajectory in the continuous time domain. (Unit same) If it is feasible within the constraint set, then the safe and feasible action level is denoted as . (Dimensionless category symbol), otherwise, based on constraint priority, it degenerates into a conservative action that satisfies the dwell and cumulative boundaries, such as a short stop, and outputs the same result. The state variables are then updated to take effect in the next control cycle, and the dwell timer is updated in a segmented manner:

[0277] ;

[0278] The unit is seconds, and the logical symbol is... Indicates "and", the cumulative pump downtime is updated to (Unit: seconds), cumulative backwash volume updated to (Unit: volume);

[0279] The final output is the safety and mobility level. Safe pump speed trajectory And the updated dwell timer Cumulative pump downtime With cumulative backwash volume So that the execution layer can call it in step S6.

[0280] In this specific embodiment, S6 specifically refers to:

[0281] The system selects the corresponding control mode and generates pump control commands to drive the pump actuator based on the safety and mobility level and the safe pump speed trajectory. The safety and mobility level is denoted as follows: This indicates that the index in the control cycle is The action category to be executed at any given time is taken from a set. These refer to the dimensionless category symbols for deceleration, short stop, long stop, and backflushing, respectively. The safe pump speed trajectory is denoted as... This indicates that the index in the control cycle is relative time The safety reference pump speed is measured in units of pump speed or volumetric flow rate, such as milliliters per hour, or rotational speed. The relative time interval is denoted as [missing information]. ,in The duration of a single control cycle is measured in seconds.

[0282] To characterize the continuous-time dynamics of the pump and pipeline, a neural ordinary differential equation is used to track and predict the reference signal. Slope constraints and safety envelope constraints are applied during the prediction process, and the continuous state vector of the system is selected. The combined state of pump rotational inertia, pipeline compliance, and frictional hysteresis is represented as a real vector, with the unit depending on the meaning of the components. A neural parameter vector is defined. For the model's trainable parameters to be dimensionless, let the reference input... For units of pump speed, a compact form of the neural constant differential equation is used. Output mapping in The first derivative of the state with respect to time is the unit of... Each portion The dimensionless function symbol is mapped to a vector field parameterized by a neural network. For continuous time prediction of pump speed, the unit is the same as pump speed. To output the symbol of the dimensionless function;

[0283] To generate an instantaneous target pump speed that satisfies the slope constraint and does not exceed the safety envelope, the pump speed at the end of the previous control cycle is defined as follows: Units with the same pump speed and a maximum rate of change upper limit set to The units are pump speed per second and rate saturation operator. real numbers Limited to the range Simultaneously using the minimum value operator The smaller of the two values ​​is used to derive the segmented target pump speed formula:

[0284]

[0285] in For the instantaneous target pump speed unit, the pump speed in deceleration mode is guaranteed not to exceed the safety envelope through a minimum value constraint. In both short-stop and long-stop modes, maintain zero speed for the respective dwell time. In backwash mode, use a negative target and apply the maximum backwash amplitude. The limiting unit is the same as the pump speed. and These are the dwell time parameters for short and long stops, respectively, in seconds. The backwashing time parameter is in seconds;

[0286] To clearly define the infusion direction and duration, a direction indicator is defined:

[0287] ;

[0288] ;

[0289] in The infusion direction indicators for the current control cycle represent forward zero velocity and reverse dimensionless velocity, respectively. To maintain the duration parameter in seconds;

[0290] When generating control messages, the instantaneous target pump speed direction identifier and the duration of hold are combined into a pump control command. ,in The dimensionless message set of instruction triplets for the current control cycle is issued through the actuator interface to achieve the following: tracking of the downward reference curve in the deceleration mode; maintaining zero speed in the short stop and long stop modes; and reverse injection in the backwash mode within a limited time period. At the same time, due to the combined constraints of rate saturation and safety envelope, the instantaneous target pump speed is kept below the safe trajectory and overshoot is suppressed.

