Individualized liquid resuscitation decision support system for critical patients

Abstract

The invention relates to the technical field of clinical decision support and medical data intelligent analysis, in particular to a critical patient individualized liquid resuscitation decision support system, which is characterized in that a data acquisition unit acquires pulse pressure variation waveform and multi-dimensional physiological data of each stroke volume; a volume benefit analysis channel of the dynamic decision-making unit segments a respiratory cycle, extracts pulse pressure fluctuation characteristics, outputs a per-stroke volume increase probability value through a machine learning classifier, an organ injury early warning channel calculates a blood lactic acid and creatinine clearance rate change rate, and generates an acute kidney injury risk sign in combination with inferior vena cava image verification; the execution control unit sends an infusion control or limit instruction according to the combined state; and the feedback optimization unit adjusts the weight of the model according to the actual value of blood lactic acid removal, and recalculates the probability value, so that individualized precise resuscitation, balance capacity benefit and organ protection are realized, and the treatment safety and effectiveness are improved.

Description

Technical Field

[0001] This invention relates to the field of clinical decision support and intelligent analysis of medical data, and more specifically, to a decision support system for individualized fluid resuscitation for critically ill patients. Background Technology

[0002] Clinical decision support and intelligent analysis of medical data is an important technology, specifically applied to the decision-making process of fluid resuscitation in critically ill patients. Its core is to achieve individualized fluid control by accurately assessing volume benefits and organ damage risks, thus meeting the core needs of critically ill patients for safety and effectiveness. In fluid resuscitation for critically ill patients, individual responses to fluid volume vary, and excessive fluid administration can lead to organ damage such as acute kidney injury. Traditional decision-making methods struggle to integrate multi-dimensional physiological data such as pulse pressure variability and blood lactate concentration in real time, making it impossible to accurately predict the potential for stroke volume improvement and organ damage risks. This can result in insufficient fluid administration affecting resuscitation outcomes or excessive fluid administration causing complications. To address this technical challenge, we have developed an individualized fluid resuscitation decision support system for critically ill patients. Summary of the Invention

[0003] The purpose of this invention is to provide a personalized fluid resuscitation decision support system for critically ill patients to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, a personalized fluid resuscitation decision support system for critically ill patients is provided, including:

[0005] The data acquisition unit connects to medical equipment in real time to acquire pulse pressure variability waveforms, continuous stroke volume measurements, and inferior vena cava images, and simultaneously extracts time-series data of blood lactate concentration and creatinine clearance rate values ​​from the medical information system.

[0006] The dynamic decision-making unit includes a volume benefit analysis channel and an organ damage early warning channel that operate in parallel. The volume benefit analysis channel segments the pulse pressure variability waveform into respiratory cycles, extracts the difference feature between the maximum and minimum pulse pressure in each respiratory cycle, inputs the difference feature into a machine learning classifier trained on historical volume response positive cases, and outputs a stroke volume increase probability value. The organ damage early warning channel calculates the hourly change rate of blood lactate concentration and the hourly change rate of creatinine clearance rate in real time. When the rate of increase of blood lactate concentration exceeds a set multiple of the rate of decrease of creatinine clearance rate, an acute kidney injury risk marker is generated.

[0007] The execution control unit generates control commands based on the combined state of the stroke volume increase probability value and the acute kidney injury risk markers, including:

[0008] (a) When the probability of increased stroke volume is higher than the preset probability threshold and no acute kidney injury risk marker is generated, a crystalloid infusion rate control signal is sent to the smart infusion pump;

[0009] (b) When the probability of increased stroke volume is lower than the preset probability threshold or there are signs of acute kidney injury, a command to limit infusion is sent to the intelligent infusion pump and the vasoactive drug infusion module is activated simultaneously to stop the infusion.

[0010] The feedback optimization unit continuously collects the actual change in stroke volume and the actual value of blood lactate clearance after the infusion is stopped. When the actual value of blood lactate clearance does not reach the expected target of the dynamic decision unit, the weight coefficient of the difference feature in the machine learning classifier is adjusted and the stroke volume increase probability value is recalculated.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0012] This invention comprehensively acquires multi-dimensional physiological data on pulse pressure variability waveforms and blood lactate concentration through a data acquisition unit. The volume benefit analysis channel of the dynamic decision-making unit, through respiratory cycle segmentation and feature extraction, accurately outputs the probability value of stroke volume increase through a gradient boosting decision tree model. The organ injury early warning channel, combined with the rate of change of blood lactate and creatinine clearance and inferior vena cava imaging verification, promptly generates acute kidney injury risk markers. The execution control unit intelligently outputs infusion control commands based on the combined state. The feedback optimization unit adjusts the model weights through treatment effect feedback, continuously improving individualized adaptability. This enables precise benefit assessment and organ risk control of fluid resuscitation in critically ill patients, avoiding insufficient or excessive infusion, improving treatment safety and effectiveness, and providing reliable intelligent support for clinical decision-making. Attached Figure Description

[0013] Figure 1 This is an overall block diagram of the present invention.

[0014] The meanings of the labels in the diagram are as follows:

[0015] 1. Data acquisition unit; 2. Dynamic decision-making unit; 3. Execution control unit; 4. Feedback optimization unit. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] This invention provides a personalized fluid resuscitation decision support system for critically ill patients. Please refer to [link / reference]. Figure 1 As shown, it includes:

[0018] The data acquisition unit connects to medical equipment in real time to acquire pulse pressure variability waveforms, continuous stroke volume measurements, and inferior vena cava images, and simultaneously extracts time-series data of blood lactate concentration and creatinine clearance rate values ​​from the medical information system.

