Intelligent vasoactive drug infusion system based on real-time physiological data

Abstract

The invention relates to the technical field of intensive care medical treatment, and discloses a vasoactive drug intelligent infusion system based on real-time physiological data. A physiological data acquisition unit of the system acquires real-time physiological data and performs preprocessing. The physiological state classification unit classifies the preprocessed data according to a preset physiological state interval, and extracts multiple types of physiological features to generate a state feature set. The abnormal state marking unit monitors the switching frequency of the physiological state interval and marks the abnormal state interval when the frequency exceeds a threshold value. The drug reaction feature decomposition unit decomposes the feature set of the non-abnormal state interval based on the similarity between the historical drug reaction stage and the current physiological feature to generate a drug reaction feature subset. And the physiological parameter coupling correction unit is used for performing parameter correction on the subset by utilizing the coupling relationship among the physiological parameters. And the medicine infusion decision-making unit reversely traverses and positions potential medicine requirements along the medicine effect path, and accordingly, the infusion priority is determined, and an infusion instruction is generated.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intensive care medical technology, in particular to an intelligent infusion system for vasoactive drugs based on real-time physiological data. BACKGROUND

[0002] Vasoactive drug infusion is a key treatment for maintaining hemodynamic stability in critically ill patients in intensive care units. Such drug treatment has a narrow window and large individual differences, and accurate dose control is crucial for patient prognosis. The traditional infusion mode mainly relies on medical staff to manually adjust the drug delivery rate according to the intermittently measured blood pressure, heart rate and other parameters, combined with clinical experience. This mode has significant limitations: the measurement of physiological parameters has intervals, and cannot capture the instantaneous dynamic changes between parameters; the experience difference of medical staff may lead to inconsistent decisions; manual adjustment has a delayed response, making it difficult to respond to rapid fluctuations in patient status. Especially in shock, severe infection and other hemodynamic instability states, the traditional method is difficult to achieve precise titration of drugs.

[0003] Existing automated infusion devices, such as target-controlled infusion pumps, mostly use simple proportional-integral-derivative control algorithms based on the deviation of a single parameter from the preset target value for adjustment. Such systems have poor adaptability to complex physiological state changes and cannot effectively handle clinical scenarios with multiple parameter coupling and nonlinear responses. For example, when a patient experiences blood pressure fluctuations and arrhythmia at the same time, a simple PID algorithm may produce contradictory adjustment instructions, or even exacerbate the condition. In addition, the system lacks the ability to learn from individual differences in drug reactions and cannot optimize current decisions based on the patient's historical reaction data. More importantly, existing technologies lack intelligent identification and processing mechanisms for "abnormal states". Rapid and frequent switching of patient physiological states may be due to worsening of the primary condition, other treatment interventions or measurement disturbances, but existing systems usually treat it as normal fluctuations and may lead to inappropriate drug adjustments, posing safety risks.

[0004] The field of intensive care medicine urgently needs an intelligent infusion system that can integrate multiple physiological parameters in real time, intelligently identify physiological states, distinguish between normal and abnormal fluctuations, and make precise decisions based on individual drug response characteristics. SUMMARY

[0005] The purpose of the present application is to provide an intelligent infusion system for vasoactive drugs based on real-time physiological data to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides an intelligent infusion system for vasoactive drugs based on real-time physiological data, which comprises: The physiological data acquisition unit, the physiological state classification unit, the abnormal state marking unit, the drug reaction feature decomposition unit, the physiological parameter coupling correction unit, and the drug infusion decision unit; The physiological data acquisition unit is configured to acquire real-time physiological data and perform a preprocessing operation. The physiological state classification unit is configured to classify the preprocessed physiological data according to preset physiological state intervals, and extract multiple physiological features in each physiological state interval to generate a state feature set. The abnormal state marking unit is configured to monitor the switching frequency of the physiological state intervals, and mark an abnormal state interval when the switching frequency exceeds a preset threshold. The drug reaction feature decomposition unit is configured to perform decomposition processing on the state feature set of the non-abnormal state interval based on the similarity matching result of the historical drug reaction stage and the current physiological feature, to generate a drug reaction feature subset. The physiological parameter coupling correction unit is configured to correct the drug reaction feature subset using the coupling relationship between physiological parameters. The drug infusion decision unit is configured to traverse the drug effect path in reverse direction according to the corrected drug reaction feature subset to locate potential drug demand, and determine the infusion priority and generate the infusion instruction based on the traversal result.

[0007] Preferably, the physiological data acquisition unit includes a signal synchronous acquisition subunit, a timestamp alignment subunit, and a data filtering and cleaning subunit. The signal synchronous acquisition subunit is configured to synchronously acquire arterial blood pressure waveform data, electrocardiogram signal data, and pulse blood oxygen saturation data. The timestamp alignment subunit is configured to perform timestamp alignment processing on the arterial blood pressure waveform data, electrocardiogram signal data, and pulse blood oxygen saturation data. The data filtering and cleaning subunit is configured to remove high-frequency fluctuation noise in the arterial blood pressure waveform data using a mean filtering method, and eliminate abnormal data points based on the time sequence correlation between the electrocardiogram signal data and the pulse blood oxygen saturation data. The processed arterial blood pressure waveform data, electrocardiogram signal data, and pulse blood oxygen saturation data are resampled at a fixed sampling frequency, and a standardized physiological data sequence is output.

[0008] Preferably, the physiological state classification unit includes a state interval division subunit and a feature fusion subunit. The state interval division sub-unit is configured to divide the standardized physiological data sequence into a low blood pressure state interval, a normal blood pressure state interval and a high blood pressure state interval according to the systolic pressure value; the feature fusion sub-unit is configured to extract time domain features and frequency domain features from the arterial blood pressure waveform data, the electrocardiogram signal data and the pulse blood oxygen saturation data in each physiological state interval respectively; and the time domain features and the frequency domain features in the same physiological state interval are fused according to a sliding time window to generate a multi-dimensional state feature set.

