Medical multi-source data fusion method based on digital twinning

Abstract

The application discloses a medical multi-source data fusion method based on digital twinning, and particularly relates to the technical field of medical data processing, and comprises the following steps: constructing a multi-scale heterogeneous medical data flow access and metadata labeling channel, synchronously accessing three types of data sources of continuous physiological waveform flow, device parameter change log and electronic medical record test value, and establishing a semantic dictionary tree; performing dynamic discovery of causal time delay and calculation of fusion weight based on time-varying transfer entropy, online calculating time-varying transfer entropy from a drug rate sequence to a blood pressure waveform variability sequence, introducing conditional transfer entropy to strip the confusion effect of heart rate fluctuation, extracting a drug-blood pressure causal response delay time, and constructing a dynamic fusion weight matrix; the application solves the defects of causal time sequence misalignment, asynchronous frequency data rhythm misfire and deduced results violating physiological principles in the prior art, and realizes full-closed loop application from multi-source data semantic fusion to autonomous tracking of treatment strategies.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and more specifically, to a method for fusing multi-source medical data based on digital twins. Background Technology

[0002] Digital twin technology, by constructing virtual mapping models of physical entities, offers a new technological path for achieving precise and personalized medical decision-making. In complex clinical scenarios such as intensive care, the construction of digital twins relies on the effective fusion of multi-source heterogeneous medical data. This data includes high-frequency continuous physiological signals such as arterial blood pressure waveforms and electrocardiogram waveforms collected by monitors, as well as discrete operation logs and parameter change records generated by life support equipment such as infusion pumps, syringe pumps, and ventilators, and low-frequency electronic medical record data such as blood gas analysis and biochemical tests. However, these multi-source data differ significantly in sampling frequency, temporal resolution, data structure, and semantic hierarchy. How to achieve deep fusion of heterogeneous data within a unified framework is a fundamental technical challenge in constructing medical digital twins with clinical credibility and physiological fidelity.

[0003] Existing methods for fusing multi-source medical data often employ a static data mapping strategy based on timestamp alignment. This involves aligning data from different sources according to physical time labels and storing them in a data lake, then extracting information through feature engineering or deep learning models. These methods have the following significant drawbacks: First, relying solely on physical timestamp alignment ignores the inherent pharmacokinetic delays and causal response lags between different physiological variables. For example, there is a physiological response delay of tens of seconds or even minutes between changes in the injection rate of vasoactive drugs and the resulting changes in blood pressure waveforms. Simple time alignment can lead to causal misalignment in the reconstructed patient state in the digital twin. Second, existing fusion methods typically use fixed time windows as data aggregation units, ignoring the inherent rhythmic characteristics of the human physiological system. When patients experience arrhythmias or decoupling of spontaneous breathing and mechanical ventilation rhythms, the fixed window method can cause "mismatched" fusion of data from different frequencies, disrupting the continuity and integrity of the physiological state representation. Third, the digital twin inference models constructed by existing methods are mostly purely data-driven, lacking constraints on the physiological conservation laws of the circulatory system. The inference results may violate basic physiological principles, making it difficult to provide reliable predictive support in high-risk clinical decision-making scenarios.

[0004] In summary, existing medical multi-source data fusion methods have shortcomings in terms of temporal alignment at the causal semantic level, rhythmic isomorphism processing of heterogeneous data, and physiological consistency constraints on inference models. These limitations hinder the practical application of digital twin technology in precision treatment within intensive care. Therefore, there is an urgent need for a medical multi-source data fusion method that can achieve breakthrough improvements in causal delay detection, rhythmic relaxation alignment, and physiological equation constraints. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a medical multi-source data fusion method based on digital twins.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The digital twin-based medical multi-source data fusion method includes the following steps: Step 1: Construct a multi-scale heterogeneous medical data stream access and metadata annotation channel: synchronously access three types of data sources: continuous physiological waveform stream, equipment parameter change log, and electronic medical record test values; attach metadata tags to each data packet and establish a semantic dictionary tree. Step 2: Perform dynamic discovery and fusion weight calculation of causal delay based on time-varying transfer entropy: Calculate the time-varying transfer entropy from the drug rate sequence to the blood pressure waveform variability sequence online, determine the significance of the causal edge through surrogate data permutation test, introduce conditional transfer entropy to remove the confusion effect of heart rate fluctuation, extract the drug-blood pressure causal response delay time, and construct a dynamic fusion weight matrix. Step 3: Perform heterogeneous data fusion and state representation based on rhythmic isomorphism relaxation alignment: extract the comprehensive physiological rhythm benchmark as a dynamic alignment unit, perform unequal-length resampling and isomorphic mapping on heterogeneous data, and generate a rhythmically consistent multivariate feature vector sequence. Step 4: Construct a differentiable neural state inferrer constrained by physiological ordinary differential equations: With the encoder-decoder network as the backbone, set up a physical constraint calculation branch with embedded blood circulation ordinary differential equations in parallel, and use the residual between the network's predicted state and the theoretical derivative of the ordinary differential equation as a regularization term to train the network. Step 5: Proactive extrapolation of the consequences of intervention and autonomous tracking of the optimal strategy: The dosing regimen change is injected into the extrapolator, multiple dose adjustment trajectories are generated through Monte Carlo tree search, and the evolution of physiological state is extrapolated. The multi-objective evaluation function is used to converge to the optimal dose adjustment trajectory.

