Cross-mechanism referral intelligent routing method and system based on medical agent middleware

Abstract

The invention provides a cross-institution referral intelligent routing method and system based on medical agent middleware, and relates to the technical field of medical informationization, and the method comprises the steps: building a physical sign data monitoring window, carrying out the time sequence mapping of physical sign index changes and medication records, constructing a drug response sensitivity curve, and calculating a treatment effective coefficient; adopting a dynamic time warping algorithm to match the medicine-physical sign association map, and generating a treatment decision tree; disease progression is predicted based on the effective treatment coefficients and specialized specialities, and the optimal referral opportunity and target mechanism are identified. According to the invention, the cross-mechanism referral efficiency is improved, the medical resource configuration is optimized, and the patient treatment effect is enhanced.

Description

Technical Field

[0001] This invention relates to medical information technology, and in particular to a cross-institutional referral intelligent routing method and system based on medical intelligent agent middleware. Background Technology

[0002] With the advancement of the tiered medical service policy, inter-institutional referrals have become an important means of optimizing the allocation of medical resources. However, the current referral process between medical institutions still suffers from problems such as information asymmetry and a lack of scientific basis for referral decisions, making it difficult to guarantee the timeliness and accuracy of referrals.

[0003] Existing referral systems primarily rely on static rules for institution matching, failing to adequately consider individualized patient treatment response characteristics, leading to unpredictable subsequent treatment outcomes. Furthermore, data silos between healthcare institutions hinder the effective flow of patient clinical information, impacting referral quality and healthcare continuity.

[0004] The lack of intelligent referral decision support systems among medical institutions hinders the dynamic assessment of patient condition evolution and the determination of optimal referral timing. Medical staff also struggle to accurately evaluate the matching degree between the treatment capabilities of different specialty medical institutions and patient needs. Therefore, there is an urgent need for an intelligent cross-institutional referral routing method based on medical intelligent agent middleware. Summary of the Invention

[0005] This invention provides a cross-institutional referral intelligent routing method and system based on medical intelligent agent middleware, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a cross-institutional referral intelligent routing method based on medical intelligent agent middleware, comprising: The referral application information is received through the medical intelligent agent middleware. The referral application information includes patient medical record information and referral department information, wherein the patient medical record information includes vital sign data and medication records. Establish a monitoring window for vital signs data, map the rate of change of vital signs indicators to the time points and dosages in the medication records, identify the response delay time and peak level of vital signs indicators to different drug doses, construct a drug response sensitivity curve, and calculate the treatment effectiveness coefficient based on the drug response sensitivity curve; The system obtains diagnosis and treatment data of candidate medical institutions from the collaborative database of medical institutions, segments the drug response sensitivity curve according to the fluctuation cycle, extracts features from each segment of the curve to obtain time-series feature vectors, constructs a drug-sign association map, uses the dynamic time warping algorithm to calculate the similarity between the drug-sign association map and the historical association map in the diagnosis and treatment data, determines the set of sign indicators based on the similarity matching results, and generates a treatment decision tree containing the disease evolution path. Based on the treatment decision tree, the patient's condition trajectory is predicted. Based on the treatment effectiveness coefficient and the specialty of candidate medical institutions, the best referral time is identified, the target medical institution is determined, and the medical intelligent agent middleware establishes a data transmission channel to complete data exchange.

[0007] In one alternative embodiment, A monitoring window for vital signs data is established, and the rate of change of vital signs indicators is time-series mapped to the dosing time points and dosages in the medication records. The response delay time and peak level of vital signs indicators to different drug doses are identified, and drug response sensitivity curves are constructed, including: Wavelet denoising was performed on the vital signs data to obtain smoothed vital signs data. The initial monitoring window was divided according to the interval between adjacent dosing time points in the medication record. The window size was dynamically adjusted based on the autocorrelation characteristics of the vital signs data to obtain the windowed vital signs sequence. The windowed vital sign sequence is calculated by time difference to obtain the vital sign change rate sequence. The dosage in the medication record is normalized to obtain the dosage sequence. The probability distribution difference between the vital sign change rate sequence and the dosage sequence is calculated as the transmission loss. The time-series mapping relationship is obtained by minimizing the transmission loss. In the time-frequency decomposition of the time-series mapping relationship, the continuous trajectory with the highest energy density is extracted as the feature ridge. The starting point of the feature ridge is marked as the response delay time, and the maximum value point of the ridge is marked as the peak response level. A feature point set is constructed based on the response delay time and peak response level. The kernel function bandwidth is dynamically adjusted based on the local distribution pattern of the feature point set. The feature point set is then interpolated and fitted using the kernel function to construct a drug response sensitivity curve.

[0008] In one alternative embodiment, The calculation of the treatment effectiveness coefficient based on the drug response sensitivity curve includes: The drug response sensitivity curve is segmented according to the curvature change point to obtain the characteristic segment reflecting the drug action process. The mapping relationship between the change trend of vital signs and the drug dosage in each characteristic segment is calculated to generate dose response characteristics. A compensation coefficient is constructed, which varies with the sensitivity of vital signs to drug dosage. The compensation coefficient is multiplied by the dose response characteristics to generate a compensated response value. Based on the compensated response values, an impulse response sequence is constructed, which reflects the response pattern of vital signs to a single dose. The superposition effect of the impulse response sequence is analyzed to obtain drug accumulation characteristics. The drug accumulation characteristics are combined with the peak values ​​of the impulse response sequence to obtain the therapeutic efficacy coefficient.

[0009] In one alternative embodiment, The drug response sensitivity curve is segmented according to its fluctuation period to extract features from each segment, resulting in a time-series feature vector. A drug-sign association map is then constructed, including: Wavelet transform is performed on the drug response sensitivity curve to obtain periodic features. Based on the periodic features, fluctuation time windows are identified, and the drug response sensitivity curve is divided into multiple subsequences according to the fluctuation time windows. Extract the temporal feature vector of the subsequence, and use principal component analysis to reduce the dimensionality of the temporal feature vector to obtain the dimensionality-reduced feature vector; An initial heterogeneous graph network is constructed based on the dimensionality-reduced feature vector, which includes drug nodes and vital sign nodes. The correlation value between the drug nodes and vital sign nodes is calculated using the Pearson correlation coefficient, and the correlation value is used as the initial weight between the drug nodes and vital sign nodes. The initial weights are dynamically adjusted using a graph attention mechanism to obtain dynamic weights. The delay response value corresponding to the dynamic weights is calculated by combining time coding. A dynamic heterogeneous graph network is constructed based on the dynamic weights and delay response value. A multi-head attention mechanism is used to extract node interaction features from the dynamic heterogeneous graph network. A drug-sign association map is constructed based on the node interaction features.