[0019] The dynamic decision-making unit includes a volume benefit analysis channel and an organ damage early warning channel that operate in parallel. The volume benefit analysis channel segments the pulse pressure variability waveform into respiratory cycles, extracts the difference feature between the maximum and minimum pulse pressure in each respiratory cycle, inputs the difference feature into a machine learning classifier trained on historical volume response positive cases, and outputs a stroke volume increase probability value. The organ damage early warning channel calculates the hourly change rate of blood lactate concentration and the hourly change rate of creatinine clearance rate in real time. When the rate of increase of blood lactate concentration exceeds a set multiple of the rate of decrease of creatinine clearance rate, an acute kidney injury risk marker is generated.

[0020] The execution control unit generates control commands based on the combined state of the stroke volume increase probability value and the acute kidney injury risk markers, including:

[0021] (a) When the probability of increased stroke volume is higher than the preset probability threshold and no acute kidney injury risk marker is generated, a crystalloid infusion rate control signal is sent to the smart infusion pump;

[0022] (b) When the probability of increased stroke volume is lower than the preset probability threshold or there are signs of acute kidney injury, a command to limit infusion is sent to the intelligent infusion pump and the vasoactive drug infusion module is activated simultaneously to stop the infusion.

[0023] The feedback optimization unit continuously collects the actual change in stroke volume and the actual value of blood lactate clearance after the infusion is stopped. When the actual value of blood lactate clearance does not reach the expected target of the dynamic decision unit, the weight coefficient of the difference feature in the machine learning classifier is adjusted and the stroke volume increase probability value is recalculated.

[0024] The volume benefit analysis channel segments the pulse pressure variability waveform into respiratory cycles. By receiving the ventilator airflow waveform signal in real time, it identifies the time-domain positions of the start and end points of the inspiratory phase and cuts the pulse pressure variability waveform into independent respiratory cycle segments based on the time-domain positions.

[0025] For the extraction of difference features, within a segmented single respiratory cycle, a peak detection algorithm is used to locate the maximum and minimum pulse pressure points. The pulse pressure fluctuation feature is generated by calculating the pressure amplitude difference between the maximum and minimum pulse pressure points. When an abnormal peak is detected in the pulse pressure fluctuation feature, interpolation of adjacent cycle data is automatically used for compensation.

[0026] The machine learning classifier was constructed using pulse pressure fluctuation features and corresponding stroke volume increase labels of historical volume response positive cases. Difference features were selected by ranking feature importance, and a gradient boosting decision tree model was trained to output the stroke volume increase probability value. Historical volume response positive cases were defined as those whose stroke volume increase reached the target after the fluid loading test.

[0027] The hourly rate of change of blood lactate concentration in the organ damage early warning channel was calculated by linearly fitting the time series data of blood lactate concentration using the sliding window method, and the slope per unit time was extracted as the rate of increase of blood lactate concentration. The rate of change of creatinine clearance was calculated by the difference method to calculate the decreasing gradient of two consecutive detection values. Both were standardized in hours.

[0028] The generation of acute kidney injury risk markers is based on the rate of increase in blood lactate concentration and the rate of decrease in creatinine clearance. The diameter and respiratory variability of the inferior vena cava image are introduced as validation parameters. When the rate of increase in blood lactate concentration exceeds a set multiple of the rate of decrease in creatinine clearance and the variability is below a critical threshold, the acute kidney injury risk marker is activated.

[0029] The crystalloid infusion rate control signal is generated by setting a base infusion rate based on the stroke volume increase probability value and adjusting the rate with a fixed step value based on the increase feedback of the real-time continuous stroke volume measurement value, while monitoring the acute kidney injury risk markers.

[0030] The activation of the infusion command is restricted. Upon receiving an acute kidney injury risk indicator, the infusion rate is immediately reduced to the maintenance level and the infusion module is activated. When the stroke volume increase probability value is lower than the preset probability threshold but no acute kidney injury risk is triggered, a gradual rate reduction strategy is adopted to simultaneously activate the infusion module.

[0031] The weighting coefficients are adjusted based on the deviation ratio between the actual change in stroke volume after infusion and the expected target, thereby reducing the decision weight of the difference feature in the machine learning classifier.

[0032] The probability of stroke volume increase is recalculated, and the feature space of the gradient boosting decision tree model is reconstructed using the adjusted weight coefficients. The pulse pressure fluctuation features extracted from the current respiratory cycle segment are input into the iteratively updated gradient boosting decision tree model for real-time probability calculation.

[0033] Further explanation is needed regarding the volume benefit analysis channel of the dynamic decision-making unit. To accurately extract respiratory cycle-related features from the pulse pressure variability waveform, which is significantly affected by respiration, the differences in pulse pressure fluctuation characteristics across different respiratory stages directly reflect the patient's volume response status. Therefore, respiratory cycle segmentation of the pulse pressure variability waveform is necessary to ensure the targeted nature of subsequent feature extraction. The volume benefit analysis channel segments the pulse pressure variability waveform according to the respiratory cycle. The pulse pressure variability waveform is a continuous curve reflecting pulse pressure fluctuations over time, acquired by the data acquisition unit from the blood pressure monitoring device. Pulse pressure is the difference between systolic and diastolic blood pressure, and its variability is closely related to the patient's blood volume status. Respiratory cycle segmentation divides the continuous pulse pressure variability waveform into independent segments according to the patient's natural respiratory cycle. Each segment corresponds to one complete inhalation and exhalation process. This avoids interference between waveform features from different respiratory stages. The starting point of the inspiratory phase and the expiratory phase are identified by real-time reception of the ventilator airflow waveform signal. The time-domain location of the bottleneck is determined by the ventilator airflow waveform signal, which is a real-time output signal from the ventilator reflecting changes in airflow within the airway. The airflow signal rises during inspiration and falls during expiration. This signal accurately marks key points in the respiratory process. The inspiratory phase start point is the moment the airflow signal begins to rise from the baseline, marking the initiation of the patient's inspiration. The expiratory phase end point is the moment the airflow signal falls back to the baseline and remains stable, marking the completion of one respiratory cycle. The time-domain location refers to the specific time coordinates of these two key points on the time axis. During identification, the volume benefit analysis channel receives the data in real-time through a dedicated data interface. The ventilator airflow waveform signal is analyzed by setting a baseline threshold (calibrated using the average airflow during the patient's quiet breathing). When the airflow signal rises from below the baseline threshold and exceeds it by 10%, this moment is recorded as the start of the inspiratory phase. When the airflow signal falls back to the baseline threshold from its peak and remains stable for 50 milliseconds without significant fluctuations, this moment is recorded as the end of the expiratory phase. Accurate identification of these two key points is ensured. The pulse pressure variability waveform is segmented into independent respiratory cycle segments based on its temporal location. Each independent respiratory cycle segment refers to the portion of the pulse pressure variability waveform between the start of the inspiratory phase and the corresponding end of the expiratory phase. Each segment corresponds completely to the pulse pressure change during a single respiratory cycle. During segmentation, the volume benefit analysis channel marks the segmentation on the continuous pulse pressure variability waveform based on the identified temporal positions of the inspiratory phase start point and the expiratory phase end point. Starting from the inspiratory phase start point and ending at the next expiratory phase end point, an independent segment is formed. At the same time, abnormal spike signals caused by equipment interference in the segment are automatically removed (signals exceeding 3 times the mean are judged as abnormal and removed by comparing with the amplitude of adjacent waveforms). This ensures that the waveform data of each respiratory cycle segment truly reflects the patient's physiological state, laying the foundation for subsequent differential feature extraction.