[0009] Preferably, the time domain features of the arterial blood pressure waveform data include a blood pressure coefficient of variation, a blood pressure trend line slope and a blood pressure fluctuation entropy value, the frequency domain features of the electrocardiogram signal data include a low frequency power spectral density and a high frequency power spectral density of heart rate variability, and the time domain features of the pulse blood oxygen saturation data include a blood oxygen saturation drop rate and a blood oxygen fluctuation amplitude.

[0010] Preferably, the abnormal state marking unit comprises a switching frequency calculation sub-unit and a state consistency evaluation sub-unit. The switching frequency calculation sub-unit is configured to calculate a switching frequency per minute based on a fixed time length window and compare the switching frequency with a frequency threshold value; if the switching frequency exceeds the frequency threshold value, the state consistency evaluation sub-unit aligns physiological features of adjacent physiological state intervals through a waveform alignment algorithm and calculates a cross-state consistency score; if the cross-state consistency score is lower than a consistency threshold value, the physiological state interval corresponding to the current window is marked as an abnormal state interval and a feature re-extraction process is triggered; otherwise, the physiological state interval is marked as a normal state interval.

[0011] Preferably, calculating the cross-state consistency score comprises calculating mutual information values of the aligned blood pressure coefficient of variation, the low frequency power spectral density of heart rate variability and the blood oxygen saturation drop rate respectively, and taking a weighted average value of the mutual information values as the cross-state consistency score.

[0012] Preferably, the drug reaction feature decomposition unit comprises a multi-modal feature fusion sub-unit and an adaptive decomposition sub-unit. The multi-modal feature fusion sub-unit is configured to align the time series features of the historical drug response stage with the current physiological features in the state feature set of the non-abnormal state interval using a dynamic time warping algorithm, and calculate a comprehensive matching degree by combining the weighted results of the cosine similarity measure and the dynamic time warping distance; the historical drug response stage label is divided by the blood pressure waveform complexity coefficient and the heart rate variability spectrum kurtosis feature; if the comprehensive matching degree is higher than the matching threshold, the adaptive decomposition sub-unit extracts the intrinsic mode function component using the ensemble empirical mode decomposition algorithm, screens the feature frequency components related to the vasoactive drug response to form a drug response feature subset; if the comprehensive matching degree is lower than the matching threshold, the energy features of the specific frequency band are extracted using the wavelet packet transform based on the correlation characteristics of the heart rate variability multifractal spectrum and the drug metabolism kinetics, and the drug response feature subset is generated by combining the nonlinear features extracted by the recursive quantitative analysis.

[0013] Preferably, the physiological parameter coupling correction unit comprises a vascular response correction sub-unit and a metabolic influence correction sub-unit. The vascular response correction sub-unit compensates and corrects the blood pressure variability coefficient in the drug response feature subset according to the linear relationship between vascular resistance and blood pressure; the metabolic influence correction sub-unit performs metabolic calibration correction on the heart rate variability low-frequency power spectral density in the drug response feature subset based on the nonlinear mapping between the basal metabolic rate and the heart rate variability; the corrected blood pressure variability coefficient and heart rate variability low-frequency power spectral density are combined with the oxygen saturation decrease rate feature to output the parameter-coupled corrected drug response feature subset.

[0014] Preferably, the drug infusion decision unit comprises an effect path mapping sub-unit, an abnormal parameter identification sub-unit, a reverse search sub-unit, and an instruction generation sub-unit. The effect path mapping sub-unit is configured to obtain a drug effect path map, which is constructed based on the physiological parameter causal chain and the drug receptor binding path, and comprises the effect propagation direction and intensity weight of the drug action node; the abnormal physiological parameters are identified from the parameter-coupled corrected drug response feature subset, and the corresponding node in the drug effect path map is matched according to the type of the abnormal physiological parameters; the potential drug target node of the abnormal physiological parameters is located by reverse search along the propagation direction of the drug effect path map; the infusion priority is calculated according to the number of abnormal parameters and the path level of the potential drug target node; the infusion instruction containing the drug type identification and the infusion flow rate is generated based on the infusion priority and the real-time physiological state.

[0015] Preferably, the drug infusion decision unit further comprises a safety boundary monitoring sub-unit configured to monitor the physiological parameter changes after the execution of the infusion instruction in real time, and if the physiological parameter exceeds the preset safety boundary, the infusion flow rate is dynamically adjusted or the instruction regeneration process is triggered.

[0016] Compared with the prior art, the present application has the beneficial effects that: The present application realizes the deep interpretation and risk screening of complex physiological signals of patients through physiological state classification and abnormal state marking. The system not only performs simple parameter monitoring, but also divides continuous physiological data flow into state intervals with clinical significance, and acutely identifies abnormal conditions with excessively frequent state switching. This capability enables it to distinguish between physiological normal fluctuations and abnormal signals that may indicate changes in the patient's condition, providing a more reliable and safer data basis for subsequent decision-making, and avoiding misjudgment and inappropriate intervention of abnormal states.

[0017] The introduction of drug response feature decomposition and historical similarity matching gives the system the ability of individualized treatment. The system can separate similar patterns from the patient's current non-abnormal physiological features to historical drug response stages to form a subset reflecting the patient's unique pharmacokinetic and pharmacodynamic characteristics. This enables the adjustment of drug dosage to be based on the patient's own historical response rather than relying on population average parameters or fixed formulas, significantly improving the accuracy and adaptability of drug administration.

[0018] Using physiological parameter coupling relationships to correct features enhances the physiological rationality and clinical relevance of the decision. Physiological parameters of the human body interact with each other to form a complex network. The system corrects the extracted drug response features by modeling the coupling relationships between parameters, making the final feature set used for decision-making better reflect the underlying physiological and pathological changes, reducing the decision bias that may be caused by isolated consideration of a single parameter, and making the infusion instructions more consistent with physiological principles.