[0007] Specifically, the three types of data sources mentioned in step one are: The first category consists of arterial blood pressure waveforms, electrocardiogram waveforms, and bi-frequency exponential continuous physiological waveform streams sampled at a high frequency of 500Hz, which are buffered as equidistant time series. The second category is parameter change logs for smart infusion pumps, syringe pumps, and ventilators triggered by irregular discrete events, including changed parameter values, operation timestamps, and operation delay indicators. The third category consists of electronic medical record test values ​​collected at a fixed low frequency; the semantic dictionary tree enables the establishment of a potential fusionable relationship between the calculated systolic blood pressure value in the continuous waveform stream and the norepinephrine rate change in the infusion pump log, with the cardiovascular-drug response as the fulcrum.

[0008] Specifically, the calculation of the time-varying transfer entropy in step two employs an adaptive kernel method: Within each sliding window, recursive graph analysis was performed on the drug rate sequence and the mean arterial pressure waveform variability sequence to obtain the deterministic retention ratio. When the ratio was higher than 0.8, a Gaussian radial basis kernel was selected, and the kernel bandwidth was adaptively determined by the Silverman empirical criterion based on the median absolute deviation of the current window data. When the ratio is below 0.8, it automatically switches to the Laplace kernel function; The surrogate data permutation test employs an iterative amplitude-adjusted Fourier transform method to randomize the phase information of the original drug rate sequence, generating surrogate data that retains the same power spectrum and autocorrelation structure as the original sequence. The causal edge is determined to be statistically significant only when the original transfer entropy value is higher than the 95th percentile threshold of the surrogate data transfer entropy distribution.

[0009] Specifically, step two involves introducing conditional transfer entropy to remove the confusion effect of heart rate fluctuations, specifically as follows: Using the synchronously extracted heart rate sequence as a conditional variable, the conditional transfer entropy from the drug rate sequence to the mean arterial pressure waveform variability sequence was calculated, and a significance threshold was constructed by applying a surrogate data permutation test to the conditional transfer entropy. The causal purification decision rule is as follows: If the original transfer entropy is significant but the conditional transfer entropy is not significant, then the drug-blood pressure causal edge of the window is determined to be a heart rate-mediated spurious association and is removed. If both are significant, the maximum value of the standardized significance of the two is taken as the final significance score, and the original time delay estimate is corrected by the peak time delay of the conditional transit entropy. If only the conditional transitive entropy is significant, then the result of the conditional transitive entropy calculation replaces the original calculation result.

[0010] Specifically, in step two, if no effective causal edge is retained after the current window undergoes causal purification decision, the causal response delay time of the previous effective window is used as the estimated value of the current window, and a causal chain interruption flag is triggered. This flag is then appended as metadata to the fusion weight matrix of the current window.

[0011] Specifically, the extraction of the comprehensive physiological rhythm benchmark in step three employs a priority decision-making logic and an anomaly tolerance mechanism: The dominant sympathetic-parasympathetic frequency identified by heart rate variability spectral density is set as the main rhythm benchmark. The frequency variation coefficient of the main rhythm is monitored in real time. If it remains below 15% for 10 consecutive cycles, the main rhythm is used as the dynamic alignment benchmark unit alone. If the frequency variation coefficient exceeds 15% or atrial fibrillation or other extremely irregular rhythm events are detected, a multi-rhythm source weighted voting mechanism is activated. The reciprocals of the main rhythm frequency, the EEG bifrequency exponential slow wave oscillation period, and the ventilator tidal volume envelope period are used as weighting factors. After multiplying by the confidence scores calculated by the free signal-to-noise ratio, the weighted average is obtained to obtain the comprehensive rhythm period. For abnormal states where central sleep apnea leads to loss of respiratory rhythm, the corresponding oscillation period term is automatically excluded from the weighted calculation, and the alignment frame is maintained only with the remaining effective rhythm sources.