[0034] After completing the respiratory cycle segmentation of the pulse pressure variability waveform, each independent segment completely corresponds to the pulse pressure change during one respiration. The difference feature, as a core indicator reflecting the patient's volume response, needs to be accurately extracted from these segments. The specific implementation method is as follows:

[0035] Extracting differential features is a crucial step in volume benefit analysis. Its core is capturing the maximum fluctuation amplitude of pulse pressure within a single respiratory cycle. Within a segmented single respiratory cycle (obtained earlier by time-domain location segmentation), the pulse pressure variability waveform segment corresponding to one inspiratory and expiratory cycle is used. The segment length is consistent with the patient's respiratory cycle, ensuring the extracted features focus on the impact of a single breath on pulse pressure. A peak detection algorithm is employed to locate the maximum and minimum pulse pressure points. This algorithm, optimized for pulse pressure waveform features, accurately distinguishes between pulse pressure extremes caused by physiological fluctuations and abnormal points caused by equipment interference. The maximum pulse pressure point refers to the sample with the highest pulse pressure value within a single respiratory cycle segment. The pulse pressure point, typically occurring in the mid-expiratory phase, is the sampling point with the lowest pulse pressure value within a segment, usually located in the mid-inspiratory phase. For localization, the pulse pressure data within the respiratory cycle segment is first smoothed using a moving average method with a window size of 5 sampling points to eliminate high-frequency noise interference. Then, an algorithm iterates through all smoothed sampling points, setting an amplitude change threshold (10% of the average pulse pressure within the segment). When the amplitude change of consecutive sampling points exceeds this threshold and the amplitude of subsequent sampling points begins to decrease, it is determined to be the maximum pulse pressure point. When the amplitude change of consecutive sampling points is below this threshold and the amplitude of subsequent sampling points begins to increase, it is determined to be the minimum pulse pressure point. This ensures that the localization error between the two extreme points does not exceed one sampling cycle. The maximum pulse pressure point and the minimum pulse pressure point are then calculated. The difference in pressure amplitude between pressure points generates a pulse pressure fluctuation characteristic quantity. This pressure amplitude difference is the difference between the pulse pressure values ​​at the maximum and minimum pulse pressure points. This difference directly reflects the range of pulse pressure fluctuation within a single respiratory cycle. The pulse pressure fluctuation characteristic quantity is a quantified indicator obtained by standardizing this difference. Standardization involves dividing by the patient's recent mean pulse pressure to eliminate the influence of individual baseline pulse pressure differences, objectively reflecting the relative intensity of pulse pressure fluctuations and providing a basis for subsequent volume response assessment. During calculation, the maximum and minimum pulse pressure values ​​are first read, and their absolute difference is directly calculated. Then, the patient's mean pulse pressure over the past 30 minutes is retrieved, and the absolute difference is divided by the mean to obtain the standardized pulse pressure fluctuation characteristic quantity, ensuring the comparability of the characteristic quantity. When an abnormal peak is detected in the pulse pressure fluctuation characteristic, automatic compensation is performed using interpolation of data from adjacent cycles. An abnormal peak refers to a pulse pressure fluctuation characteristic in a respiratory cycle that significantly exceeds the normal range. The criterion for judgment is that the characteristic is greater than three times the average of the characteristic values ​​of the three adjacent respiratory cycles. Such abnormalities are often caused by non-physiological factors such as transient equipment interference, patient coughing, or agitation. Interpolation of data from adjacent cycles involves using the pulse pressure fluctuation characteristic of two normal respiratory cycles before and after the abnormal cycle to calculate a reasonable value to replace the abnormal peak through linear interpolation. Compensation replaces the abnormal peak, ensuring the continuity and reliability of the characteristic value sequence. During detection, the pulse pressure fluctuation characteristic of the current cycle is compared in real time with the average of the three adjacent cycles. If an abnormal peak is determined, compensation is performed.The compensation mechanism is immediately activated. Normal features are extracted from the two cycles preceding and following the abnormal cycle. Linear interpolation is performed using time weights (normal cycles closer to the abnormal cycle have higher weights) to obtain a compensation value. This compensation value replaces the abnormal peak value. Simultaneously, the abnormal event is marked and recorded in the system log, providing a reference for subsequent data traceability and ensuring that the features input to the machine learning classifier accurately reflect the patient's physiological state.