[0019] Using decision logic that traverses in reverse along the drug effect path, the system realizes the transition from "phenomenon-driven" to "mechanism-driven". Traditional methods directly calculate the dosage adjustment based on the difference between the current blood pressure and the target. The present system, however, attempts to reason in reverse, starting from the current physiological features and working backward along the known action path of the drug to locate the most likely source of demand, thereby determining the infusion priority. This mechanism-based exploration helps to better understand the nature of drug demand in the current state, making the decision more forward-looking and targeted.

[0020] The system builds a complete closed loop from data perception, state recognition, feature extraction to intelligent decision-making. Each unit works in layers, transforming raw physiological signals into precise infusion instructions. This not only reduces the over-reliance on medical staff experience and reduces the inconsistency of human decision-making, but also enables 24-hour uninterrupted real-time monitoring and rapid response, providing more stable and safer life support for critically ill patients. The intelligent and individualized features of the system represent the development direction of precision medicine in the field of intensive care. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The working principle diagram of the intelligent infusion system of vasoactive drugs based on real-time physiological data according to the present application; Figure 2 The working flow diagram of the physiological data acquisition unit; Figure 3 The flow diagram of the cross-state consistency score calculation; Figure 4 The diagram of the drug response feature decomposition and comprehensive analysis. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0023] Please refer to Figure 1 The present application provides an intelligent infusion system of vasoactive drugs based on real-time physiological data. The system includes multiple functional units integrated to achieve intelligent infusion control of vasoactive drugs. The physiological data acquisition unit is responsible for collecting real-time physiological data of the patient, including arterial blood pressure waveform data, electrocardiogram signal data and pulse blood oxygen saturation data, and performing data preprocessing operations to eliminate noise and outliers. The preprocessed data is transmitted to the physiological state classification unit, which classifies the physiological data according to the preset physiological state interval and extracts multiple physiological features in each interval to generate a state feature set. The abnormal state marking unit monitors the switching frequency of the physiological state interval. When the switching frequency exceeds the preset threshold, the unit marks the abnormal state interval by calculating the cross-state consistency score, thereby avoiding misjudgment. For non-abnormal state intervals, the drug response feature decomposition unit decomposes the state feature set based on the similarity matching results of historical drug response stages and current physiological features to generate a drug response feature subset. The physiological parameter coupling correction unit uses the coupling relationship between physiological parameters to correct the drug response feature subset to improve feature accuracy. Finally, the drug infusion decision unit traverses the potential drug demand along the drug effect path based on the corrected drug response feature subset to determine the infusion priority and generate the infusion instruction, thereby achieving precise drug infusion control.

[0024] Embodiment 1: Please refer to Figure 2The physiological data acquisition unit synchronously acquires the arterial blood pressure waveform data, the electrocardiogram signal data and the pulse blood oxygen saturation data through a signal synchronization acquisition subunit. This synchronization process relies on a precise clock signal at the hardware level, so that the multiple physiological signals can be recorded under the same time reference. A time stamp alignment subunit performs time stamp alignment processing on the data. The subunit corrects millisecond-level deviations caused by transmission delay or slight differences in the starting time of sampling by using an interpolation algorithm, thereby ensuring that all data streams strictly correspond on the time axis. A data filtering and cleaning subunit removes high-frequency fluctuation noise in the arterial blood pressure waveform data by using a mean filtering method. The filtering method smoothes the signal by calculating the arithmetic mean of adjacent sampling points. For the electrocardiogram signal data and the pulse blood oxygen saturation data, the subunit uses the inherent time sequence correlation, such as the physiological connection between the electrocardiogram R wave and the blood oxygen pulse wave peak, to identify and eliminate abnormal data points caused by motion artifacts or transient device interference. The arterial blood pressure waveform data, the electrocardiogram signal data and the pulse blood oxygen saturation data processed in the above manner are then resampled at a fixed sampling frequency. The frequency is usually set to 125 Hz to be compatible with the standards of most medical monitoring devices. Finally, a standardized and time-dimension-unified physiological data sequence is output.

[0025] After receiving the standardized physiological data sequence, a state interval division subunit of the physiological state classification unit divides the systolic pressure value into three clear physiological state intervals according to clinical consensus. The low blood pressure state interval is defined as the systolic pressure being continuously lower than 90 mmHg. The normal blood pressure state interval is defined as the systolic pressure being between 90 and 140 mmHg. The high blood pressure state interval is defined as the systolic pressure being continuously higher than 140 mmHg. The design of the interval boundaries takes into account appropriate hysteresis bands, such as a narrow buffer region between 89 mmHg and 91 mmHg, to prevent frequent switching of states caused by slight fluctuations in blood pressure values near the threshold, thereby enhancing the stability of system determination. A feature fusion subunit then analyzes the multi-modal data in each physiological state interval divided. For the arterial blood pressure waveform data, the extracted time-domain features include the blood pressure coefficient of variation, which is obtained by calculating the ratio of the standard deviation of the systolic pressure to its mean value in a period of time, to quantify the volatility of blood pressure. The blood pressure trend line slope describes the direction and rate of change of blood pressure in the short term through linear fitting. The blood pressure fluctuation entropy value evaluates the irregularity and complexity of the blood pressure sequence by borrowing the concept of information theory.

[0026] For electrocardiogram data, the focus is on the frequency domain features, especially the low frequency power spectral density and high frequency power spectral density of heart rate variability, which are calculated after the electrocardiogram signal is converted from time domain to frequency domain through fast Fourier transform. These features reflect the balance of the autonomic nervous system by calculating the power distribution in a specific frequency band. For pulse oximetry data, the time domain features extracted include the oxygen saturation decrease rate, which is the decrease in oxygen saturation value per unit time, and the oxygen fluctuation amplitude, which is the difference between the maximum and minimum values of oxygen saturation within a certain time window. After extracting these features, the feature fusion subunit uses a sliding time window technique to fuse the time domain and frequency domain features from different data sources within the same physiological state interval. The width of this sliding window can be adjusted according to the clinical scenario, typically set to 30 seconds to 2 minutes, to capture clinically meaningful physiological change patterns. The fused results generate a multi-dimensional state feature set, which is organized in matrix form, with each row representing a time point and each column representing a physiological feature dimension.