[0010] In one alternative embodiment, The dynamic time warping algorithm is used to calculate the similarity between the drug-sign association map and the historical association map in the diagnosis and treatment data. Based on the similarity matching results, the set of sign indicators is determined, and a treatment decision tree containing the disease evolution path is generated, including: The drug-sign association map is constructed as a time series matrix. A time series matching window is set, and a dynamic time warping algorithm is executed within the time series matching window to calculate the initial distance value between the time series matrix and the historical association map in the diagnosis and treatment data. Construct a strip penalty function, and perform weighted correction on the initial distance value according to the strip penalty function to obtain the corrected distance value. Calculate the corrected distance values ​​at different time scales and perform weighted combination to obtain the map similarity value. Historical cases are clustered and grouped according to the map similarity value. The pattern of vital signs change is extracted from the clusters to obtain the vital sign feature sequence. The information correlation between features in the vital sign feature sequence is calculated. The importance of the vital sign feature sequence is quantitatively scored based on the information correlation. Features with importance quantitative scores higher than the preset score are selected to construct a set of vital sign indicators. Decision tree nodes are constructed based on the set of vital signs indicators. In the decision tree nodes, indicator judgment intervals, drug response probabilities, and time window parameters are set to generate an initial decision tree. The initial decision tree is pruned and optimized based on the principle of minimum information content. Structural constraint terms are introduced to limit the branch size of the initial decision tree to obtain the treatment decision tree of the disease evolution path.

[0011] In one alternative embodiment, Based on the principle of minimum information content, the initial decision tree is pruned and optimized. Structural constraints are introduced to limit the branch size of the initial decision tree, resulting in a treatment decision tree for the disease evolution path, including: The classification results, node transition time, and subtree information are extracted from the non-leaf nodes of the initial decision tree to construct a node information vector containing temporal consistency constraints. The node information vector is then weighted and combined based on the temporal consistency constraints to generate the node information quantity. The information change value of the non-leaf node under the principle of minimum information is calculated, and an adaptive threshold curve is constructed based on the depth distribution and sample distribution of the initial decision tree. The information change value is projected onto the adaptive threshold curve for comparison. When the projection result is lower than the adaptive threshold curve, pruning is performed to obtain the pruned decision tree. Extract the number of leaf nodes, tree depth, and path complexity of the pruning decision tree, and combine them to construct structural constraints. Based on the structural constraints, perform dynamic programming search on the pruning decision tree to generate multiple candidate decision trees. Feature statistics are performed on the nodes of the candidate decision trees, and the upper bound of information gain is calculated in combination with the disease evolution pattern. The candidate decision trees are then screened according to the upper bound of information gain, and the decision tree with the smallest information gain is selected as the optimal decision tree. Boundary optimization is performed on the branches of the optimal decision tree to generate a treatment decision tree containing the disease evolution path.

[0012] In one alternative embodiment, Based on the prediction of patient disease progression using a treatment decision tree, and considering the treatment effectiveness coefficient and the specialty strengths of candidate medical institutions, the optimal referral timing is identified, and target medical institutions are determined, including: The pre-acquired set of vital signs indicators is decomposed temporally and features are extracted. The results are then input into the treatment decision tree to calculate the state transition probability and transition time of each node. A multidimensional state space is constructed based on the state transition probability and transition time. In the multidimensional state space, Monte Carlo sampling is performed on the state transition sequence in combination with the treatment effectiveness coefficient to generate a weighted set of disease progression trajectories; Historical treatment cases of candidate medical institutions are extracted from the collaborative database of medical institutions. These historical treatment cases are clustered according to disease type and treatment stage. The treatment success rate and disease duration of each category are calculated to generate a specialty matrix. The disease progression trajectory set is mapped to a specialty matrix, the trajectory-specialty matching degree is calculated, and the candidate medical institutions are dynamically sorted based on the matching degree. At the same time, a referral trigger threshold is set according to the treatment effectiveness coefficient. When the treatment effectiveness coefficient is lower than the referral trigger threshold and there are candidate medical institutions with a matching degree higher than the preset matching threshold, the current moment is determined as the best time for referral, and the medical institution with the highest matching degree is selected as the target medical institution.

[0013] A second aspect of this invention provides a cross-institutional referral intelligent routing system based on medical intelligent agent middleware, comprising: The first unit is used to receive referral application information through the medical intelligent agent middleware. The referral application information includes patient medical record information and referral department information, wherein the patient medical record information includes vital sign data and medication records. The second unit is used to establish a monitoring window for vital signs data, to time-series map the rate of change of vital signs indicators with the time points and doses in the medication records, to identify the response delay time and peak level of vital signs indicators to different drug doses, to construct a drug response sensitivity curve, and to calculate the treatment effectiveness coefficient based on the drug response sensitivity curve. The third unit is used to obtain the diagnosis and treatment data of candidate medical institutions from the collaborative database of medical institutions, segment the drug response sensitivity curve according to the fluctuation cycle, extract the features of each segment of the curve to obtain the time-series feature vector, construct the drug-sign association map, use the dynamic time warping algorithm to calculate the similarity between the drug-sign association map and the historical association map in the diagnosis and treatment data, determine the set of sign indicators based on the similarity matching results, and generate a treatment decision tree containing the disease evolution path; The fourth unit is used to predict the patient's disease progression trajectory based on the treatment decision tree, identify the best referral time based on the treatment effectiveness coefficient and the specialty of candidate medical institutions, determine the target medical institution, and complete the data exchange by establishing a data transmission channel through the medical intelligent agent middleware.