[0012] Specifically, the isomorphism mapping process described in step three is as follows: The high-frequency waveforms within each physiological rhythm cycle are compressed into a feature vector through extreme value sampling and morphological feature encoding, and the causal effects of drugs acting within the same rhythm cycle are integrated into an intensity scalar. Meanwhile, in order to preserve the differences in drug action at different phases of the cardiac cycle, the comprehensive physiological rhythm cycle is used as a benchmark to divide it into multiple phase intervals. The drug causal influence curve obtained in step two is converted into a phase domain representation and integrated in each phase interval to obtain the drug effect phase vector. The drug effect phase vector, together with the eigenvector and the effect intensity scalar, constitutes a complete fusion representation of the rhythmic cycle, and anchors the discrete blood gas analysis results at the end of the rhythmic cycle.

[0013] Specifically, the structure of the physical constraint calculation branch in step four is as follows: The encoder-decoder network outputs a six-dimensional state vector at each future time step, with each dimension representing the predicted arterial blood pressure, cardiac output, systemic vascular resistance index, heart rate, drug effect-site concentration, and total blood volume estimate, respectively. The embedded blood circulation ordinary differential equation includes three coupled equations reflecting the drug-vascular tension relationship, the cardiac output-peripheral resistance relationship, and the baroreflex sensitivity modulation. The initial value of the penalty factor in the regularization term is set to 1.0, and it is automatically adjusted according to the rate of decrease of the physical residual on the validation set after each preset number of training rounds. If the rate of decrease of physical residuals is lower than the preset proportion of the rate of decrease of data prediction error, the penalty factor is amplified; if the physical residuals have decreased to below the preset threshold, the penalty factor remains unchanged.

[0014] Specifically, the multi-objective evaluation function described in step five includes: The calculation maximizes the proportion of time the mean arterial pressure is maintained within the preset target window, minimizes the integral of the total exposure dose of vasoactive drugs, and penalizes the risk of reperfusion injury due to excessive blood pressure variability. The extrapolation process is adjusted by the patient-specific drug-blood pressure causal response delay time parameter discovered in step two, so that the extrapolated blood pressure response curve has a precise individualized drug efficacy lag.

[0015] The technical effects and advantages of this invention are as follows: This invention addresses the core shortcomings of existing medical multi-source data fusion methods—such as causal time sequence misalignment, rhythmic isomorphism relaxation alignment, and physiological ordinary differential equation constraint deduction—through synergistic innovation at three levels: dynamic discovery of causal delays, rhythmic misalignment in heterogeneous data, and deduction results that violate physiological principles. At the causal level, through joint analysis of time-varying transfer entropy and conditional transfer entropy, it can discover individualized causal delay times between drugs and physiological responses online and automatically remove the mediating effects of confounding variables such as heart rate. This avoids the problems of inverted causal relationships or spurious associations in traditional timestamp alignment methods, giving the reconstructed patient state of the digital twin an interpretable physiological causal semantic basis. The fusion weights are adjusted in real time according to the patient's state rather than being preset fixed values. At the rhythm level, an adaptively extracted comprehensive physiological rhythm benchmark is used instead of a fixed time window as the data aggregation unit. Under abnormal states such as arrhythmia and human-machine interaction, the alignment framework can be dynamically maintained through a multi-rhythm source weighted voting mechanism. At the same time, the causal influence of drugs is decomposed into a phase sensitivity vector within the rhythm cycle, preserving the pharmacodynamic differences in different cardiac phases. This fundamentally eliminates the misstep phenomenon and the loss of temporal information in multi-source heterogeneous frequency data fusion, providing a continuous, complete, and high-fidelity state representation sequence for the digital twin.

[0016] At the extrapolation level, the blood circulation physiological ordinary differential equation is embedded as a physical constraint branch into a data-driven neural network extrapolator. This ensures that the state prediction results are strictly constrained by the physiological conservation laws, eliminating the possibility of predictions that violate physiological principles from pure black-box models and improving the credibility of the extrapolation results in high-risk clinical decision-making scenarios. At the application level, the fused patient-specific causal delay time parameters and phase-sensitive pharmacodynamic characterization are injected into the proactive extrapolation process. Through Monte Carlo tree search and Bayesian optimization, the system autonomously follows the optimal treatment strategy in the continuous space of drug dosage. The extrapolated blood pressure response curve contains precise individualized pharmacodynamic lag information, and the objective function simultaneously considers multiple balances of treatment effect, drug exposure dose, and reperfusion injury risk. These technical effects together constitute a closed-loop technical chain from multi-source data semantic fusion to autonomous tracking of treatment strategies, significantly enhancing the practical application value of digital twin technology in precision treatment in intensive care. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] 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.