[0036] After extracting and compensating for pulse pressure fluctuation characteristics, these characteristics need to be transformed into stroke volume increase probability values ​​that reflect the patient's benefit from fluid resuscitation through a specially constructed machine learning classifier. The specific implementation method is as follows:

[0037] The construction of a machine learning classifier is the core of achieving quantitative assessment of volume benefit. Its core logic is to learn the correlation between pulse pressure variability characteristics and stroke volume increase through historical case data. The construction process first uses the pulse pressure variability characteristics and corresponding stroke volume increase labels of historical volume-responsive positive cases. Historical volume-responsive positive cases refer to critically ill patients selected from the hospital's intensive care unit's historical diagnosis and treatment database who have received fluid resuscitation and have a good volume response. Each case includes complete pulse pressure variability monitoring data, continuous stroke volume measurement data, and treatment records. The pulse pressure variability characteristics are the standardized values ​​of the maximum pulse pressure fluctuation within a single respiratory cycle extracted above and compensated for for anomalies. Multiple characteristics are extracted from each case according to the respiratory cycle to form a feature sequence. Stroke volume increase labels are used to label the treatment effect corresponding to the pulse pressure fluctuation feature for each case, i.e., the actual increase in stroke volume after the fluid overload test. This serves as the target output for the classifier's learning. Difference features are selected by prioritizing features based on their contribution to the model's prediction results. This prioritization is achieved using a recursive feature elimination algorithm, which effectively removes redundant features and retains key information. Difference features refer to the core features among the pulse pressure fluctuation features that have the highest correlation with stroke volume increase, not all extracted features. During selection, the pulse pressure fluctuation features of historical cases are first combined with their corresponding stroke volume increase labels to form a training dataset. Then, the algorithm calculates the contribution weight of each feature to the stroke volume increase prediction and selects the features according to their weights. The features are sorted from highest to lowest weight, and a feature retention ratio is set. Features with the highest weight are retained as the final input difference features to the model. This reduces the computational load of the model and improves prediction accuracy. The gradient boosting decision tree model outputs the stroke volume increase probability value. The gradient boosting decision tree model is an ensemble learning model that constructs multiple weak decision trees and iteratively optimizes them in the gradient descent direction. It can accurately fit the nonlinear relationship between features and target values ​​and is suitable for the large individual differences among critically ill patients. The stroke volume increase probability value is the probability output by the model that the patient's stroke volume will reach the expected increase after receiving fluid resuscitation. The value ranges from 0 to 1. The closer it is to 1, the higher the probability of volume benefit. During training, the selected difference features and stroke volume increase labels are input into the model, and the number of decision trees is set. Hyperparameters such as (typically 100 trees) and maximum tree depth (typically 5 layers) are used to minimize the error between the predicted value and the actual stroke volume increase label through iterative training. After training, the model can output the corresponding stroke volume increase probability value based on the input difference features. A 5-fold cross-validation method is used to evaluate the model's performance, ensuring that the model's prediction accuracy is not less than 85%, meeting clinical decision-making needs. Historically positive volume response cases are defined as those whose stroke volume increase reaches the target after a fluid overload test. The fluid overload test is a standard method for clinically assessing a patient's volume response, which involves infusing a certain volume (typically 500 ml) of crystalloid solution into the patient over a short period (usually 15-30 minutes) and observing changes in stroke volume. Achieving the target stroke volume increase means that after the fluid overload test...A stroke volume increase exceeding 10% of the baseline value is a universally accepted criterion for a positive volume response in critical care medicine. This standard ensures a unified and scientific basis for screening historical cases, preventing model training bias due to inconsistent criteria. During case screening, historical fluid overload test records are retrieved from the clinical database, and cases with a stroke volume increase of ≥10% are identified as historically positive volume responses. Cases with severe arrhythmias, acute heart failure, or other conditions affecting volume response assessment are excluded, ensuring the reliability and representativeness of the training dataset.

[0038] While the volume benefit analysis channel operates in parallel, the organ injury early warning channel needs to monitor in real time whether the patient is at risk of organ injury due to fluid resuscitation. The core is to determine the risk trend by calculating the dynamic rate of change of blood lactate concentration and creatinine clearance rate. The specific implementation method is as follows:

[0039] Calculating the hourly rate of change of blood lactate concentration in the organ damage early warning channel is a crucial step in risk assessment. Blood lactate concentration is a core indicator reflecting tissue hypoxia and metabolic disorders; a sustained increase often suggests insufficient organ perfusion or functional impairment. The hourly rate of change quantifies how quickly blood lactate concentration changes over time, accurately capturing its dynamic trend and offering greater early warning value than values ​​at a single time point. A sliding window method is used to linearly fit the time-series data of blood lactate concentration. The sliding window method involves setting a fixed-length time window and processing the time-series data in segments by continuously sliding the window. This effectively smooths short-term fluctuations and highlights the overall trend. Here, based on the clinical monitoring frequency, the window length is set to 1 hour, and the sliding step is 15 minutes, ensuring timely response to changes while avoiding oversensitivity. The time-series data of blood lactate concentration consists of blood lactate test values ​​extracted from the medical information system by the data acquisition unit and arranged in chronological order, typically recorded every 15-30 minutes. Linear fitting uses mathematical methods to find a straight line that best fits the data points within the window; the slope of this line represents the blood lactate concentration within the corresponding time period. For the rate of change of concentration, during the fitting process, the least squares method is used to calculate the fitted straight line for the blood lactate concentration data within each sliding window to eliminate the influence of accidental measurement errors and ensure that the slope can truly reflect the trend of concentration change. The slope per unit time is extracted as the rate of increase of blood lactate concentration. The slope per unit time is the slope value of the fitted straight line. Its positive or negative sign represents the increase or decrease of concentration, and the absolute value represents the speed of change. The rate of increase of blood lactate concentration refers to the numerical increase of blood lactate concentration per unit time. When the slope is positive, the slope value is directly used as the rate of increase, and when the slope is negative, the rate of increase is determined to be 0. Focusing on the risk-related concentration increase trend, the slope value obtained by linear fitting is converted to hourly units during extraction. For example, the slope of concentration change within 30 minutes in the window is converted to the rate of change per hour to ensure the consistency of the time dimension of the indicator and facilitate subsequent comparison with other rate indicators. The rate of change of creatinine clearance rate is calculated by the difference method to calculate the decreasing gradient of two consecutive detection values. Creatinine clearance rate is a core indicator for assessing kidney function, reflecting the kidney's ability to clear metabolic waste in the body. Its decrease suggests that kidney function may be impaired.The differential method is a simple and efficient way to obtain the change in creatinine clearance rate by calculating the difference between two adjacent test values. It is suitable for the relatively low-frequency detection characteristics of creatinine clearance rate. The decreasing gradient between two consecutive test values ​​refers to the difference between the creatinine clearance rate value of the later test and the value of the earlier test. If the difference is negative, it means that the clearance rate is decreasing. The larger the absolute value of the gradient, the faster the decrease. In the calculation, the creatinine clearance rate test data is first arranged in chronological order to ensure the temporal continuity of the data. Then, the difference between two adjacent data is calculated one by one to obtain the change in clearance rate within each time interval, which is then converted into the hourly decreasing gradient. For example, if the interval between two tests is 1 hour, the difference is directly used as the hourly decreasing gradient; if the interval is 2 hours, the difference is used as the hourly decreasing gradient. Dividing by 2 yields the hourly gradient. Both are standardized to an hourly unit. Standardization involves converting the rate of increase in blood lactate concentration and the rate of decrease in creatinine clearance into indicators with an hourly time unit. This eliminates the influence of differences in detection intervals and ensures the comparability of the two indicators over time, laying the foundation for subsequent rate comparison and risk assessment. During standardization, the changes are converted according to the time interval between two tests. The final output of the rate of increase in blood lactate concentration is in hourly concentration units, and the rate of decrease in creatinine clearance is in hourly clearance rate units. This ensures that the values ​​of both directly reflect the degree of change within an hour, facilitating the subsequent setting of multiple relationship thresholds for risk assessment.