[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] In this embodiment, intelligent routing for cross-institutional referrals is achieved through a medical intelligent agent middleware. This enables the construction of drug response sensitivity curves and drug-sign correlation maps, accurately assessing patients' response characteristics to treatment, improving the precision of referral decisions, and avoiding treatment delays and resource waste caused by blind referrals. By dynamically monitoring the temporal relationship between changes in sign data and medication records, combined with a dynamic time warping algorithm, this invention achieves precise identification of individualized treatment response patterns in patients, scientifically predicting disease progression trajectories, and providing data-driven decision support for patient triage between medical institutions, significantly improving the rationality and timeliness of the referral process. The medical intelligent agent middleware of this invention establishes a secure and efficient data transmission channel between different medical institutions, achieving seamless exchange of medical data while protecting patient privacy, promoting the optimal allocation of medical resources, and providing technical support for building a hierarchical medical system and improving the overall quality of medical services. It has significant social benefits and clinical application value. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the intelligent routing method for cross-institutional referrals based on medical intelligent agent middleware, as described in an embodiment of the present invention. Figure 2 A schematic diagram of the system architecture for constructing drug-sign association maps; Figure 3 This is a simulation diagram illustrating the similarity between drug-sign association maps and the importance of sign indicators. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the cross-institutional referral intelligent routing method based on medical intelligent agent middleware according to an embodiment of the present invention. Figure 1 As shown, the method includes: The referral application information is received through the medical intelligent agent middleware. The referral application information includes patient medical record information and referral department information, wherein the patient medical record information includes vital sign data and medication records. Establish a monitoring window for vital signs data, map the rate of change of vital signs indicators to the time points and dosages in the medication records, identify the response delay time and peak level of vital signs indicators to different drug doses, construct a drug response sensitivity curve, and calculate the treatment effectiveness coefficient based on the drug response sensitivity curve; The system obtains diagnosis and treatment data of candidate medical institutions from the collaborative database of medical institutions, segments the drug response sensitivity curve according to the fluctuation cycle, extracts features from each segment of the curve to obtain time-series feature vectors, constructs a drug-sign association map, uses the dynamic time warping algorithm to calculate the similarity between the drug-sign association map and the historical association map in the diagnosis and treatment data, determines the set of sign indicators based on the similarity matching results, and generates a treatment decision tree containing the disease evolution path. Based on the treatment decision tree, the patient's condition trajectory is predicted. Based on the treatment effectiveness coefficient and the specialty of candidate medical institutions, the best referral time is identified, the target medical institution is determined, and the medical intelligent agent middleware establishes a data transmission channel to complete data exchange.

[0021] The medical intelligent agent middleware receives referral requests from medical institutions via a standardized interface protocol. This request information is encapsulated in JSON format and includes a unique patient identifier, basic information fields, complete medical record data, and the target referral department identifier. The vital signs data in the patient's medical record includes time-series records of physiological parameters such as blood pressure, heart rate, body temperature, and blood oxygen saturation. Each record includes a measurement timestamp, value, measuring device identifier, and data quality identifier. Medication records detail the drug name, administration time, dosage, administration method, and patient response, forming a complete medication timeline.

[0022] In one optional implementation, a monitoring window for vital signs data is established, and the rate of change of vital signs is time-series mapped to the dosing time points and doses in the medication records. The response delay time and peak level of vital signs to different drug doses are identified, and a drug response sensitivity curve is constructed, including: Wavelet denoising was performed on the vital signs data to obtain smoothed vital signs data. The initial monitoring window was divided according to the interval between adjacent dosing time points in the medication record. The window size was dynamically adjusted based on the autocorrelation characteristics of the vital signs data to obtain the windowed vital signs sequence. The windowed vital sign sequence is calculated by time difference to obtain the vital sign change rate sequence. The dosage in the medication record is normalized to obtain the dosage sequence. The probability distribution difference between the vital sign change rate sequence and the dosage sequence is calculated as the transmission loss. The time-series mapping relationship is obtained by minimizing the transmission loss. In the time-frequency decomposition of the time-series mapping relationship, the continuous trajectory with the highest energy density is extracted as the feature ridge. The starting point of the feature ridge is marked as the response delay time, and the maximum value point of the ridge is marked as the peak response level. A feature point set is constructed based on the response delay time and peak response level. The kernel function bandwidth is dynamically adjusted based on the local distribution pattern of the feature point set. The feature point set is then interpolated and fitted using the kernel function to construct a drug response sensitivity curve.

[0023] In one embodiment, wavelet denoising is performed on the vital signs data to obtain smoothed data. Specifically, the db4 wavelet basis function is used to decompose the original vital signs data into five levels. A threshold of 0.05 is set for high-frequency coefficients, and soft thresholding is applied to coefficients exceeding the threshold. Then, wavelet reconstruction is performed to obtain smoothed vital signs data. For example, for a patient's blood pressure data, the original value is [135, 142, 138, 146, 140, 143, 141, 138] mmHg. After wavelet denoising, the values ​​are [136.5, 140.2, 139.3, 143.8, 141.2, 142.1, 140.5, 138.8] mmHg, effectively eliminating measurement noise.

[0024] The initial monitoring window is defined based on the interval between adjacent dosing times in the medication record. For example, if a patient administers 100 mg of drug A at 08:00 and 50 mg of drug A at 14:00, the initial monitoring window is set to 6 hours. Considering the potential individual differences in drug metabolism rates, the window size is dynamically adjusted based on the autocorrelation characteristics of the vital signs data. Specifically, the autocorrelation function of the vital signs data is calculated, and the time point when the autocorrelation coefficient first drops below 0.3 is used as the actual window boundary. In practical applications, if the autocorrelation coefficient of blood pressure data drops to 0.28 after 5 hours, the monitoring window is adjusted to 5 hours. In this way, multiple windowed vital sign sequences are generated for different dosing events, each sequence corresponding to a complete drug response process.