[0291] In this specific embodiment, S7 specifically refers to:

[0292] The system executes control modes such as deceleration, short stop, long stop or backwashing within the current control cycle according to the pump control command, and uses the system clock to time-mark and sample to form flow and pressure samples with unified timestamps to support closed-loop feedback calculation.

[0293] The current control cycle index is denoted as The dimensionless discrete index is denoted as and the control cycle duration is denoted as . The sampling interval is denoted as a fixed time length in seconds. The sampling step size is fixed in seconds, and the start time of the period is denoted as . The system clock time identifier is expressed in seconds, and the number of discrete sampling points is denoted as:

[0294] ;

[0295] Operator This represents the floor operation and is a dimensionless operator symbol;

[0296] The unified timestamp sequence is defined as:

[0297] ;

[0298] in To control the cycle The Each unified sampling time unit is seconds, index This refers to the dimensionless sampling sequence number;

[0299] At each uniform timestamp, collect and record the flow and pressure sensor readings as follows:

[0300] and ;

[0301] in For at any time The instantaneous volumetric flow rate is measured in volumes per second, such as milliliters per second. For at any time Instantaneous pipeline pressure is measured in units such as Pascals or millimeters of mercury. and The measurement mappings of flow rate and pressure are respectively represented by dimensionless function symbols, indicating the process by which the sensor returns a reading at a given time.

[0302] To calculate the injection volume for the current control cycle, a time integral is performed on the flow rate samples with a uniform timestamp, excluding the volume increment during the backwash period. A discrete integral expression is used:

[0303] ;

[0304] in To control the cycle The unit of forward infusion volume is volume, such as milliliters, symbol For real numbers The positive part operator is dimensionless. The backwash period indicator is dimensionless and takes a value of 1 when the action is backwashing and the sample time does not exceed the backwash period parameter; otherwise, it is 0.

[0305] To explicitly depict the backflushing window, you can make:

[0306] ;

[0307] in The safe and feasible action category for the current period is assigned a value to the set. For dimensionless category symbols, The backwash time parameter is in seconds, and the symbol is... The function returns 1 if the condition is true and 0 otherwise; logical symbol. Indicates "and";

[0308] At the end of the control period, periodic statistics are performed on samples with a uniform timestamp to obtain a representative value, and the actual flow is defined as the time average:

[0309] ;

[0310] in To control the cycle The actual flow unit is the same Volume per second;

[0311] Define pipeline pressure as the median over time:

[0312] ;

[0313] in To control the cycle The pipeline pressure unit is the same For pressure, operator The statistical function representing the median position of a sample after sorting by size is a dimensionless function.

[0314] Final output actual flow Pipeline pressure With infusion volume In the next control cycle, it serves as a closed-loop feedback to update the risk budget tank value and the nutrient target progress value to maintain stable system operation.

[0315] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0316] This invention addresses the technical problem of achieving aspiration risk response and closed-loop control of nutrition pumps within clinical safety boundaries. It proposes a combined algorithm integrating multimodal temporal risk estimation, strategy-level gating, quadratic programming safety shielding (combined with control barrier functions), and neural network constant differential equation execution. By weighting different quality signals by evidence strength and implementing conformal prediction calibration at the action level, a verifiable upper bound and coverage width of risk are obtained. This confidence information, along with the risk budget tank and nutritional target progress, jointly drives strategy rewards and gating, enabling the system to prioritize deceleration, short-term stoppage, long-term stoppage, or backflushing when risk increases, and to promptly resume infusion when risk decreases and budget allows. Subsequently, quadratic programming and control barrier functions are used to unify the maximum pump speed, minimum pump stoppage time, and cumulative constraints into the safety solution. Finally, neural network constant differential equations are used to match pump-pipe dynamics and slope constraints to generate traceable, non-overshoot control commands. The resulting technical effects are: improved reliability and interpretability of mis-aspiration identification, significant suppression of motion jitter and pump speed overshoot, ensuring the forward invariance of the safety set while ensuring the timely achievement of nutritional targets, and maintaining a robust closed loop in scenarios with signal quality fluctuations such as probe detachment and saturation.