[0027] During the generation of the state feature set, the system also considers the internal relationship between different physiological parameters, such as the physiological coupling relationship between the change of R-R interval in electrocardiogram and the pulse wave transmission time in arterial blood pressure waveform. This relationship can be used to verify the credibility of the extracted features. The division of the state interval is not fixed, and the system allows for fine-tuning based on the individualized baseline blood pressure value of the patient. For example, for patients with known long-term hypertension, the upper limit of the normal blood pressure interval may be adjusted upward. The sliding time window used for feature extraction typically has a step size of half the window length. This overlapping sampling method helps to reduce information loss caused by window boundary cutting and ensures the continuity of state changes.

[0028] The calculation of blood pressure coefficient of variation requires sufficient data points to ensure statistical reliability, so each analysis window typically contains at least tens of heart cycles of blood pressure data. The calculation of blood pressure trend line slope uses linear regression with least squares method to minimize fitting error. The calculation of blood pressure fluctuation entropy involves estimating the probability distribution, which requires dividing the blood pressure values into intervals and calculating their frequency. The calculation of heart rate variability power spectral density requires a relatively stable electrocardiogram signal, so segments with no obvious arrhythmia and significant noise interference are usually selected for analysis. The calculation of oxygen saturation decrease rate focuses on the rapid change phase, and its time derivative is approximated by difference operation. In the feature fusion stage, to avoid bias in subsequent analysis due to large differences in dimension and numerical range, the features before fusion are usually standardized, such as converting them to a distribution with mean zero and standard deviation one using Z-score standardization method. The generated multi-dimensional state feature set constitutes a quantitative description of the patient's current physiological state.

[0029] Embodiment 2: Refer to Figure 3 The implementation of the abnormal state flagging unit relies on the continuous monitoring and analysis of the physiological state interval switching pattern. The switching frequency calculation sub-unit counts the number of interval switches based on a fixed time window, which is usually set to five minutes to balance the real-time and stability requirements of detection. The system updates the window content at each sampling time and calculates the total number of state interval switching events that occur within the window. For example, switching from the low blood pressure interval to the normal interval or from the normal interval to the high blood pressure interval is counted as one valid switch. The switching frequency per minute is obtained by dividing the total number by the window length. The calculated frequency value is compared with a preset frequency threshold, which is usually set to three switches per minute. This value is based on the common requirement for blood pressure stability in clinical practice. If the observed switching frequency continues to exceed this threshold, it indicates that the patient's hemodynamic state may be in an unstable oscillation phase.

[0030] When the switching frequency exceeds the threshold, the state consistency assessment sub-unit is activated for deeper verification. This sub-unit extracts the physiological feature sequences corresponding to adjacent physiological state intervals, such as the two state interval features before and after the last switch in the current window. Waveform alignment algorithm is used to align these features in time. The waveform alignment algorithm usually selects dynamic time warping technology, which can flexibly stretch or compress time series to eliminate small time axis shifts caused by individual differences in physiological response speed or transient signal delays, achieving optimal matching of feature points. After aligning the feature sequences, the sub-unit starts calculating the cross-state consistency score. This calculation process focuses on several core physiological features, including blood pressure coefficient of variation, heart rate variability low-frequency power spectral density, and blood oxygen saturation decline rate. For each feature, the mutual information value between different state intervals is calculated. Mutual information value measures the degree of correlation of the information contained in a feature before and after state change. When calculating, the joint probability distribution and marginal probability distribution of the feature need to be constructed. Histogram method or kernel density estimation method is used to approximate these distributions, and finally a quantitative correlation index is obtained.

[0031] The multiple mutual information values calculated are not treated equally, they are given different weights according to their importance in clinical judgment, the coefficient of variation of blood pressure is usually given the highest weight of 0.5 due to its direct correlation with vascular tension and cardiac output, the low frequency power spectral density of heart rate variability is given a medium weight of 0.3 as a reflection of sympathetic nervous activity, and the oxygen saturation drop rate is given a weight of 0.2, and the weighted average of these weights is finally determined as the cross-state consistency score. This score is compared with a preset consistency threshold, which is usually set to 0.7, if the score is lower than the threshold, it indicates that the physiological characteristics of adjacent state intervals lack internal coherence, and the physiological state interval corresponding to the current time window is therefore marked as an abnormal state interval, the system will trigger the feature re-extraction process and inform the physiological state classification unit to reanalyze the original data in this time period to obtain a more reliable feature set, if the score is higher than or equal to the threshold, it indicates that the state transition may be due to real physiological changes, the current interval is marked as a normal state interval, and subsequent processing will continue.

[0032] The accuracy of waveform alignment is crucial to the entire evaluation process, so the algorithm parameters need to be carefully adjusted, the path constraint of dynamic time warping is usually set in a symmetric form to ensure the rationality of alignment, when calculating the mutual information value, the number of bins used for probability estimation needs to be adaptively selected according to the data volume, when the data volume is sufficient, fine-grained binning can be used to improve accuracy, when the data volume is small, coarse-grained binning is used to ensure the stability of estimation, sometimes the mutual information values are normalized before weighted averaging to make their scales consistent. The state interval marked as abnormal is recorded by the system and temporarily excluded from subsequent drug response feature analysis, so as to avoid making decisions based on unreliable data, if the system detects abnormal states continuously for multiple windows, it will generate an alarm to prompt clinical medical staff for manual verification, ensuring system safety redundancy. The setting of frequency threshold is not completely fixed, the system allows fine-tuning according to the patient's specific medical history and current monitoring stage, for example, for patients in the early postoperative period, the threshold may be temporarily relaxed to adapt to the expected blood pressure fluctuations, the calculation of switching frequency uses a rolling window mechanism, updating the statistical results every second to ensure the continuity of monitoring, state consistency evaluation is not limited to directly adjacent intervals, under certain configurations, it also examines the overall consistency pattern of all state transitions within a window.