[0025] The windowed vital sign sequence is calculated by time difference to obtain the vital sign change rate sequence. For blood pressure data with a sampling interval of 10 minutes, the difference between two adjacent measurements is calculated and divided by the time interval to obtain the blood pressure change rate sequence, in mmHg / hour. For example, the original blood pressure values ​​[140, 142, 145, 143, 140] mmHg at adjacent 10-minute sampling points are converted to a change rate of [12, 18, -12, -18] mmHg / hour. At the same time, the dosage in the medication record is normalized, for example, the dosage of drug A is divided by the maximum recommended dose of 200 mg to obtain the normalized dose sequence [0.5, 0.25].

[0026] The difference in probability distribution between the rate of change of vital signs sequence and the dose sequence is calculated as the transmission loss. Specifically, kernel density estimation is used to estimate the probability distributions of the rate of change of vital signs and the normalized dose, respectively, and the Wasserstein distance between these two distributions is calculated as the transmission loss. By iteratively adjusting the time offset, the time offset value that minimizes the transmission loss is found, establishing a temporal mapping relationship between the rate of change of vital signs and the drug dose. In the example, for drug A, it was found that the transmission loss is minimized when the time offset is 45 minutes, meaning that the drug's effect on blood pressure begins to appear approximately 45 minutes after administration.

[0027] Feature ridges were extracted from the time-frequency decomposition of the temporal mapping relationship. Continuous wavelet transform was used to perform time-frequency analysis on the mapped vital sign change rate sequence, using Morlet wavelets as the mother wavelet with a scale parameter set to 1-64. Feature ridges were extracted by identifying the continuous region with the highest energy density in the transform result. In the above patient case, the starting point of the feature ridges on the time axis was located at 45 minutes after drug administration, marked as the response delay time; the maximum point of the ridges occurred at 2.5 hours after drug administration, at which time the blood pressure change rate reached 20 mmHg / hour, marked as the peak response level.

[0028] Based on the analysis of multiple dosing records, multiple sets of response delay times and peak response levels are collected to construct a feature point set. For example, for three dosing records of different doses (50 mg, 100 mg, 150 mg), the extracted feature point set might be {(50, 50, 15), (100, 45, 20), (150, 40, 28)}, representing the dose, response delay time (minutes), and peak response level (mmHg / hour). The kernel function bandwidth is dynamically adjusted based on the local distribution pattern of the feature point set. Specifically, the average distance between adjacent points in the feature point set is calculated, and the bandwidth is increased in sparse regions and decreased in dense regions. An adaptive bandwidth Gaussian kernel function is used to interpolate and fit the feature point set to construct a drug response sensitivity curve.

[0029] Drug response sensitivity curves are represented by two functions: the dose-response delay function and the dose-peak response function. These two functions can predict the drug response characteristics at any dose. For example, for drug A, the predicted response delay at a dose of 75 mg is approximately 47 minutes, and the peak response level is approximately 17.5 mmHg / hour. Healthcare professionals can use these curves to customize personalized dosing regimens for patients. For example, for patients who need rapid blood pressure reduction, a dose that quickly reaches the peak response can be selected; for patients with long-term stable blood pressure control, a dose with a moderate response delay and a mild peak response can be selected.

[0030] In this embodiment, high-precision modeling of the response relationship between patient vital signs data and drug dosage can be achieved, thereby accurately depicting the dynamic impact of drugs on individual vital signs. Wavelet denoising improves the stability of vital sign data, and adaptive windowing enhances the specificity and sensitivity of time-series analysis. By minimizing the transmission loss between the rate of change of vital signs and the normalized dose, automatic correlation between the time lag and intensity response of different drug actions is achieved, thus avoiding the subjectivity of traditional experience-based judgments. Extracting high-energy feature ridges from time-frequency decomposition accurately locates key time points and response amplitudes of vital sign responses, aiding in the analysis of the rhythmicity of drug efficacy. Furthermore, the introduction of a kernel function interpolation method based on adaptive bandwidth according to local distribution patterns makes the constructed drug response sensitivity curve more continuous and interpretable, effectively supporting subsequent individualized treatment effect evaluation and dynamic referral decisions, and improving the intelligence and response efficiency of cross-institutional medical collaboration.

[0031] In one optional implementation, calculating the treatment effectiveness coefficient based on the drug response sensitivity curve includes: The drug response sensitivity curve is segmented according to the curvature change point to obtain the characteristic segment reflecting the drug action process. The mapping relationship between the change trend of vital signs and the drug dosage in each characteristic segment is calculated to generate dose response characteristics. A compensation coefficient is constructed, which varies with the sensitivity of vital signs to drug dosage. The compensation coefficient is multiplied by the dose response characteristics to generate a compensated response value. Based on the compensated response values, an impulse response sequence is constructed, which reflects the response pattern of vital signs to a single dose. The superposition effect of the impulse response sequence is analyzed to obtain drug accumulation characteristics. The drug accumulation characteristics are combined with the peak values ​​of the impulse response sequence to obtain the therapeutic efficacy coefficient.

[0032] In this embodiment, the therapeutic efficacy can be more accurately assessed by calculating the treatment efficacy coefficient based on the drug response sensitivity curve. This method includes steps such as segmented analysis of the drug response sensitivity curve, calculation of dose-response characteristics, construction of compensation coefficients, generation of impulse response sequences, and analysis of drug accumulation characteristics.

[0033] Segmentation of the drug response sensitivity curve is fundamental to calculating the therapeutic efficacy coefficient. First, vital sign data of patients under different drug doses are collected, and drug response sensitivity curves are plotted. By calculating the second derivative of the curve, points of significant curvature change are identified as segmentation points. For example, for antihypertensive drugs, when the dose increases from 5 mg to 10 mg, the rate of blood pressure decrease may change from 2 mmHg per hour to 3.5 mmHg per hour; the inflection point of this rate of change can be used as a segmentation point. In practical applications, data analysis of a group of hypertensive patients shows that a typical antihypertensive drug response curve can be divided into three characteristic segments: an initial slow response segment (0-5 mg), a rapid response segment (5-15 mg), and a stable response segment (above 15 mg).