[0317] Compared to existing technologies, this invention makes targeted improvements to the algorithm structure to address the specific problems:

[0318] First, it introduces action conditional conformal prediction calibration, which refines risk calibration to four types of actions: deceleration, short stop, long stop and backwashing. This directly provides a coverable upper bound and coverage width for graded response, avoiding unverifiable uncertainty caused by a single threshold or black box scoring.

[0319] Secondly, by balancing the dual objectives of risk budget tank and nutrient progress with the stickiness of the dwell timer and the switching penalty, an endogenous hysteresis is formed to suppress high-frequency switching and oscillation.

[0320] Third, a quadratic programming safety shield layer is adopted and a control barrier function is introduced to unify the instantaneous and cumulative clinical boundaries under the same optimization framework and ensure forward invariance, and project the reference trajectory into an executable safe pump speed trajectory.

[0321] Fourth, introduce a divine constant differential equation at the execution layer and impose slope constraints and a safety envelope to make reference-execution consistent and in line with physical realizability.

[0322] The aforementioned structural improvements work synergistically to form a closed-loop chain that is "detectable, decision-making, provably safe, and stably executable," further reducing false alarms and false negatives, decreasing the number of switching operations and peak overshoot, and improving usability and reliability in complex clinical scenarios.

Claims

1. A method for identifying aspiration and implementing closed-loop nutrient pump based on strategy optimization, characterized in that, include: S1. Collect and preprocess bedside multimodal time-series data to obtain respiratory waveforms, blood oxygen saturation, acoustic signals, body position signals, pump line pressure and pump line flow rate with a unified timestamp; S2. Calculate the probability of accidental aspiration and the strength of evidence at the current moment based on the six types of signals, and perform conformal prediction calibration of action conditionalization at the action level to obtain the risk upper bound and coverage width of each action level; S3. Based on the probability of aspiration risk, the strength of evidence, and the risk upper bound and coverage of each action level, and combined with the execution feedback of the previous control cycle, update the risk budget tank value and the nutritional target progress value. S4. Based on the updated risk budget tank value and nutrient target progress value, and combined with the risk upper bound and coverage width of each action level, hierarchical gating is implemented, and action stickiness constraints and switching penalties are applied according to the dwell timer in the system status register to generate unfiltered target action level and pump speed reference trajectory. S5. Based on the unfiltered target action level and pump speed reference trajectory, combined with the dwell timer, cumulative pump stop duration and cumulative backwash volume in the system status register, apply the clinical safety boundary, use the safety shield layer of quadratic programming and introduce the control barrier function to solve the safety constraints, obtain the safe action level and safe pump speed trajectory, and update the dwell timer, cumulative pump stop duration and cumulative backwash volume. S6. Based on the safe action level and safe pump speed trajectory, combined with the execution dynamics and slope limits of the pump and pipeline, the pump control command is generated using a normal differential equation model and output to the pump actuator. S7. The pump actuator completes the actual injection according to the pump control command, generates execution feedback, and uses it in the next control cycle to maintain closed-loop operation.

2. The method for aspiration identification and nutrient pump closed-loop based on strategy optimization according to claim 1, characterized in that, S1 specifically refers to: Based on the system clock, timestamps and clock deviation corrections are applied to respiratory waveforms, blood oxygen saturation, acoustic signals, body position signals, pump line pressure, and pump line flow to eliminate time drift in each channel and achieve time alignment. Denoising, drift removal, outlier removal, missing measurement interpolation, and resampling are performed on the six types of signals respectively to form a unified timestamp with a fixed sampling interval; Amplitude and unit calibrations are performed on the six types of signals to ensure consistency of physical quantities; And generate signal quality markers for the six types of signals to indicate states such as probe detachment, saturation, over-range, and strong interference; Output respiratory waveforms, blood oxygen saturation, acoustic signals, body position signals, pump line pressure, pump line flow rate, and corresponding signal quality markers with a unified timestamp.