[0033] Embodiment 3: The implementation of the drug response feature decomposition unit begins with the processing of the state feature set of the non-anomalous state interval, and the multi-modal feature fusion sub-unit focuses on the current physiological features in the set, which constitute a multi-dimensional time series. The core task of this sub-unit is to perform similarity matching between these current features and the pre-constructed historical drug response phase database. The labels of the historical drug response phases are not randomly assigned, but are based on two specific quantitative indicators: the blood pressure waveform complexity coefficient and the heart rate variability spectrum kurtosis feature. The blood pressure waveform complexity coefficient evaluates the irregularity of the waveform sequence by calculating the approximate entropy, while the heart rate variability spectrum kurtosis feature is extracted through high-order spectrum analysis, reflecting the peak characteristics of the heart rate signal in the frequency domain. Using these indicators, a large amount of historical physiological data after drug use is divided into several representative response phases by an unsupervised clustering algorithm, and each phase corresponds to a drug response mode.

[0034] To perform matching, the time series of current physiological features needs to be compared with the feature sequences of historical response phases, which involves the problem of time series alignment. Because even the same physiological response may differ in time scale and rhythm due to individual differences, the dynamic time warping algorithm is used to solve this problem. This algorithm finds a nonlinear time warping path to achieve optimal alignment of two sequences while maintaining shape similarity as much as possible. After alignment, the similarity between the sequences needs to be calculated. Here, a comprehensive matching degree measurement method is used, which combines the weighted results of cosine similarity and dynamic time warping distance. Cosine similarity focuses on the directional consistency of feature vectors in multi-dimensional space, while dynamic time warping distance focuses on the closeness of the entire time series shape. The weighted fusion of the two forms a comprehensive matching degree score , which is calculated as: where: represents the cosine similarity value, represents the distance after warping, is the maximum distance constant, and are the weights given to the two, and satisfy .

[0035] If the calculated comprehensive matching degree is higher than the preset matching threshold (for example, 0.8), it means that the current physiological state is highly similar to a certain historical drug response stage, and the adaptive decomposition subunit starts the ensemble empirical mode decomposition process accordingly, which is performed on the current characteristic signal. Ensemble empirical mode decomposition is a noise-assisted data analysis method that adds different proportions of white noise to the original signal multiple times, performs empirical mode decomposition on each noise-added signal to generate a group of intrinsic mode function components, and then averages the corresponding components obtained from multiple decompositions to effectively suppress modal aliasing. From this group of intrinsic mode function components, the characteristic frequency components related to the response of vasoactive drugs need to be selected, for example, the components with a frequency range of 0.04 Hz to 0.15 Hz are usually closely related to the contraction and relaxation activities of blood vessels. These selected components collectively constitute the drug response characteristic subset.

[0036] If the comprehensive matching degree is lower than the matching threshold, it indicates that the current physiological characteristics cannot be well matched with the known historical response stages, representing a new or atypical response pattern. In this case, the decomposition strategy will change, and the system will analyze the known correlation between heart rate variability multifractal spectrum and drug metabolism kinetics. Heart rate variability multifractal spectrum can reveal the self-similarity and complexity of heart rate fluctuations at different time scales, which is inherently related to the metabolism and distribution of drugs in the body. At this time, the adaptive decomposition subunit will use wavelet packet transform to perform time-frequency analysis on the signal. Wavelet packet transform can provide more detailed frequency band division than standard wavelet transform, thereby extracting energy distribution features in specific frequency bands. At the same time, recursive quantitative analysis is applied to extract nonlinear features of the signal. Recursive quantitative analysis can quantify the determinism and recurrence rate of signal dynamics on a recurrence plot. Finally, the frequency band energy features obtained by wavelet packet transform and the nonlinear features obtained by recursive quantitative analysis are combined to generate the drug response characteristic subset.

[0037] During the entire decomposition process, the historical drug response stage database needs to be updated regularly to include new patient data, thereby enhancing the adaptability and generalization ability of the system. The path constraint of dynamic time warping algorithm is usually set to symmetric mode to balance the computational complexity and alignment accuracy. When calculating the comprehensive matching degree, the weights of and can be adjusted according to the reliability of different features, for example, the weight of the warping distance can be appropriately increased for features with high signal-to-noise ratio. The standard deviation of white noise added in ensemble empirical mode decomposition and the number of ensemble averages are important parameters that affect the decomposition effect and need to be optimized according to the signal characteristics. The selection of wavelet basis function and the determination of decomposition level in wavelet packet transform directly affect the effectiveness of frequency band separation. The settings of parameters such as embedding dimension and delay time in recursive quantitative analysis depend on the preliminary judgment of the chaotic characteristics of the signal.

[0038] Referring to Figure 4 The upper subgraph demonstrates the real-time monitoring results of multiple physiological parameters, including the systolic pressure signal (black solid line), diastolic pressure signal (dark gray solid line), and heart rate variability signal (gray dashed line). In the time series, the light gray shaded area marks the high-matching interval of drug response, where the physiological characteristics have a high similarity to the patterns in the historical drug response database. The dark gray shaded area identifies the abnormal state interval, which may cause difficulty in feature extraction due to signal interference or physiological state mutation. The continuous monitoring of these physiological signals provides the raw data basis for drug response feature decomposition. The systolic and diastolic pressure signals reflect the vascular tone state, and the heart rate variability signal provides information on the activity of the autonomic nervous system. By aligning these time series using the dynamic time warping algorithm and combining cosine similarity calculations to comprehensively match the degree, the system can accurately identify the similarity to the historical drug response stage. The lower subgraph demonstrates the signal decomposition results based on the ensemble empirical mode decomposition (EEMD) algorithm. The figure shows the first three main intrinsic mode function (IMF) components, represented by different depths of gray lines. The black solid line represents the reconstructed signal, i.e., the physiological signal reconstructed based on the filtered feature components, verifying the effectiveness of feature extraction. The gray boxes and text on the right side of the graph highlight the importance of the feature frequency range, which usually contains the most relevant physiological information for the vasodilator drug response.