[0034] Piecewise linear fitting was used to calculate the dose-response characteristics for each characteristic segment. Within each segment, the ratio of the rate of change in vital signs to the dose was calculated. Taking antihypertensive drugs as an example, in the rapid response segment (5-15 mg), the rate of blood pressure reduction might be 1.2 mmHg per milligram dose, while in the steady response segment (above 15 mg), this ratio might drop to 0.5 mmHg per milligram. In this way, dose-response characteristic values ​​were generated for each segment, reflecting the pharmacodynamic characteristics of different dose ranges.

[0035] The compensation coefficients were constructed to account for individual differences in drug response. Compensation models were built based on patients' physiological parameters (such as age, weight, and liver and kidney function). The compensation coefficients were adjusted according to changes in the sensitivity of these physiological indicators to drug dosage. For example, for patients with hepatic impairment, whose drug metabolism capacity is reduced, the compensation coefficient might be between 1.2 and 1.5; while for patients with renal impairment, whose drug excretion is impaired, the compensation coefficient might reach 1.8 to 2.0. Practical application data shows that the average compensation coefficient for elderly patients over 60 years of age is 1.35, while the compensation coefficient for healthy adults aged 30-40 years is close to 1.0. Multiplying these compensation coefficients by the previously calculated dose-response characteristics yields the compensated response value, making the assessment results more consistent with individual circumstances.

[0036] A pulse response sequence was constructed based on the compensated response values ​​to reflect the dynamic response of vital signs to a single drug administration. Time series analysis was used to record the changing patterns of vital signs over time. Taking oral hypoglycemic agents as an example, the pulse response sequence of blood glucose changes after a single administration may exhibit a pattern of "first decrease - reaching the lowest point - slowly recovering," with a complete response cycle of 8 hours. The lowest point is reached 2.5 hours after administration, with a blood glucose decrease of 2.3 mmol / L. The system constructed a complete pulse response sequence by sampling and recording blood glucose values ​​at time points of 0.5h, 1h, 2h, 4h, 6h, and 8h.

[0037] Analyzing drug accumulation characteristics is crucial for evaluating long-term drug efficacy. This study simulates the superposition of residual effects from previous doses and the effects of new doses under multiple-dose scenarios. Through convolution operations, the impulse response sequences of a single dose are time-shifted and superimposed to obtain the cumulative effect curve for multiple doses. Actual data show that for a drug with a half-life of 12 hours, a twice-daily dosing regimen reaches steady-state plasma concentration by day five, with an accumulation coefficient of approximately 1.8, meaning the steady-state plasma concentration is 1.8 times the peak concentration of a single dose.

[0038] The therapeutic efficacy coefficient (TEC) calculation integrates all the above factors. It is generated by weighting the drug accumulation characteristics with the peak values ​​of the impulse response sequence. This coefficient comprehensively reflects both the immediate intensity of the drug's effect and its long-term cumulative effect. For example, for a lipid-lowering drug, if the peak impulse response after a single dose results in a 0.8 mmol / L reduction in total cholesterol and the drug accumulation characteristic value is 1.5, then the calculated TEC is 1.2 (derived through a specific weighting algorithm), indicating that the drug has a good therapeutic effect at the recommended dose.

[0039] Based on the above technical solutions, quantitative assessment of drug treatment effects at the individual patient level can be achieved, thereby assisting in intelligent treatment pathway planning and referral decisions. By segmenting the curvature change points of the drug response sensitivity curve, key dynamic features in the drug action process are extracted, enhancing the ability to express nonlinear drug response patterns. Furthermore, combining dose response characteristics with compensation coefficients for response sensitivity adjustment effectively improves the model's adaptability to individual differences, making the calculation results closer to real clinical responses. By constructing impulse response sequences, the system characterizes the response patterns of vital signs to a single dose, and overlays and analyzes the cumulative effect of the drug during continuous administration, thereby comprehensively judging the correlation between peak vital signs and cumulative drug efficacy, forming a quantifiable treatment effectiveness coefficient. This coefficient reflects both treatment intensity and considers the delay and cumulative changes in vital sign response, which can be used to support more accurate efficacy prediction and institution selection, improving the timeliness and scientific rigor of cross-institutional referrals.

[0040] In one optional implementation, the drug response sensitivity curve is segmented according to its fluctuation period, and features are extracted from each segment to obtain a time-series feature vector. The drug-sign association map is then constructed, including: Wavelet transform is performed on the drug response sensitivity curve to obtain periodic features. Based on the periodic features, fluctuation time windows are identified, and the drug response sensitivity curve is divided into multiple subsequences according to the fluctuation time windows. Extract the temporal feature vector of the subsequence, and use principal component analysis to reduce the dimensionality of the temporal feature vector to obtain the dimensionality-reduced feature vector; An initial heterogeneous graph network is constructed based on the dimensionality-reduced feature vector, which includes drug nodes and vital sign nodes. The correlation value between the drug nodes and vital sign nodes is calculated using the Pearson correlation coefficient, and the correlation value is used as the initial weight between the drug nodes and vital sign nodes. The initial weights are dynamically adjusted using a graph attention mechanism to obtain dynamic weights. The delay response value corresponding to the dynamic weights is calculated by combining time coding. A dynamic heterogeneous graph network is constructed based on the dynamic weights and delay response value. A multi-head attention mechanism is used to extract node interaction features from the dynamic heterogeneous graph network. A drug-sign association map is constructed based on the node interaction features.

[0041] Figure 2 A schematic diagram of the system architecture for constructing a drug-sign correlation map is provided. For example, wavelet transform is applied to the drug response sensitivity curve to obtain periodic features. Specifically, discrete wavelet transform is used to decompose the original sensitivity curve, and a suitable wavelet basis function, such as the Daubechies wavelet, is selected to decompose the signal into coefficients in different frequency bands. Wavelet coefficients are filtered by setting a threshold to retain significant fluctuation information. For example, for a drug response curve containing 1000 time points, a fourth-level wavelet decomposition is used to obtain the wavelet coefficient matrix. By analyzing the energy distribution of the wavelet coefficients, the main periodic components can be identified. Based on the distribution of local extrema of the wavelet coefficients, fluctuation periods are identified; for example, for a certain antibiotic, two main fluctuation periods, 24 hours and 72 hours, may be identified. The size of the time window is determined according to the identified fluctuation characteristics; for example, a 12-hour window unit is set to divide the entire curve into several subsequences.