3. The method for aspiration detection and nutrient pump closed-loop based on strategy optimization according to claim 1, characterized in that, S2 specifically refers to: Based on respiratory waveform, blood oxygen saturation, acoustic signal, body position signal, pump line pressure and pump line flow, time series features are extracted and consistency is checked so that various signals participate in risk estimation under a unified timestamp; The contribution weights of various signals based on signal quality labels are weighted according to the strength of evidence to form the strength of evidence, which is then combined with time-series features to calculate the probability of false absorption risk. In the candidate action level set, the action conditional conformal prediction calibration is performed for deceleration, short stop, long stop and backwash respectively. Based on the preset coverage target, the risk upper bound and coverage width are generated for deceleration, short stop and coverage width, long stop and coverage width, and backwash. Output the probability of accidental aspiration risk, the strength of evidence, the upper bound and coverage of risk for deceleration, the upper bound and coverage of risk for short stops, the upper bound and coverage of risk for long stops, and the upper bound and coverage of risk for backwashing.

4. The method for aspiration detection and nutrient pump closed-loop based on strategy optimization according to claim 1, characterized in that, S3 specifically refers to: The risk intensity value is calculated based on the probability of aspiration risk and the strength of evidence. A conservative risk upper bound is synthesized based on the risk upper bound for deceleration, the risk upper bound for short stops, the risk upper bound for long stops and the risk upper bound for backwashing. The conservative risk upper bound is used to determine the risk consumption rate of the current control cycle by combining the coverage width with a preset weight. Based on the actual flow rate, pipeline pressure, and injection volume generated in the previous control cycle, the risk budget tank value is deducted and restored. The deduction amount is calculated as the product of the risk consumption rate and the evidence strength, and additional deduction is applied when rapid fluctuations in pipeline pressure or actual flow rate are detected. The restoration amount takes effect according to a preset time constant when the risk strength value is below the threshold and the signal is stable. The updated risk budget tank value is limited to between zero and the budget upper limit. The nutritional target progress value is calculated based on the infusion volume generated in the previous control cycle, the target infusion volume specified in the doctor's order, and the target time window. It does not increase during the pump stoppage period when the actual flow rate is zero, but increases according to the ratio of the cumulative infusion volume to the target infusion volume during effective infusion. The first control cycle initializes the risk budget tank value to the budget limit, initializes the nutrient target progress value to zero, and initializes the residence timer, cumulative pump stop duration, and cumulative backwash volume to zero; Output the updated risk budget tank value and the updated nutrition target progress value.

5. The method for misaspiration identification and nutrient pump closed-loop based on strategy optimization according to claim 1, characterized in that, S4 specifically refers to: Based on the updated risk budget tank value and the updated nutrition target progress value, the strategy reward weights for deceleration, short stop, long stop and backwash in the candidate action level set are calculated respectively, and the reward distribution of each action level is formed. Based on the risk upper bound and coverage width of deceleration, short stop, long stop and backwashing, a graded gating threshold is set, and graded gating is executed so that the action level obtains effective strategy preference only when the corresponding risk upper bound does not exceed the gating threshold and the coverage width meets the preset coverage target. Based on the dwell timer in the system status register, action stickiness constraints and switching penalties are applied. When the dwell timer has not reached the minimum dwell time parameter, the switching of action level is restricted and the switching-related policy preferences are reduced to form endogenous hysteresis. Based on the strategy preference after hierarchical gating and viscosity constraint processing, the unfiltered target action level is selected, and a pump speed reference trajectory is generated based on the selected unfiltered target action level and the updated nutrient target progress value. The pump speed reference trajectory satisfies the preset change range and time characteristics, wherein the deceleration corresponds to a downward slope-limited curve, short stop and long stop correspond to zero speed holding curves during their respective residence time, and backwash corresponds to the reverse infusion curve during the backwash period. Output unfiltered target motion level and pump speed reference trajectory.