[0039] Example 4: Implementation process of the physiological parameter coupling correction unit revolves around its two core subunits, namely the vascular response correction subunit and the metabolic influence correction subunit, which sequentially calibrate subsets of drug response features. Suppose the system receives a feature subset containing the following values: blood pressure coefficient of variation is 0.15, heart rate variability low-frequency power spectral density is 620 ms² / Hz, and blood oxygen saturation decline rate is 0.9% / s. The vascular response correction subunit is activated first, and its operation is based on a recognized physiological principle that, under relatively stable cardiac output, the level of mean arterial pressure is positively correlated with peripheral vascular resistance. To perform the correction, this subunit needs to obtain two key vascular resistance values: the patient's individual baseline vascular resistance and the current instant vascular resistance.

[0040] The baseline vascular resistance is an individualized reference benchmark derived from the patient's own historical data record in a recent hemodynamic stable state, calculated by the system through analysis of the mean arterial pressure and cardiac output monitoring values in that time period, assuming this baseline value as 1.0 PRU. The instantaneous vascular resistance is dynamically calculated through real-time data, using the mean arterial pressure in the current monitoring period and the continuously estimated cardiac output based on pulse waveform analysis technology, substituted into the hemodynamic formula to obtain, assuming that the calculated instantaneous value is 1.3 PRU. The sub-unit calculates the resistance compensation factor, which is the ratio of the instantaneous resistance to the baseline resistance, which in this example is . The original blood pressure coefficient of variation will be multiplied by this compensation factor to obtain the corrected value: The intention of this correction is that when the vascular resistance is elevated, the cardiovascular system may be in a higher tension state, at which the same absolute blood pressure fluctuation amplitude may contain greater clinical instability risk. By amplifying the coefficient of variation through the compensation factor, it aims to more sensitively reflect this state change.

[0041] The metabolic influence correction sub-unit then calibrates the heart rate variability low-frequency power spectral density, which is significantly modulated by the body's basal metabolic level. The correction logic is that an increase in basal metabolic rate is usually accompanied by an increase in the basal level of the sympathetic nervous system, which will raise the baseline value of the heart rate variability low-frequency power to some extent. The sub-unit estimates the patient's current basal metabolic rate, which is usually done using a standardized formula containing height, weight, age, and gender information, assuming the calculation result is 1620 kcal per day. The system internally presets a reference metabolic rate value for calibration, which may be based on population standards or specific clinical scenarios, assuming 1500 kcal per day. Instead of directly using linear proportion, a nonlinear conversion function is introduced to better fit the physiological relationship, such as the square root of the ratio. In this example, the ratio is , and its square root is approximately 1.039. The original heart rate variability low-frequency power spectral density value will be multiplied by this calibration coefficient to obtain the corrected value: 620 ms² / Hz * 1.039 ≈ 644 ms² / Hz. The purpose of this calibration operation is to partially offset the lifting effect of low-frequency power spectral density caused by high basal metabolism in the background, so that the corrected value can more purely reflect the changes in autonomic nervous activity caused by drug intervention or pathophysiological changes.

[0042] The blood oxygen saturation rate feature, whose variation is mainly affected by the lung ventilation / gas exchange function and blood oxygen transport efficiency, has relatively weak direct coupling relationship with vascular tension and basal metabolic rate, so it is not adjusted in this revision example, and the original value of -0.9% / s is maintained. After all the revision steps are completed, the three features are combined and output to form the final drug response feature subset after parameter coupling revision. Referring to Table 1, the complete data flow and calculation logic from the original input to the revised output are clearly shown.

[0043] Table 1: Physiological parameter coupling revision process data record The performance of the vascular response revision subunit depends on the accuracy of its individualized parameter profile, which is established during the stabilization period after the patient accesses the system and is continuously verified and updated. The real-time estimation of vascular resistance is highly dependent on the accuracy of invasive arterial blood pressure monitoring and the effectiveness of the cardiac output calculation model. The reference metabolic rate value used by the metabolic influence revision subunit can be configured according to specific clinical needs, for example, in applications related to metabolic disease management, more individualized longitudinal reference values may be used. The system usually sets reasonable upper and lower threshold values for the revision coefficients. If the calculated resistance compensation factor or metabolic calibration coefficient exceeds this range (for example, less than 0.6 or more than 1.8), a data quality review flag will be triggered, prompting clinical staff to pay attention to possible signal interference or extreme physiological states, preventing the generation of misleading revision results. The entire revision process aims to de-mix the original drug response features by introducing vascular resistance and metabolic level, two important physiological background information, to improve the accuracy of the subsequent decision unit in judging the true effect of the infused vasoactive drugs.

[0044] Example 5: The workflow of the drug infusion decision unit can be specifically presented through a simulated clinical scenario. The effect path mapping subunit internally stores a pre-constructed drug effect path map, which is a directed network structure. The nodes represent physiological states or drug action targets, and the edges represent the causal influence relationship between the nodes and are labeled with the weight of the influence strength. For example, there may be a path in the map that norepinephrine acting on the vascular alpha receptor node leads to the vascular contraction node, which in turn leads to the mean arterial pressure rise node, where the edge from vascular contraction to mean arterial pressure rise is labeled as strong effect. Another path may be that dopamine acting on the heart beta1 receptor node leads to the myocardial contractility enhancement node, which in turn leads to the cardiac output enhancement node, and ultimately also affects the mean arterial pressure node.