[0040] After calculating the hourly rate of change of hemorrhagic lactate concentration and the hourly rate of change of creatinine clearance in the organ injury early warning channel, relying solely on these two rate indicators to determine the risk of acute kidney injury is still somewhat one-sided. It is necessary to introduce intuitive physiological indicators reflecting the patient's volume status for cross-validation to improve the accuracy of risk marker generation. The specific implementation method is as follows:

[0041] The generation of acute kidney injury risk markers is the core output of the organ injury early warning system. Its purpose is to promptly capture the potential risk of kidney injury during fluid resuscitation. Based on the rate of increase in blood lactate concentration and the rate of decrease in creatinine clearance, these two rate indicators reflect risk trends from the perspectives of metabolic disturbance and changes in renal function, respectively. The rate of increase in blood lactate concentration suggests insufficient tissue perfusion or metabolic abnormalities, while the rate of decrease in creatinine clearance directly reflects the degree of renal function impairment. The combination of these two indicators provides a basis for risk assessment, but further validation through physiological indicators is needed to avoid misjudgment based on a single indicator. The diameter and respiratory variability of the inferior vena cava image are introduced as a validation parameter. The inferior vena cava image is acquired by the data acquisition unit from ultrasound equipment. Imaging data reflecting the central venous volume status of patients is used. The change in the diameter of the inferior vena cava with respiration can directly reflect whether the blood volume is sufficient. Diameter respiratory variability refers to the ratio of the change in the diameter of the inferior vena cava during inspiration and expiration. The calculation formula is (maximum diameter during expiration - minimum diameter during inspiration) / maximum diameter during expiration. This indicator is closely related to the patient's volume response. The variability increases when the blood volume is insufficient and decreases when the blood volume is excessive or renal function is impaired. Validation parameters are used to corroborate the results of the first two rate indicators. By working together, the reliability of risk assessment is improved. When introduced, the system receives inferior vena cava imaging data in real time, performs image preprocessing (including noise reduction and edge detection), and automatically identifies... The diameter of the inferior vena cava during inspiration and expiration is used to calculate the respiratory variability of diameter using a formula. This parameter is ensured to be consistent with the time dimension of the two rate indicators, both being updated synchronously in hourly increments. When the rate of increase in blood lactate concentration exceeds a set multiple of the rate of decrease in creatinine clearance and the variability is below a critical threshold, an acute kidney injury risk marker is activated. The set multiple is a risk assessment threshold determined through extensive clinical data, typically set at 2 times. That is, when the rate of increase in blood lactate concentration reaches 2 times or more of the rate of decrease in creatinine clearance, it indicates a significantly increased risk of metabolic disturbances and kidney function impairment. The critical threshold is the risk cutoff value for the respiratory variability of diameter, clinically validated and set at 10%. A value below this threshold indicates... The small change in the diameter of the vena cava with respiration suggests possible hypervolemia or decreased renal excretion function, further supporting the risk of acute kidney injury. The acute kidney injury risk marker is a digital signal (active high) generated by the system to trigger subsequent fluid restriction instructions. Upon activation, the system compares three indicators simultaneously: first, it determines whether the rate of increase in blood lactate concentration exceeds a set multiple of the rate of decrease in creatinine clearance; second, it checks whether the respiratory variability in the diameter is below a critical threshold. Only when both conditions are met is the risk of acute kidney injury determined, the risk marker is immediately activated, and transmitted to the execution control unit. This ensures that the generation of the risk marker is based on objective data and has undergone multi-dimensional verification, avoiding false triggering or missed triggering.