[0019] like Figure 1 As shown, the steps of the medical multi-source data fusion method based on digital twins are as follows: Step 1: Construct a multi-scale heterogeneous medical data stream access and metadata annotation channel.

[0020] A multi-protocol adapter gateway is deployed in the hospital bedside information system to simultaneously access three types of data sources: The first type is continuous physiological waveform streams such as human arterial blood pressure waveforms, electrocardiogram waveforms, and bispectral index of electroencephalogram (EEG) sampled at a high frequency of 500Hz, cached as equidistant time series; the second type is parameter change logs of intelligent infusion pumps, syringe pumps, and ventilators triggered by irregular discrete events, including variable parameter values, operation timestamps, and operation delay identifiers; the third type is electronic medical record test values ​​collected at a fixed low frequency, such as arterial blood gas analysis results once per hour. A unified five-tuple metadata tag is attached to each access data packet, including patient identification, device physical channel identification, NTP synchronization timestamp from the hospital time center, data semantic type encoding, and sampling rhythm feature value. A preliminary semantic dictionary is established based on the data semantic type encoding, enabling a potential fusion relationship between the calculated systolic blood pressure value in the continuous waveform stream and the norepinephrine rate change in the infusion pump log, centered on "cardiovascular-drug response," rather than relying solely on timestamp concatenation features. This step provides a high-quality data source foundation with well-defined semantic boundaries for subsequent dynamic causal alignment.

[0021] Step 2: Perform dynamic discovery of causal delay and calculation of fusion weights based on time-varying propagation entropy.

[0022] An online analysis framework with a 30-second sliding window and a 5-second step size is constructed for the continuous physiological waveform stream and discrete event logs obtained in step one.

[0023] Within each window, the discrete parameter change logs of the infusion pump are resampled into a stepped time series with 5-second intervals. An adaptive kernel-based transfer entropy algorithm is constructed to address the nonlinear and non-stationary characteristics of intensive care data. Specifically, within each 30-second sliding window, recursive graph analysis is first performed on the "drug rate sequence" and the "mean arterial pressure waveform variability sequence" to obtain the deterministic retention ratio. When this ratio is higher than 0.8, the sequence is considered to have high local stationarity, and a Gaussian radial basis kernel is used to calculate the transfer entropy. The kernel bandwidth is adaptively determined using the Silverman empirical criterion based on the median absolute deviation of the current window's data. When the deterministic retention ratio is lower than 0.8, the sequence is considered to have strong local non-stationarity, and a Laplace distribution-based kernel function is automatically switched to tolerate heavy-tailed noise. Within a time delay range of 0 to 120 seconds, the time-varying transfer entropy from the "drug rate sequence" to the "mean arterial pressure waveform variability sequence" is calculated. A surrogate data permutation test is introduced during the calculation process. The randomization method for this test is as follows: while maintaining the amplitude distribution of the original sequence, an iterative amplitude-adjusted Fourier transform method is used to randomize the phase information of the original drug rate sequence, generating 100 sets of surrogate data that retain the same power spectrum and autocorrelation structure as the original sequence. Only when the transfer entropy value calculated between the original sequences is higher than the 95th percentile threshold of the transfer entropy distribution among these 100 sets of surrogate data is the causal edge considered statistically significant, and its transfer entropy value is retained for subsequent weight calculations.

[0024] After obtaining the significance determination results of the drug-blood pressure transfer entropy, an indirect causal stripping mechanism based on conditional transfer entropy is further introduced to eliminate the potential mediating effect of heart rate fluctuation as a confounding variable on the above causal relationship. Let the heart rate sequence extracted synchronously within the current window be... , calculation When it is a conditional variable, from the drug rate sequence To mean arterial pressure waveform variability sequence Conditional transitive entropy Its definition is: ,in To determine the conditional transfer entropy from variable X to variable Y, given the variable Z, Let be the value of the blood pressure variability sequence at time t+1. Let be the joint probability distribution function. Let be the conditional probability distribution function. It is a logarithmic function. The historical embedding dimensions for blood pressure variability, drug rate, and heart rate sequence are adaptively determined using nearest neighbor estimation based on mutual information. The historical embedding vectors for blood pressure variability, drug rate, and heart rate are respectively used, and each embedding dimension is adaptively determined using nearest neighbor estimation based on mutual information. A permutation test of 100 sets of phase-randomized surrogate data is applied to the conditional transit entropy to construct its significance threshold. Subsequently, a causal purification decision is performed: if the original transit entropy is significant but the conditional transit entropy is not significant, the drug-blood pressure causal edge in that window is determined to be a spurious association mediated by heart rate and is removed; if both are significant, the maximum of the standardized significance of the two is taken as the final significance score of the causal edge, and the original time delay estimate is corrected by the peak time delay of the conditional transit entropy; if only the conditional transit entropy is significant, the calculated result of the conditional transit entropy replaces the original calculation result.