[0045] The abnormal parameter identification subunit analyzes the feature subset sent by the physiological parameter coupling correction unit. Suppose the corrected feature subset shows a blood pressure coefficient of variation of 0.28, which is significantly higher than the baseline level of 0.11 for this patient. The heart rate variability low-frequency power spectral density is 350 ms2 / Hz, which is significantly lower than the baseline of 600 ms2 / Hz. The blood oxygen saturation rate of decline is 0.3% / s, which is within the normal range. The system compares each feature value with the individualized normal range established according to the patient's historical data, identifying the blood pressure coefficient of variation and the heart rate variability low-frequency power spectral density as abnormal physiological parameters that deviate from the normal range and are marked as abnormal physiological parameters. According to the type of abnormal parameters identified, the system matches in the drug effect pathway map. The blood pressure coefficient of variation anomaly is mainly associated with nodes reflecting blood pressure stability, such as mean arterial pressure nodes. The heart rate variability low-frequency power spectral density anomaly may be associated with nodes reflecting autonomic nervous function, which may be located in upstream pathways affecting cardiac function or vascular tone. The reverse search subunit is then started, with its search direction opposite to the natural propagation direction of physiological effects. Instead of tracking effects downstream from drug targets, it retraces upstream from the identified abnormal parameter nodes against the direction of directed edges, seeking potential drug targets that can correct these abnormalities through effect chains. The search algorithm traverses all possible upstream paths and records the nodes and edges passed through each path. For example, reverse searching from the abnormal mean arterial pressure node may find vasoconstriction and cardiac output change nodes as its direct upstream impact factors. The algorithm evaluates the length of these paths and the effect strength weights of the edges in the paths.

[0046] After locating the potential drug target nodes, the instruction generation subunit begins to calculate the infusion priority of different target drugs. The priority score takes into account multiple factors, including the number of abnormal parameters pointing to the same target node, the length of the path, and the weight of the effect strength on the path. Suppose the calculation result shows that the intervention path for alpha receptors is shorter and can improve two abnormal parameters at the same time, with the highest priority score. The intervention path for beta1 receptors is longer and mainly only indirectly related to heart rate variability abnormalities, with a relatively low score. Based on this priority ranking, the current real-time physiological state, and the pharmacological characteristics of the drug, the instruction generation subunit generates specific infusion instructions. The instruction content explicitly includes drug category identification such as preferentially selecting norepinephrine that mainly stimulates alpha receptors, and recommended initial infusion flow rate such as 0.04 mcg / kg / min. The setting of this flow rate refers to the standard medication guidelines and takes into account the patient's body weight.

[0047] The safety boundary monitoring subunit enters the continuous monitoring state immediately after the infusion instruction starts execution. It sets dynamic safety boundary parameters for the patient, such as a systolic pressure lower limit of 85 mmHg and an upper limit of 160 mmHg, and a heart rate upper limit of 125 beats per minute. This subunit monitors the trends of these key parameters in real time. Suppose that after infusion of norepinephrine, the systolic pressure is gradually increased from 88 mmHg to 105 mmHg, but the heart rate is increased from 90 beats per minute to 120 beats per minute, close to the pre-set safety upper limit. At this time, the safety boundary monitoring subunit will intervene according to the built-in control logic. It may dynamically reduce the infusion flow rate from 0.04 micrograms per kilogram per minute to 0.03 micrograms per kilogram per minute, in order to maintain the effect of increasing blood pressure while avoiding the risk of increasing tachycardia. If the parameters change abnormally sharply or exceed the safety boundary, such as a sudden increase in systolic pressure exceeding 170 mmHg or the occurrence of ventricular arrhythmia, the subunit will immediately trigger the instruction regeneration process. It will pause the current infusion and force the decision unit to perform a full analysis and decision-making cycle based on the latest real-time physiological data. The drug effect path graph itself is also not completely static. The system is designed with a slow adaptive learning mechanism, which can adjust the weights of the edges in the graph according to the long-term cumulative infusion events and actual physiological response effects data, so that the decision logic of the system can gradually approach the specific response characteristics of the individual. The entire decision-making process constitutes a closed-loop system that includes perception, analysis, decision-making, execution, and feedback. The core design principle is to pursue the effect of intervention while placing patient safety at the highest priority.

[0048] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application. The scope of the application is defined by the appended claims and their equivalents.

Claims

1. A smart infusion system for vasoactive drugs based on real-time physiological data, characterized in that, The system integrates a physiological data acquisition unit, a physiological state classification unit, an abnormal state marking unit, a drug response characteristic decomposition unit, a physiological parameter coupling correction unit, and a drug infusion decision unit. The physiological data acquisition unit is configured to acquire real-time physiological data and perform preprocessing operations. The physiological state classification unit is configured to classify the preprocessed physiological data according to a preset physiological state interval, and extract multiple physiological features within each physiological state interval to generate a set of state features. The abnormal state marking unit is configured to monitor the switching frequency of physiological state intervals and mark the abnormal state interval when the switching frequency exceeds a preset threshold. The drug response feature decomposition unit is configured to decompose the set of state features of non-abnormal state intervals based on the similarity matching results between historical drug response stages and current physiological features, and generate a subset of drug response features. The physiological parameter coupling correction unit is configured to use the coupling relationship between physiological parameters to correct the parameters of a subset of drug response features. The drug infusion decision unit is configured to perform a reverse traversal along the drug effect path based on the modified drug response feature subset to locate potential drug needs, and determine the infusion priority and generate infusion instructions based on the traversal results.