[0042] For each subsequence, a time-series feature vector is extracted. The extracted features fall into two main categories: statistical features and morphological features. Statistical features include indicators such as mean, standard deviation, kurtosis, skewness, and quartiles; morphological features include the number of peaks, fluctuation amplitude, rate of increase, rate of decrease, and curve curvature. For example, for a 12-hour response window of a certain antihypertensive drug, the extracted feature vector might contain over 20 dimensions of features, such as the average blood pressure reduction of 2.3 mmHg, standard deviation of 1.1 mmHg, maximum reduction of 4.7 mmHg, and rate of increase of 0.15 mmHg / hour. Due to the high dimensionality of the original features, principal component analysis (PCA) is used for dimensionality reduction. First, all features are standardized, the feature covariance matrix is ​​calculated, and the eigenvalues ​​and eigenvectors of this matrix are solved. An appropriate number of principal components are selected based on the cumulative contribution rate (usually those with a cumulative contribution rate exceeding 85%), and the original features are mapped to the principal component space. For example, the original 20-dimensional features are reduced to an 8-dimensional representation, retaining 88.7% of the original data's information.

[0043] An initial heterogeneous graph network is constructed based on the dimensionality-reduced feature vectors. The graph network contains two types of nodes: drug nodes and sign nodes. Each drug node represents a drug, with attributes including drug molecular fingerprint, mechanism of action classification, and half-life. Each sign node represents a physiological indicator, such as blood pressure, heart rate, or blood glucose, with attributes including measurement units and normal range values. To establish connections between nodes, the Pearson correlation coefficient is calculated for each drug-sign pair. Specifically, a large amount of data on changes in signs after drug administration is collected. For drug A and sign B, the sequence of sign B changes in all patients treated with drug A is extracted, and the correlation coefficient is calculated between the dimensionality-reduced feature vectors and the original sign values. For example, the correlation coefficient between the drug mebendazole and body temperature might be 0.73, and the correlation coefficient with heart rate might be 0.32. A correlation threshold (e.g., 0.3) is set, and only connections with correlations higher than the threshold are retained. The correlation coefficient is then used as the initial weight of the edges.

[0044] A graph attention mechanism is employed to dynamically adjust the initial weights. A multi-layer graph attention network is constructed, with each layer containing multiple attention heads. Each attention head independently learns the importance weights between nodes. For drug and vital sign nodes, a query vector, key vector, and value vector are first generated through a linear transformation. The dot product of the query vector and key vector is calculated to obtain the original attention score. The softmax function is then applied to normalize the score to obtain the attention weight. This attention weight represents the importance between different nodes and, combined with the initial network weights, yields the dynamically adjusted edge weights. For example, the initial weight of drug A and vital sign C is 0.45, which is adjusted to 0.58 after learning through the graph attention mechanism, indicating that this association is more important than initially assessed.

[0045] The delayed response value is calculated using time coding. To capture the temporal dynamics of drug action, a time coding mechanism is introduced. Absolute time positions are encoded into high-dimensional vectors, and different frequencies of sine and cosine functions are used to capture changes at different time scales. The onset time and peak time of drug action are calculated, constructing a time difference matrix. For example, drug B has an onset time of 1.5 hours, a peak time of 4 hours, and a duration of 10 hours for its effect on symptom D. These temporal features are combined with dynamic weights to obtain the temporal correlation strength containing delayed response information.

[0046] A dynamic heterogeneous graph network is constructed based on dynamic weights and delayed response values. In this network, nodes represent drugs or vital signs, edges represent the relationships between them, and the edge weights contain information about both the strength of the relationship and the temporal response characteristics. A multi-head attention mechanism is applied to this network, with each attention head responsible for capturing different types of node interaction patterns. The outputs of the multi-head attention are aggregated to obtain a comprehensive node representation vector. For example, using an 8-head attention mechanism, each attention head outputs 64-dimensional features, which, after aggregation, result in 512-dimensional node interaction features.

[0047] A final drug-sign association map was constructed based on node interaction features. Cluster analysis was performed on the node interaction features to identify drug groups with similar action patterns and associated sign groups. A threshold for association strength was set to retain significant associations. The final map contains 80 drug nodes, 35 sign nodes, and 320 valid edges. Each edge is labeled with attributes such as association strength, onset delay, peak time, and duration. The map structure is presented using visualization technology, with edges of different thicknesses and colors representing associations of different strengths, facilitating physicians to quickly identify the patterns and temporal characteristics of drug effects on specific signs.

[0048] In this embodiment, by introducing a fluctuation period division and a multi-level feature extraction mechanism, the accuracy and expressive power of modeling the relationship between drug action process and vital sign response are improved. Existing technologies often use static statistical features or fixed time windows to model the correlation between drug and vital sign data, making it difficult to fully capture the dynamic changes in vital sign response under different periods. Furthermore, when constructing drug-vital sign maps, the influence of response delay and weight dynamics is generally ignored, resulting in rigid map structures and insufficient expression of individual differences. This application obtains the periodic change features in the drug response sensitivity curve through wavelet transform and constructs a fluctuation time window accordingly to achieve adaptive recognition of the rhythm of drug efficacy response. Further extraction and dimensionality reduction of the temporal feature vectors of each subsequence enhances the compactness and generalization ability of the feature representation. During map construction, Pearson correlation is first used as static weights to establish the basic connection between drug and vital signs. Then, a graph attention mechanism is used to dynamically adjust the weights, and time coding is combined to extract delayed response information, enabling the finally constructed dynamic heterogeneous graph network to simultaneously express structural correlation and response temporality. By further extracting node interaction features through a multi-head attention mechanism, the constructed drug-sign association map not only retains key associations but also enhances the ability to express individual treatment response patterns, significantly improving its application value in referral pathway optimization and treatment strategy selection.