[0042] After the dynamic decision-making unit outputs the probability value of increased stroke volume and confirms that no risk marker for acute kidney injury has been generated, the execution control unit needs to generate a precise crystalloid infusion rate control signal to maximize volume benefit while avoiding the risk of over-infusion. The specific implementation method is as follows:

[0043] The generation of the crystalloid infusion rate control signal is a core operation in performing fluid resuscitation. This signal is a digital signal used to instruct the intelligent infusion pump to adjust the infusion rate. The signal format is compatible with the control interface of the intelligent infusion pump, directly driving the pump to deliver crystalloids at a set rate. These fluids can rapidly replenish blood volume and are commonly used in fluid resuscitation for critically ill patients. The baseline infusion rate is set based on the stroke volume increase probability value. This probability value, output by a machine learning classifier, represents the likelihood of an increase in stroke volume after the patient receives infusion, ranging from 0 to 1. The baseline infusion rate serves as the rate benchmark in the initial stage of infusion and needs to be matched with the probability value to match the volume replenishment intensity, avoiding rates that are too high or too low that could negatively impact resuscitation effectiveness. During setting, the system incorporates a rate-probability mapping table, calibrated through clinical trials. When the stroke volume increase probability value is between 0.7 and 1.0 (high benefit probability), the baseline infusion rate is set at [value missing] per hour. The basic infusion rate is set at 500-600 ml per hour. When the probability value is between 0.5 and 0.7 (medium probability of benefit), the rate is set at 300-500 ml per hour. When the probability value is below 0.5 (low probability of benefit), the basic infusion rate is not started (this needs to be determined in conjunction with other conditions). This ensures that the basic rate is accurately matched with the patient's volume response potential. The rate is adjusted with a fixed step value based on the increase feedback of the real-time continuous stroke volume measurement value. The real-time continuous stroke volume measurement value is the stroke volume data obtained by the data acquisition unit from the cardiac output monitoring device and updated once per second. It can reflect the impact of infusion on cardiac function in real time. The increase feedback refers to the ratio of the difference between the current stroke volume measurement value and the baseline value before infusion. For example, if the stroke volume increases from 50 ml to 60 ml after infusion, the increase is 20%. The fixed step value is the fixed range of each adjustment of the infusion rate, set at 50 ml per hour. This range can quickly respond to changes in the increase while avoiding cyclical fluctuations caused by sudden rate changes. During adjustment, the system calculates the stroke volume increase every 5 minutes: if the increase reaches 10% or more (achieving the expected benefit target), the current rate is maintained; if the increase is less than 5% (insufficient benefit), the rate is increased by a fixed step value; if the increase exceeds 20% (approaching volume saturation), the rate is decreased by a fixed step value, forming a closed-loop feedback adjustment mechanism. Simultaneously, the system monitors the acute kidney injury risk markers, which are real-time risk signals output by the organ injury early warning channel. Their status (activated / inactive) directly determines whether to continue infusion. During monitoring, the execution control unit establishes a real-time data link with the organ injury early warning channel, acquiring the risk marker status every second. If the marker is activated during infusion, the rate adjustment is immediately paused and the infusion restriction process is triggered to prevent further infusion from aggravating kidney damage. If the marker remains inactive, the infusion rate is continuously adjusted according to the increase feedback to ensure that fluid resuscitation achieves maximum benefit within safe boundaries. The resulting crystalloid infusion rate control signal is sent to the intelligent infusion pump in real-time via the communication interface, driving the pump to precisely execute the infusion operation.

[0044] After the execution control unit determines the status based on a combination of the stroke volume increase probability value and acute kidney injury risk markers, it needs to promptly activate the infusion restriction command to ensure patient safety when insufficient infusion benefit or organ damage risk is detected. The specific implementation method is as follows:

[0045] Activating the infusion restriction command is a crucial step in balancing the benefits of fluid resuscitation with organ protection. The infusion restriction command is a control signal sent by the execution control unit to the intelligent infusion pump to reduce the infusion rate or pause the infusion. Its core purpose is to avoid excessive fluid intake that burdens organs, or to reduce unnecessary fluid input when there is no volume benefit. Upon receiving an acute kidney injury risk indicator (ARDI), a high-priority risk signal output by the organ injury early warning channel, activation indicates that the patient's renal function may be at risk of damage due to excessive fluid intake or abnormal perfusion. Immediate intervention is required, immediately reducing the infusion rate to a maintenance rate and activating the infusion module. The maintenance rate is the minimum infusion rate required to maintain the patient's baseline blood volume requirement, clinically calibrated to be 10-20 ml per hour. This rate avoids circulatory fluctuations caused by abrupt fluid cessation and minimizes renal metabolic stress. The infusion module is a dedicated module for infusing vasoactive drugs. Norepinephrine constricts blood vessels and raises blood pressure, maintaining effective circulatory perfusion after infusion restriction, avoiding... To prevent tissue hypoxia due to insufficient blood volume, during implementation, upon receiving an acute kidney injury risk indicator, the execution control unit generates a rate adjustment signal within 1 second and sends it to the intelligent infusion pump to rapidly reduce the current rate to a maintenance dose. Simultaneously, it sends a start signal to the infusion module, which begins infusion at a preset initial dose (0.05 micrograms per kilogram of body weight per minute). During infusion, the patient's blood pressure is monitored in real time. If the systolic blood pressure is below 90 mmHg, the dose is increased in fixed increments (0.01 micrograms per kilogram of body weight per minute) to ensure that the blood pressure is maintained within a safe range. When the stroke volume increase probability value is lower than the preset probability threshold but no acute kidney injury risk is triggered, the preset probability threshold is the critical value for determining whether fluid resuscitation has a volume benefit. It has been calibrated to 0.5 based on a large amount of clinical data. That is, when the stroke volume increase probability value is lower than 0.5, it indicates that the patient's response to infusion is poor, and continuing infusion at the original rate will not yield the expected benefit. The absence of an acute kidney injury risk indicates that the patient's renal function is not significantly affected temporarily, and there is no need to immediately stop the infusion. A gradual adjustment can be made to ensure a smooth transition.A gradual rate reduction strategy is employed, simultaneously activating the infusion module. This strategy involves progressively decreasing the infusion rate at fixed time intervals and by varying the rate of decrease to avoid circulatory fluctuations caused by sudden rate changes. The rate reduction interval is set at 5 minutes, with each reduction increment being 50 ml per hour, until the rate reaches the maintenance dose or is stopped based on the patient's response. The infusion module here refers to the vasoactive drug infusion module, which functions identically to the infusion module and is used to maintain circulatory stability during the rate reduction process. During implementation, the execution control unit first calculates the difference between the current rate and the maintenance dose, breaks down the rate reduction into multiple reduction increments according to the rate of decrease, and sends a rate reduction signal to the intelligent infusion pump every 5 minutes. Simultaneously, the infusion module is activated to infuse the vasoactive drug at a low initial dose (0.02 micrograms per kilogram of body weight per minute). Stroke volume and blood pressure are monitored in real time during the rate reduction process. If the stroke volume increases (more than 5%), the rate reduction can be paused while maintaining the current rate for observation. If blood pressure remains stable, monitoring continues for 1 hour after the rate is reduced to the maintenance dose. If no abnormalities are found, the infusion can be stopped, ensuring a smooth and safe rate reduction process.