2. The intelligent infusion system for vasoactive drugs according to claim 1, characterized in that, The physiological data acquisition unit includes a signal synchronization acquisition subunit, a timestamp alignment subunit, and a data filtering and cleaning subunit; The signal synchronization acquisition subunit is configured to synchronously acquire arterial blood pressure waveform data, electrocardiogram signal data, and pulse oximetry data; the timestamp alignment subunit is configured to perform timestamp alignment processing on the arterial blood pressure waveform data, electrocardiogram signal data, and pulse oximetry data. The data filtering and cleaning subunit is configured to use the mean filtering method to remove high-frequency fluctuation noise in the arterial blood pressure waveform data, and to remove abnormal data points based on the temporal correlation between electrocardiogram signal data and pulse oxygen saturation data; the processed arterial blood pressure waveform data, electrocardiogram signal data and pulse oxygen saturation data are resampled at a fixed sampling frequency to output a standardized physiological data sequence.

3. The intelligent infusion system for vasoactive drugs according to claim 1, characterized in that, The physiological state classification unit includes a state interval division subunit and a feature fusion subunit; The state interval division subunit is configured to divide the standardized physiological data sequence into hypotension state interval, normal blood pressure state interval and hypertension state interval according to the systolic blood pressure value; the feature fusion subunit is configured to extract time domain features and frequency domain features from the arterial blood pressure waveform data, electrocardiogram signal data and pulse oxygen saturation data in each physiological state interval, respectively. The time-domain and frequency-domain features within the same physiological state interval are fused using a sliding time window to generate a multi-dimensional set of state features.

4. The intelligent infusion system for vasoactive drugs according to claim 3, characterized in that, The time-domain characteristics of arterial blood pressure waveform data include blood pressure coefficient of variation, blood pressure trend line slope, and blood pressure fluctuation entropy. The frequency-domain characteristics of electrocardiogram signal data include low-frequency power spectral density and high-frequency power spectral density of heart rate variability. The time-domain characteristics of pulse oximetry data include the rate of decrease in blood oxygen saturation and the amplitude of blood oxygen fluctuation.

5. The intelligent infusion system for vasoactive drugs according to claim 1, characterized in that, The abnormal state marking unit includes a switching frequency calculation subunit and a state consistency evaluation subunit; The switching frequency calculation subunit is configured to count the number of switching intervals within a fixed duration window, calculate the switching frequency per minute, and compare it with a frequency threshold. If the switching frequency exceeds the frequency threshold, the state consistency evaluation subunit will align the physiological characteristics of adjacent physiological state intervals using a waveform alignment algorithm and calculate the cross-state consistency score. If the cross-state consistency score is lower than the consistency threshold, the physiological state interval corresponding to the current window is marked as an abnormal state interval and the feature re-extraction process is triggered; otherwise, it is marked as a normal state interval.

6. The intelligent infusion system for vasoactive drugs according to claim 5, characterized in that, The calculation of the cross-state consistency score involves calculating the mutual information values ​​for the aligned blood pressure coefficient of variation, heart rate variability low-frequency power spectral density, and blood oxygen saturation decrease rate, and then using the weighted average of the mutual information values ​​as the cross-state consistency score.

7. The intelligent infusion system for vasoactive drugs according to claim 1, characterized in that, The drug response feature decomposition unit includes a multimodal feature fusion subunit and an adaptive decomposition subunit; The multimodal feature fusion subunit is configured to align the current physiological features in the state feature set of non-abnormal state intervals with the time series features of historical drug response stages using a dynamic time warping algorithm, and calculate the comprehensive matching degree by combining the weighted result of cosine similarity metric and dynamic time warping distance; the historical drug response stage labels are divided by blood pressure waveform complexity coefficient and heart rate variability spectrum kurtosis feature; if the comprehensive matching degree is higher than the matching threshold, the adaptive decomposition subunit uses the ensemble empirical mode decomposition algorithm to extract intrinsic mode function components, and selects feature frequency components related to vasoactive drug response to form a subset of drug response features; If the overall matching degree is lower than the matching threshold, then based on the correlation characteristics between heart rate variability multifractal spectrum and drug metabolism kinetics, wavelet packet transform is used to extract energy features of specific frequency bands, and nonlinear features extracted by recursive quantitative analysis are combined to generate a subset of drug response features.

8. The intelligent infusion system for vasoactive drugs according to claim 1, characterized in that, The physiological parameter coupling correction unit includes a vascular response correction subunit and a metabolic effect correction subunit; The vascular response correction subunit corrects the blood pressure variability coefficient in the drug response feature subset based on the linear relationship between vascular resistance and blood pressure. The metabolic effect correction subunit corrects the low-frequency power spectral density of heart rate variability in the drug response feature subset based on the nonlinear mapping between basal metabolic rate and heart rate variability. The corrected blood pressure variability coefficient and heart rate variability low-frequency power spectral density are combined with the blood oxygen saturation decrease rate feature, and the output parameters are coupled with the corrected drug response feature subset.

9. The intelligent infusion system for vasoactive drugs according to claim 1, characterized in that, The drug infusion decision unit includes an effect path mapping subunit, an abnormal parameter identification subunit, a reverse search subunit, and an instruction generation subunit. The effect path mapping subunit is configured to acquire a drug effect path map, which is constructed based on the physiological parameter causal chain and drug receptor binding path, and includes the effect propagation direction and intensity weight of the drug action node; abnormal physiological parameters are identified from the parameter coupling corrected drug response feature subset, and the corresponding node in the drug effect path map is matched according to the type of abnormal physiological parameter; By searching backwards along the propagation direction of the drug effect pathway map, potential drug target nodes with abnormal physiological parameters can be located. The infusion priority is calculated based on the number of abnormal parameters and path hierarchy of potential drug target nodes; based on the infusion priority and real-time physiological status, an infusion command containing drug type identification and infusion flow rate is generated.

10. The intelligent infusion system for vasoactive drugs according to claim 9, characterized in that, The drug infusion decision unit also includes a safety boundary monitoring subunit, configured to monitor changes in physiological parameters in real time after the execution of the infusion command. If the physiological parameters exceed the preset safety boundary, the infusion rate is dynamically adjusted or the command regeneration process is triggered.

Images