[0049] In one optional implementation, a dynamic time warping algorithm is used to calculate the similarity between the drug-sign association map and the historical association map in the diagnostic and treatment data. Based on the similarity matching results, a set of sign indicators is determined, and a treatment decision tree containing the disease evolution path is generated, including: The drug-sign association map is constructed as a time series matrix. A time series matching window is set, and a dynamic time warping algorithm is executed within the time series matching window to calculate the initial distance value between the time series matrix and the historical association map in the diagnosis and treatment data. Construct a strip penalty function, and perform weighted correction on the initial distance value according to the strip penalty function to obtain the corrected distance value. Calculate the corrected distance values ​​at different time scales and perform weighted combination to obtain the map similarity value. Historical cases are clustered and grouped according to the map similarity value. The pattern of vital signs change is extracted from the clusters to obtain the vital sign feature sequence. The information correlation between features in the vital sign feature sequence is calculated. The importance of the vital sign feature sequence is quantitatively scored based on the information correlation. Features with importance quantitative scores higher than the preset score are selected to construct a set of vital sign indicators. Decision tree nodes are constructed based on the set of vital signs indicators. In the decision tree nodes, indicator judgment intervals, drug response probabilities, and time window parameters are set to generate an initial decision tree. The initial decision tree is pruned and optimized based on the principle of minimum information content. Structural constraint terms are introduced to limit the branch size of the initial decision tree to obtain the treatment decision tree of the disease evolution path.

[0050] In practice, the dynamic time warping algorithm is used to calculate the similarity between the drug-sign association map and the historical association map in the diagnosis and treatment data. Based on the similarity matching results, the set of sign indicators is determined, and finally a treatment decision tree containing the disease evolution path is generated.

[0051] The system acquires patient medication records and vital sign data during diagnosis and treatment to construct a drug-vital sign correlation graph, reflecting the influence of specific drugs on vital sign indicators. This correlation graph is then converted into a time series matrix, where each row represents a drug, each column represents a time point, and the matrix element values ​​indicate the degree of influence of the drug on each vital sign indicator at that time point. For example, if drug A has an influence value of 0.75 on vital sign X at time point t1, it means there is a 75% probability of causing a change in vital sign X. The system sets the time series matching window to 14 days. Within this window, a dynamic time warping algorithm is executed to calculate the initial distance between the current time series matrix and the correlation graph in historical diagnosis and treatment data. This distance is obtained by accumulating the point-to-point numerical differences; for example, for two sequences of length n, the square root of the sum of the squares of the differences between each pair of corresponding points is used as the initial distance value.

[0052] A strip-shaped penalty function is constructed to weight and correct the initial distance value. This penalty function is defined as a weight function that increases with the time difference between matching points. When the time difference between matching points is less than 3 days, the penalty coefficient is 1.0; when the time difference is between 3 and 7 days, the penalty coefficient is 1.5; and when the time difference exceeds 7 days, the penalty coefficient is 2.0. The corrected distance value is obtained by multiplying the initial distance value by the corresponding penalty coefficient. The system calculates the corrected distance value at different time scales, including daily, weekly, and monthly scales, assigning weights of 0.5, 0.3, and 0.2 respectively for weighted combination, ultimately obtaining the map similarity value. For example, for patient case P, the corrected distance value with historical case H is 0.25 at the daily scale, 0.32 at the weekly scale, and 0.40 at the monthly scale. Therefore, the weighted map similarity value is 1 - (0.25 × 0.5 + 0.32 × 0.3 + 0.40 × 0.2) = 0.695.

[0053] Based on the calculated map similarity values, historical cases are clustered. A hierarchical clustering algorithm is used, with a similarity threshold of 0.7; cases with similarity higher than this threshold are grouped together. From the clustering results, the system extracts the vital sign change patterns within each group, forming a vital sign feature sequence. For example, in a certain cluster, blood pressure, heart rate, and blood oxygen saturation are found to have significant temporal change patterns. The system calculates the information correlation between these features, using mutual information as the evaluation index; higher mutual information indicates a stronger correlation between the two features. For each vital sign feature, an importance quantification score is performed based on its mutual information with other features and its predictive ability in treatment outcomes. The score uses a standardized score of 0-100; for example, blood pressure has an importance score of 85, heart rate 78, and blood oxygen saturation 92. The system selects features with importance scores higher than a preset score of 80 to construct a set of vital sign indicators; in this example, blood pressure and blood oxygen saturation are included.

[0054] Decision tree nodes are constructed based on the selected set of vital signs indicators. Each node is configured with three key parameters: indicator judgment interval, drug response probability, and time window parameter. For example, for blood pressure, the judgment interval is set to "systolic blood pressure > 140 mmHg and diastolic blood pressure > 90 mmHg", the drug response probability is "antihypertensive drug A: 78%, antihypertensive drug B: 65%", and the time window parameter is "expected response time: 24-48 hours". After generating the initial decision tree, the system performs pruning optimization based on the principle of minimum information gain. This principle calculates the information gain of each node, and prunes nodes when the information gain is below 0.05. Simultaneously, the system introduces structural constraints to limit the branch size of the decision tree, such as limiting the maximum depth of the decision tree to 5 and each node to a maximum of 3 child nodes. Through these optimization steps, a treatment decision tree containing the disease evolution path is finally generated, which can provide personalized treatment plan suggestions based on the patient's real-time vital signs data.

[0055] In this embodiment, by transforming the drug-sign association map into a time series matrix and introducing a dynamic time warping algorithm, sequence alignment is performed within a limited matching window, effectively identifying the dynamic trends of different historical maps. The robustness of similarity calculation is enhanced through a banded penalty function and a multi-scale weighting mechanism, making map similarity more discriminative. Based on this, cluster analysis is used to mine patterns of sign changes in historical cases, and importance is quantified by combining information relevance, effectively screening sign indicators that highly contribute to disease evolution, enhancing the interpretability of the subsequent decision-making model and the reliability of the decision-making basis. The final constructed treatment decision tree not only reflects the key nodes and response patterns of the disease evolution path but also controls model complexity through pruning and structural constraints, improving the stability and generalization ability of the decision path, providing an efficient and reliable support for disease trajectory prediction and precision treatment in cross-institutional referrals.