[0046] After the control unit completes the infusion operation or restricts the infusion command, the feedback optimization unit needs to continuously track the treatment effect. When the actual blood lactate clearance value does not reach the expected target of the dynamic decision unit, it indicates that the difference feature weights of the machine learning classifier may not match the individual patient response. It is necessary to optimize the model prediction accuracy by adjusting the weight coefficients. The specific implementation method is as follows:

[0047] The adjustment of weight coefficients is the core function of the feedback optimization unit. Its purpose is to make the machine learning classifier more closely reflect the patient's actual physiological response. The adjustment is based on the deviation ratio between the actual change in stroke volume after infusion and the expected target. The actual change in stroke volume after infusion is the difference between the measured stroke volume obtained by the data acquisition unit from the cardiac output monitoring device after infusion is stopped and the baseline value before infusion. This accurately reflects the actual effect of fluid resuscitation. The expected target is the target increase set by the dynamic decision-making unit based on the stroke volume increase probability value output by the machine learning classifier. The deviation ratio is the percentage of the difference between the actual change and the expected target relative to the expected target. For example, if the expected increase is 15% and the actual increase is 5%, the deviation ratio is 66.7%. This ratio quantifies the degree of deviation between the predicted value and the actual effect, reducing the decision weight of the difference feature in the machine learning classifier. The difference feature, namely the pulse pressure fluctuation feature extracted earlier, is the core input for the classifier to predict the stroke volume increase probability. The decision weight is the weight used by the classifier to measure the deviation of each... The parameter representing the degree of influence of a feature on the prediction result has a higher weight, and the greater the contribution of the feature to the prediction result. When the deviation ratio is large, it indicates that the current weight of the difference feature is too high, causing the prediction result to deviate from the actual result. Therefore, it is necessary to reduce its weight. During adjustment, the system has a built-in weight adjustment formula to determine the adjustment range according to the deviation ratio. When the deviation ratio is between 30% and 50%, the weight of the difference feature is reduced by 20%; when the deviation ratio is between 50% and 80%, it is reduced by 40%; when the deviation ratio exceeds 80%, it is reduced by 60%. At the same time, it is ensured that the adjusted weight is not lower than 30% of the initial weight to avoid completely ignoring the value of the feature. The weight change log is recorded synchronously during the adjustment process to facilitate subsequent tracking of the optimization process. The probability value of stroke volume increase is recalculated. The feature space of the gradient boosting decision tree model is reconstructed using the adjusted weight coefficients. The probability value of stroke volume increase is recalculated based on the optimized model to re-predict the physiological characteristics of the current patient and obtain a more accurate assessment of volume benefit.The adjusted weight coefficients are the difference feature weights after deviation ratio calibration, along with the weights of other features (remaining unchanged, only the difference feature weights are optimized). The feature space of the gradient boosting decision tree model is a multi-dimensional space composed of all input features and their weights. Changes in feature weights alter the spatial structure, thus affecting the model's prediction logic. Reconstructing the feature space involves substituting the adjusted weight coefficients into the model, redefining the contribution ratio of the difference feature to other features, and tilting the model's decision logic towards the patient's actual response. During implementation, the feedback optimization unit transmits the adjusted weight coefficients to the machine learning classifier, replacing the original weight values ​​through the model parameter update interface, and reconstructing the feature space according to the structural rules of the gradient boosting decision tree. This ensures that the new feature space accurately reflects the actual impact of the difference feature. The pulse pressure fluctuation features extracted from the current respiratory cycle segmentation are input into the iteratively updated gradient boosting decision tree model for real-time probability analysis. The pulse pressure fluctuation feature extracted during the current respiratory cycle segmentation refers to the pulse pressure fluctuation feature latest collected by the data acquisition unit after respiratory cycle segmentation and anomaly compensation, after weight adjustment. This ensures the real-time nature and accuracy of the input data. The iteratively updated gradient boosting decision tree model refers to the optimized model after reconstructing the feature space. It retains the core structure of the original model and achieves iterative optimization only through weight adjustment. Real-time probability calculus is the process by which the model recalculates the stroke volume increase probability value based on the new feature space and input features. The calculation logic is consistent with the original model, but due to the optimization of feature weights, the prediction results are more in line with the patient's actual situation. During the calculation, the input pulse pressure fluctuation feature is first standardized and then input into the iteratively updated model. The model outputs a new stroke volume increase probability value through the integrated operation of multiple decision trees. This value will serve as the basis for the next fluid resuscitation decision, continuously improving the system's adaptability to individualized patients and prediction accuracy.

[0048] In this invention, the data acquisition unit acquires multi-dimensional physiological data of pulse pressure variability waveform and stroke volume; the dynamic decision-making unit's volume benefit analysis channel segments the respiratory cycle, extracts pulse pressure fluctuation characteristics, and outputs a stroke volume increase probability value through a machine learning classifier; the organ injury early warning channel calculates the rate of change of blood lactate and creatinine clearance, and generates acute kidney injury risk markers by combining inferior vena cava imaging; the execution control unit sends infusion control or restriction commands according to the combined state; and the feedback optimization unit adjusts the model weights based on the actual blood lactate clearance value, recalculates the probability value, achieves individualized and precise resuscitation, balances volume benefit and organ protection, and improves treatment safety and effectiveness.