6. The method for aspiration detection and nutrient pump closed-loop based on strategy optimization according to claim 1, characterized in that, S5 specifically refers to: Based on the unfiltered target action level and pump speed reference trajectory, combined with the resident timer, cumulative pump stop duration and cumulative backflushing volume in the system status register, clinical safety boundaries of maximum pump speed, minimum pump stop time, maximum cumulative pump stop duration and maximum cumulative backflushing volume are applied to construct a quadratic programming problem and introduce a control barrier function to ensure the forward invariance of the safety set; The goal of the quadratic programming is to minimize the weighted square deviation of the safe pump speed trajectory from the pump speed reference trajectory while satisfying the above constraints, and to suppress the excessively rapid changes in pump speed by limiting the slope. The constraints include: the pump speed does not exceed the maximum pump speed; for short and long stops, the pump speed is kept at zero for the respective stop time to meet the minimum stop time; the current action level remains unchanged when the dwell timer has not reached the minimum dwell time; the cumulative stop time does not exceed the maximum cumulative stop time; the cumulative backwash volume does not exceed the maximum cumulative backwash volume; and the control barrier function keeps the system state from leaving the safe set. The safe action level and safe pump speed trajectory are obtained by solving the problem, and the state variables are updated: when the safe action level is different from the action level of the previous control cycle and the dwell timer reaches the minimum dwell time, the dwell timer is reset to zero; otherwise, it is accumulated according to the control cycle duration. When the safe pump speed trajectory includes a zero-speed segment, the pump stop time is accumulated by adding the duration of the zero-speed segment. When the safe pump speed trajectory includes a reverse injection section, the cumulative backwash volume is added to the injection volume of the reverse injection section. Outputs the safe action level and safe pump speed trajectory, as well as the updated dwell timer, updated cumulative pump downtime, and updated cumulative backwash volume.

7. The method for aspiration identification and nutrient pump closed-loop based on strategy optimization according to claim 1, characterized in that, S6 specifically refers to: The control mode is selected based on the level of safe and feasible action, so that the deceleration corresponds to the downward tracking reference curve, the short stop and long stop correspond to the zero speed maintenance during their respective residence time, and the backwash corresponds to the reverse injection curve during the backwash period. Based on the safe pump speed trajectory as a reference signal, the execution dynamics of the pump and pipeline are continuously predicted using a neural network constant differential equation model. The execution dynamics cover the dynamic characteristics of pump rotational inertia, pipeline compliance and frictional hysteresis. In the prediction process, a slope limit is applied to constrain the rate of change of pump speed and suppress overshoot, and to ensure that the instantaneous target pump speed does not exceed the envelope of the safe pump speed trajectory. During the zero-speed maintenance phase of short and long stops, maintain zero speed until the end of the corresponding residence time. During the backwashing phase, generate the target pump speed curve for reverse infusion and maintain it during the backwashing period. The instantaneous target pump speed, forward or reverse indication, and holding time are combined into a pump control command and output to the pump actuator.

8. The method for aspiration detection and nutrient pump closed-loop based on strategy optimization according to claim 1, characterized in that, S7 specifically refers to: Based on the pump control command, execute the control mode of speed reduction, short stop, long stop or backwashing within the current control cycle; During execution, the system clock is used for time stamping, and the readings of the flow sensor and pressure sensor are continuously collected to form flow samples and pressure samples with the same timestamp. At the end of the control cycle, the flow samples with the same timestamp are integrated over time according to the sampling interval to generate the infusion volume, and the infusion volume is not included in the backwash period; At the end of the control cycle, periodic statistics are performed on flow samples and pressure samples with the same timestamp to generate actual flow and pipeline pressure. The system outputs actual flow rate, pipeline pressure, and injection volume, which are used to update the risk budget tank value and nutrient target progress value in the next control cycle and maintain closed-loop operation.

Images