[0025] If no effective causal edge is retained after the current window undergoes causal purification decision, the causal response delay time of the previous effective window is used as the estimated value of the current window, and the "causal chain interruption" flag is triggered. This flag is then appended as metadata to the fusion weight matrix of the current window.

[0026] The time delay value corresponding to the maximum transfer entropy detected within each window is used as the real-time drug-blood pressure causal response delay time under the current physiological state. Then, a dynamic fusion weight matrix is ​​constructed based on the causal delay times of each parameter: for drug parameters with a causal delay of 30 seconds, the blood pressure waveform feature points from 30 seconds prior are weighted and fused, with the weight proportional to the standardized significance of the transfer entropy; for parameters without significant causal correlation, independent time series mappings are maintained. Simultaneously, a time-perceptron network based on an attention mechanism is employed to learn a differentiable continuous time-delay influence kernel function, smoothing discrete causal delay information into a continuous influence decay curve on the time axis. This allows heterogeneous data with different time resolutions to be strictly aligned at the causal semantic level, replacing traditional physical timestamps, and achieving a realistic reconstruction of multidimensional data based on physiological response dynamics within the twin space.

[0027] Step 3: Perform heterogeneous data fusion and state representation based on rhythmic isomorphism relaxation alignment.

[0028] For the heterogeneous data with causal influence decay curves obtained in step two, the data streams are fused according to their intrinsic physiological rhythms rather than absolute frequencies. The instantaneous heart rate variability spectral density is extracted in real time from the electrocardiogram waveform to identify the dominant frequency of the current sympathetic-parasympathetic tone. The slow wave oscillation period is extracted from the bispectral index of the electroencephalogram and used together with the periodic envelope of the ventilator tidal volume waveform as the current physiological rhythm benchmark.

[0029] In this step, when extracting the "comprehensive physiological rhythm benchmark," the following priority decision logic and anomaly tolerance mechanism are maintained: The sympathetic-parasympathetic dominant frequency identified by heart rate variability spectral density is set as the primary rhythm benchmark. The frequency variation coefficient of this primary rhythm is continuously monitored in real time. If it remains below 15% for 10 consecutive cycles, this primary rhythm is used alone as the dynamic alignment benchmark unit. If the frequency variation coefficient exceeds 15% or extremely irregular rhythm events such as atrial fibrillation are detected, a multi-rhythm source weighted voting mechanism is activated. The reciprocals of the primary rhythm frequency, the bispectral exponential slow-wave oscillation period of EEG, and the tidal volume envelope period of the ventilator are used as weighting factors. These factors are multiplied by their respective confidence scores and then weighted averaged to obtain the comprehensive rhythm cycle. The confidence score is calculated from the signal-to-noise ratio of the current segment of each waveform. For abnormal states such as central sleep apnea that cause respiratory rhythm loss, the corresponding oscillation period term is automatically excluded from the weighted calculation, and only the remaining effective rhythm sources are used to maintain the alignment framework. After obtaining the aforementioned comprehensive physiological rhythm baseline, a piecewise aggregation approximation method is employed. Using this comprehensive physiological rhythm cycle as the dynamic alignment reference unit, rather than a fixed time window, unequal-length resampling and isomorphic mapping are performed on the arterial pressure waveform segments, the drug causal influence curve, and the discrete points of the test indicators. Specifically, the high-frequency waveforms within each physiological rhythm cycle are compressed into a feature vector through extreme value sampling and morphological feature encoding. The causal influence of the drugs acting within the same rhythm cycle is integrated into an intensity scalar. Simultaneously, to preserve the differences in drug action across different phases of the cardiac cycle, the comprehensive physiological rhythm cycle is divided into M equal phase intervals (M=12). The drug causal influence curve obtained in step two is then... Convert to phase domain representation The drug effect phase vector is obtained by integrating within each phase interval. , ,in Let be the m-th component of the drug effect phase vector D, representing the cumulative intensity of the drug effect within the m-th phase interval. Let m be the total number of phase intervals, and m be the index of the phase interval. The integration domain is... Indicates the first One phase interval, For time delay variables, For phase variables, In the phase interval definite integrals on, Let be an integral infinitesimal element, representing the phase variable. Integration. This phase vector, together with the eigenvector and the intensity scalar, constitutes a complete fusion representation of the rhythm cycle, and anchors the discrete blood gas analysis results at the end of the rhythm cycle. All heterogeneous data are suppressed into a rhythmically consistent multivariate eigenvector sequence under a unified relaxation rhythm framework, completely eliminating the misstep phenomenon caused by heterogeneous sampling frequencies. The resulting isomorphic sequence directly constitutes the state update trajectory of the digital twin virtual heart-circulatory system under nonlinear physiological rhythm modulation.