[0049] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A severe patient individualized fluid resuscitation decision support system, characterized in that, Comprise: The data acquisition unit (1) is connected with medical equipment to obtain pulse pressure variability waveform, stroke volume continuous measurement value and inferior vena cava image in real time, and extract blood lactic acid concentration time series data and creatinine clearance rate value from medical information system synchronously; The dynamic decision unit (2) contains parallel operation capacity benefit analysis channel and organ damage early warning channel, the capacity benefit analysis channel carries out respiratory cycle segmentation on pulse pressure variability waveform, extracts the difference value characteristics of maximum pulse pressure and minimum pulse pressure in each respiratory cycle, inputs the difference value characteristics into the machine learning classifier trained by historical capacity response positive cases, and outputs the stroke volume improvement probability value, the organ damage early warning channel calculates the hourly change rate of blood lactic acid concentration and the hourly change rate of creatinine clearance rate in real time, and generates acute kidney injury risk sign when the rising rate of blood lactic acid concentration exceeds the set multiple of the falling rate of creatinine clearance rate; The execution control unit (3) generates control instructions according to the combined state of the stroke volume improvement probability value and the acute kidney injury risk sign, including: (a) when the stroke volume improvement probability value is higher than the preset probability threshold value and the acute kidney injury risk sign is not generated, a crystal liquid infusion rate control signal is sent to the intelligent infusion pump; (b) when the stroke volume improvement probability value is lower than the preset probability threshold value or the acute kidney injury risk sign exists, a limited infusion instruction is sent to the intelligent infusion pump and the infusion module is activated to stop infusion synchronously; The feedback optimization unit (4) continuously collects the actual change value of stroke volume and the actual value of blood lactic acid clearance after stopping infusion, and adjusts the weight coefficient of the difference value characteristics in the machine learning classifier and recalculates the stroke volume improvement probability value when the actual value of blood lactic acid clearance does not reach the expected target of the dynamic decision unit (2).

2. The individualized fluid resuscitation decision support system for critically ill patients of claim 1, wherein: The capacity benefit analysis channel carries out respiratory cycle segmentation on pulse pressure variability waveform, identifies the time domain position of the start point of inspiration phase and the end point of expiration phase by real-time receiving ventilator airflow waveform signal, and cuts the pulse pressure variability waveform into independent respiratory cycle segments by the time domain position.

3. The individualized fluid resuscitation decision support system for critically ill patients of claim 2, wherein: The extraction of the difference value characteristics, in the segmented single respiratory cycle segment, adopts peak detection algorithm to locate the maximum pulse pressure point and the minimum pulse pressure point, and generates pulse pressure fluctuation characteristic quantity by calculating the pressure amplitude difference between the maximum pulse pressure point and the minimum pulse pressure point.

4. The individualized fluid resuscitation decision support system for critically ill patients of claim 3, wherein: The construction of the machine learning classifier uses the pulse pressure fluctuation characteristic quantity and the corresponding stroke volume improvement label of historical capacity response positive cases, selects the difference value characteristics by feature importance sorting, trains gradient boosting decision tree model to output stroke volume improvement probability value, and the gradient boosting decision tree model takes the machine learning classifier as the model architecture, wherein the historical capacity response positive case is defined as the one whose stroke volume increases by a certain amount after fluid loading test.

5. The critically ill patient individualized fluid resuscitation decision support system of claim 1, wherein: The calculation of the hourly change rate of blood lactic acid concentration of the organ damage early warning channel adopts sliding window method to linearly fit the blood lactic acid concentration time series data, extracts the unit time slope as the rising rate of blood lactic acid concentration, and the falling gradient of the continuous two detection values is calculated by difference method to obtain the change rate of creatinine clearance rate, both of which are standardized by hour.

6. The individualized fluid resuscitation decision support system for critically ill patients of claim 5, wherein: The acute kidney injury risk marker is generated based on the rate of increase in blood lactate concentration and the rate of decrease in creatinine clearance rate, and the diameter respiratory variability of the inferior vena cava image is introduced as a verification parameter; when the rate of increase in blood lactate concentration exceeds the set multiple of the rate of decrease in creatinine clearance rate and the diameter respiratory variability is lower than the critical threshold, the acute kidney injury risk marker is activated.

7. The critically ill patient individualized fluid resuscitation decision support system of claim 1, wherein: The crystal liquid infusion rate control signal is generated, the basic infusion rate is set according to the stroke increase probability value, the rate is adjusted by a fixed step value based on the amplitude feedback of the real-time stroke continuous measurement value, and the acute kidney injury risk marker is monitored.

8. The critically ill patient individualized fluid resuscitation decision support system of claim 6, wherein: The limited infusion instruction is activated, the infusion rate is immediately reduced to a preset maintenance amount and the infusion module is started when the acute kidney injury risk marker is received, and when the stroke increase probability value is lower than a preset probability threshold but the acute kidney injury risk is not triggered, a gradual speed reduction strategy is adopted to simultaneously enable the infusion module.

9. The individualized fluid resuscitation decision support system for critically ill patients of claim 3, wherein: The weight coefficient is adjusted according to the deviation proportion of the actual value and the expected target of blood lactate clearance after infusion, and the decision weight of the difference feature in the machine learning classifier is reduced.

10. The individualized fluid resuscitation decision support system for critically ill patients of claim 9, wherein: The recalculated stroke increase probability value uses the adjusted weight coefficient to reconstruct the feature space of the gradient boosting decision tree model, and the pulse pressure fluctuation feature quantity extracted in the current respiratory cycle is input into the iteratively updated gradient boosting decision tree model for real-time probability calculation.

Images