[0030] Step 4: Construct a differentiable neural state extrapolator constrained by physiological ordinary differential equations, serving as the core power engine for digital twins.

[0031] The design incorporates a hybrid computational graph, with the backbone being an encoder-decoder temporal extrapolation network composed of long short-term memory units, receiving the isomorphic multivariate feature vector sequence generated in step three as input. Each time the decoder predicts the future system state, a parallel physical constraint computation branch is added. This branch embeds a set of parameter-identified ordinary differential equations for blood circulation, including three coupled equations reflecting the drug-vascular tone relationship, the cardiac output-peripheral resistance relationship, and the baroreflex sensitivity modulation.

[0032] The specific structure of the physical constraint computation branch is as follows: The long short-term memory encoder-decoder network outputs a six-dimensional vector as the predicted state vector at each future time step. The dimensions are, in order, the predicted arterial blood pressure, cardiac output, systemic vascular resistance index, heart rate, drug effect-site concentration, and total blood volume estimate. This six-dimensional vector is simultaneously fed in two directions: firstly, into the decoder state transition for the next time step; and secondly, into the ordinary differential equation (ODE) physical computation branch. This ODE branch contains three coupled equations, specifically: ,in Let be the derivative of arterial blood pressure with respect to time, representing the instantaneous rate of change of arterial blood pressure. For cardiac output, The systemic vascular resistance index, Arterial blood pressure, Arterial time constant; ,in Heart rate, Baseline stroke volume, which is the baseline stroke volume unaffected by medication. This is the coefficient representing the effect of drug effect-site concentration on stroke volume. The drug effect-site concentration is calculated from the first-order rate constant of pharmacokinetics. This is a venous return curve function, whose input VR is the dimensionless value of the normalized total blood volume estimate. , Item by current The integral of the deviation from the set point is obtained. This is the coefficient representing the effect of drug concentration in the effect-site area on vascular tone. This refers to the drug effect-site concentration. Let AP, CO, and SVR be the pressure reflection sensitivity modulation coefficients. The theoretical derivative is calculated by substituting the three terms AP, CO, and SVR from the six-dimensional state vector predicted by the neural network into the above ordinary differential equation. (Theoretical derivative of arterial blood pressure calculated from ordinary differential equations) (Theoretical derivative of cardiac output calculated from ordinary differential equations). (The theoretical derivative of the systemic vascular resistance index calculated from the ordinary differential equation) and the mean square error between this and the three state change rates directly output by the decoder network are calculated. This mean square error is then multiplied by an externally set penalty factor. As a physical consistency regularization term, it is added to the total loss function, which is composed of prediction error and physical residual. The initial value is set to 1.0. During training, it is automatically adjusted every 10 epochs based on the rate of decrease of the physical residuals on the validation set: if the rate of decrease of the physical residuals is less than 20% of the rate of decrease of the data prediction error, then... Multiply by 1.5 times; if the physical residual has decreased to below the preset threshold, then maintain. This ensures that the neural network output is both capable of fitting complex individual patient data and does not violate the qualitative dynamic relationships defined by the physiological differential equations. When extrapolating the combined use of fluid resuscitation and vasoactive drugs in septic shock patients, the predictive curves of future blood pressure, cardiac output, and systemic vascular resistance index output by this extrapolator not only conform to the individualized patient response patterns learned through data-driven learning but are also strictly constrained by the parameter boundaries of the cardiovascular physical model, preventing abrupt state changes that violate physiological principles. This extrapolator constitutes the core power engine of a digital twin with high dynamic responsiveness and physical consistency.

[0033] Step 5: Proactive extrapolation of the consequences of intervention and autonomous tracing of the optimal strategy. In the digital twin engine, an intervention injection mechanism is implemented on the encoder-decoder extrapolation network defined in Step 4. When a clinician's proposed change in dosing regimen (e.g., "reduce the norepinephrine dose from 0.2 μg / kg / min to 0.1 μg / kg / min after 15 minutes") is input into the system, the extrapolator in Step 4 injects the drug change information as an additional control vector into the state transition calculation at each time step during its decoding phase. The simulation process employs Monte Carlo tree search, automatically generating 500 test dose adjustment trajectories in the extrapolator's continuous drug dose space using Bayesian optimization. Each trajectory extrapolates the physiological state evolution over the next 60 minutes. The extrapolation process is adjusted by the patient-specific drug-blood pressure causal response delay time parameter discovered in Step 2, ensuring that each extrapolated blood pressure response curve carries a precise and individualized drug hysteresis. During the search process, a multi-objective evaluation function is used to calculate the utility value of each item from the extrapolation results. The objective function simultaneously includes maximizing the proportion of time during which the mean arterial pressure is maintained within the target window of 65-75 mmHg in the future, minimizing the integral of the total exposure dose of vasoactive drugs, and penalizing the risk of reperfusion injury due to excessive blood pressure variability. Finally, a Monte Carlo tree search converges to an optimal dose adjustment trajectory in the policy space, which is directly transformed into an infusion pump-readable sequence of injection rate control instructions and output to the clinical auxiliary decision-making interface, completing a closed-loop application from multi-source data semantic fusion to autonomous tracking of treatment strategies.

[0034] The above formulas are all dimensionless calculations. Dimensionless calculations can be performed using various methods such as standardization, which will not be elaborated here. The formulas are derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas can be set by those skilled in the art according to the actual situation.

[0035] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0036] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0037] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0038] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0039] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0040] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0041] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0042] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A medical multi-source data fusion method based on digital twins, characterized in that, Includes the following steps: Step 1: Synchronously access three types of data sources: continuous physiological waveform stream, device parameter change log, and electronic medical record test values; attach metadata tags to each data packet and build a semantic dictionary tree; Step 2: Calculate the time-varying transfer entropy from the drug rate sequence to the blood pressure waveform variability sequence online, determine the significance of the causal edge through surrogate data permutation test, introduce conditional transfer entropy to remove the confusion effect of heart rate fluctuation, extract the drug-blood pressure causal response delay time, and construct a dynamic fusion weight matrix. Step 3: Extract the comprehensive physiological rhythm benchmark as a dynamic alignment unit, perform unequal-length resampling and isomorphic mapping on heterogeneous data, and generate a multi-dimensional feature vector sequence with consistent rhythm; Step 4: Using the encoder-decoder network as the backbone, set up a physical constraint calculation branch with embedded blood circulation ordinary differential equation in parallel, and use the residual between the network prediction state and the theoretical derivative of the ordinary differential equation as a regularization term to train the network. Step 5: Inject the change in the dosing regimen into the extrapolator, generate multiple dose adjustment trajectories through Monte Carlo tree search, extrapolate the evolution of physiological state, and converge to the optimal dose adjustment trajectory using a multi-objective evaluation function.

2. The medical multi-source data fusion method based on digital twins according to claim 1, characterized in that, The three types of data sources mentioned in step one are as follows: The first category consists of arterial blood pressure waveforms, electrocardiogram waveforms, and bi-frequency exponential continuous physiological waveform streams sampled at a high frequency of 500Hz, which are buffered as equidistant time series. The second category is parameter change logs for smart infusion pumps, syringe pumps, and ventilators triggered by irregular discrete events, including changed parameter values, operation timestamps, and operation delay indicators. The third category consists of electronic medical record test values ​​collected at a fixed low frequency; the semantic dictionary tree enables the establishment of a potential fusionable relationship between the calculated systolic blood pressure value in the continuous waveform stream and the norepinephrine rate change in the infusion pump log, with the cardiovascular-drug response as the fulcrum.

3. The medical multi-source data fusion method based on digital twins according to claim 1, characterized in that, The calculation of the time-varying transfer entropy in step two employs an adaptive kernel method: Within each sliding window, recursive graph analysis was performed on the drug rate sequence and the mean arterial pressure waveform variability sequence to obtain the deterministic retention ratio. When the ratio was higher than 0.8, a Gaussian radial basis kernel was selected, and the kernel bandwidth was adaptively determined by the Silverman empirical criterion based on the median absolute deviation of the current window data. When the ratio is below 0.8, it automatically switches to the Laplace kernel function; The surrogate data permutation test employs an iterative amplitude-adjusted Fourier transform method to randomize the phase information of the original drug rate sequence, generating surrogate data that retains the same power spectrum and autocorrelation structure as the original sequence. The causal edge is determined to be statistically significant only when the original transfer entropy value is higher than the 95th percentile threshold of the surrogate data transfer entropy distribution.

4. The medical multi-source data fusion method based on digital twins according to claim 1, characterized in that, Step two, which introduces conditional transfer entropy to remove the confusion effect of heart rate fluctuations, specifically involves: Using the synchronously extracted heart rate sequence as a conditional variable, the conditional transfer entropy from the drug rate sequence to the mean arterial pressure waveform variability sequence was calculated, and a significance threshold was constructed by applying a surrogate data permutation test to the conditional transfer entropy. The causal purification decision rule is as follows: If the original transfer entropy is significant but the conditional transfer entropy is not significant, then the drug-blood pressure causal edge of the window is determined to be a heart rate-mediated spurious association and is removed. If both are significant, the maximum value of the standardized significance of the two is taken as the final significance score, and the original time delay estimate is corrected by the peak time delay of the conditional transit entropy. If only the conditional transitive entropy is significant, then the result of the conditional transitive entropy calculation replaces the original calculation result.

5. The medical multi-source data fusion method based on digital twins according to claim 4, characterized in that, In step two, if no effective causal edge is retained after the causal purification decision of the current window, the causal response delay time of the previous effective window is used as the estimated value of the current window, and the causal chain interruption flag is triggered. This flag is then attached to the fusion weight matrix of the current window as metadata.

6. The medical multi-source data fusion method based on digital twins according to claim 1, characterized in that, The extraction of the comprehensive physiological rhythm benchmark in step three employs a priority decision-making logic and an anomaly tolerance mechanism: The dominant sympathetic-parasympathetic frequency identified by heart rate variability spectral density is set as the main rhythm benchmark. The frequency variation coefficient of the main rhythm is monitored in real time. If it remains below 15% for 10 consecutive cycles, the main rhythm is used as the dynamic alignment benchmark unit alone. If the frequency variation coefficient exceeds 15% or atrial fibrillation or other extremely irregular rhythm events are detected, a multi-rhythm source weighted voting mechanism is activated. The reciprocals of the main rhythm frequency, the EEG bifrequency exponential slow wave oscillation period, and the ventilator tidal volume envelope period are used as weighting factors. After multiplying by the confidence scores calculated by the free signal-to-noise ratio, the weighted average is obtained to obtain the comprehensive rhythm period. For abnormal states where central sleep apnea leads to loss of respiratory rhythm, the corresponding oscillation period term is automatically excluded from the weighted calculation, and the alignment frame is maintained only with the remaining effective rhythm sources.

7. The medical multi-source data fusion method based on digital twins according to claim 1, characterized in that, The specific process of isomorphism mapping described in step three is as follows: The high-frequency waveforms within each physiological rhythm cycle are compressed into a feature vector through extreme value sampling and morphological feature encoding, and the causal effects of drugs acting within the same rhythm cycle are integrated into an intensity scalar. Meanwhile, in order to preserve the differences in drug action at different phases of the cardiac cycle, the comprehensive physiological rhythm cycle is used as a benchmark to divide it into multiple phase intervals. The drug causal influence curve obtained in step two is converted into a phase domain representation and integrated in each phase interval to obtain the drug effect phase vector. The drug effect phase vector, together with the eigenvector and the effect intensity scalar, constitutes a complete fusion representation of the rhythmic cycle, and anchors the discrete blood gas analysis results at the end of the rhythmic cycle.

8. The medical multi-source data fusion method based on digital twins according to claim 1, characterized in that, The specific structure of the physical constraint calculation branch in step four is as follows: The encoder-decoder network outputs a six-dimensional state vector at each future time step, with each dimension representing the predicted arterial blood pressure, cardiac output, systemic vascular resistance index, heart rate, drug effect-site concentration, and total blood volume estimate, respectively. The embedded blood circulation ordinary differential equation includes three coupled equations reflecting the drug-vascular tension relationship, the cardiac output-peripheral resistance relationship, and the baroreflex sensitivity modulation. The initial value of the penalty factor in the regularization term is set to 1.0, and it is automatically adjusted according to the rate of decrease of the physical residual on the validation set after each preset number of training rounds. If the rate of decrease of physical residuals is lower than the preset proportion of the rate of decrease of data prediction error, the penalty factor is amplified; if the physical residuals have decreased to below the preset threshold, the penalty factor remains unchanged.

9. The medical multi-source data fusion method based on digital twin according to claim 1, characterized in that, The multi-objective evaluation function mentioned in step five includes: The calculation maximizes the proportion of time the mean arterial pressure is maintained within the preset target window, minimizes the integral of the total exposure dose of vasoactive drugs, and penalizes the risk of reperfusion injury due to excessive blood pressure variability. The extrapolation process is adjusted by the patient-specific drug-blood pressure causal response delay time parameter discovered in step two, so that the extrapolated blood pressure response curve has a precise individualized drug efficacy lag.

Images