[0056] In one optional implementation, the initial decision tree is pruned and optimized based on the principle of minimum information content. A structural constraint term is introduced to limit the branch size of the initial decision tree, resulting in a treatment decision tree for the disease evolution path, including: The classification results, node transition time, and subtree information are extracted from the non-leaf nodes of the initial decision tree to construct a node information vector containing temporal consistency constraints. The node information vector is then weighted and combined based on the temporal consistency constraints to generate the node information quantity. The information change value of the non-leaf node under the principle of minimum information is calculated, and an adaptive threshold curve is constructed based on the depth distribution and sample distribution of the initial decision tree. The information change value is projected onto the adaptive threshold curve for comparison. When the projection result is lower than the adaptive threshold curve, pruning is performed to obtain the pruned decision tree. Extract the number of leaf nodes, tree depth, and path complexity of the pruning decision tree, and combine them to construct structural constraints. Based on the structural constraints, perform dynamic programming search on the pruning decision tree to generate multiple candidate decision trees. Feature statistics are performed on the nodes of the candidate decision trees, and the upper bound of information gain is calculated in combination with the disease evolution pattern. The candidate decision trees are then screened according to the upper bound of information gain, and the decision tree with the smallest information gain is selected as the optimal decision tree. Boundary optimization is performed on the branches of the optimal decision tree to generate a treatment decision tree containing the disease evolution path.

[0057] For example, after the initial decision tree is constructed, key information is extracted from all non-leaf nodes for subsequent pruning and optimization. Each non-leaf node contains classification result data, specifically recording the disease classification label, classification confidence, and sample distribution corresponding to that node. The node transition time records the cumulative time information from the root node to the current node, reflecting the time dimension characteristics of disease evolution. The subtree information includes the number and depth distribution of all child nodes under that node, as well as the statistical information of the sample categories contained in the leaf nodes.

[0058] The temporal consistency constraint is constructed based on the disease evolution patterns in the medical field, requiring that the time transitions between adjacent nodes in the decision tree must conform to the natural progression of the disease. For decision trees of patients with acute myocardial infarction, the transition time from the symptom onset node to the electrocardiogram abnormality node is typically between 30 minutes and 2 hours. Node transitions exceeding this time range are considered to violate the temporal consistency constraint. By checking whether the transition time of each node is within the medical standard range, a temporal consistency weight value is assigned to the node. Nodes that comply with the constraint have a weight of 1.0, partially compliant nodes have a weight between 0.5 and 0.9, and non-compliant nodes have a weight below 0.5.

[0059] The construction of node information vectors encodes classification results, node transfer times, and subtree information into multidimensional vector forms. The classification results use one-hot encoding to represent disease categories, node transfer times are standardized to values ​​between 0 and 1, and subtree information is represented by statistical feature vectors, including the number of child nodes, average depth, and sample distribution entropy. These vector components are weighted and combined based on temporal consistency constraints. Weight allocation follows medical expertise, with key diagnostic nodes receiving higher weights and auxiliary examination nodes receiving relatively lower weights.

[0060] The generation of node information content is achieved by calculating the comprehensive complexity of the weighted node information vector. Information content reflects the importance and complexity of the node in the decision-making process. Nodes containing multiple disease classification possibilities have higher information content, while nodes corresponding to only a single, definitive diagnosis have lower information content. The system calculates the variance contribution of each vector component and obtains the overall information content value based on the weighting coefficients.

[0061] The change in information content is calculated by comparing the difference in information content of nodes before and after pruning. The system simulates the removal of a specific node and calculates the change in the overall information content of the decision tree. If the overall information content of the decision tree decreases significantly after removing a node, it indicates that the node plays an important role in the decision-making process and should not be pruned. Conversely, nodes whose information content changes very little after removal can be considered for pruning to simplify the decision tree structure.

[0062] The adaptive threshold curve is constructed based on the depth and sample distribution characteristics of the initial decision tree. Nodes at different depth levels have different pruning threshold standards. Shallow nodes typically contain more important diagnostic information, so they are given higher pruning thresholds, while deeper nodes may contain overfitting details, and are given relatively lower pruning thresholds. Sample distribution also affects the threshold setting; nodes with uniform sample distribution receive higher thresholds, while nodes with highly uneven sample distribution receive lower thresholds. The curve is constructed using a piecewise linear interpolation method to provide a smooth threshold transition between different depth levels of the tree.

[0063] The process of projecting the information content change value onto the adaptive threshold curve is achieved through coordinate mapping. The node's position is determined in two-dimensional space using node depth as the x-axis and the information content change value as the y-axis. When a node's position is below the threshold curve corresponding to its depth, the node is deemed to be pruned. For example, if a node at depth 5 has an information content change value of 0.3, and the threshold for that depth is 0.5, then the node is marked as a pruning candidate.

[0064] The pruning decision tree is generated by traversing the initial decision tree level by level and applying pruning rules. Pruning operations are performed from leaf nodes to the root node, ensuring that the pruning process does not disrupt the basic structure of the tree. The pruning operation replaces the selected subtree with a leaf node, and the leaf node's label is the most frequently occurring category label in that subtree. The integrity of the decision path is maintained during pruning, ensuring that each path from the root node to a leaf node represents a complete disease evolution trajectory.

[0065] The construction of structural constraints comprehensively considers multiple structural feature parameters of the pruned decision tree. The number of leaf nodes reflects the final classification ability of the decision tree; too many leaf nodes may lead to overfitting, while too few leaf nodes may lead to underfitting. Tree depth affects the complexity of the decision-making process; excessively deep trees increase decision-making time and computational complexity. Path complexity is measured by calculating the average length and variance of all paths from the root node to the leaf nodes; a decision tree with a uniform distribution of path lengths has better balance. The structural constraints combine these parameters into a comprehensive evaluation index to guide subsequent decision tree optimization.

[0066] Dynamic programming search further optimizes the pruned decision tree based on structural constraints, seeking the structurally optimal tree shape while maintaining decision accuracy. The search algorithm maintains a state transition table, recording the optimal substructure solution under different tree configurations. For each node, the algorithm calculates the overall objective function value under the two options of keeping or removing the node, selecting the scheme with the better objective function value. The search process generates multiple candidate decision trees satisfying different structural constraints, providing alternative solutions for subsequent selection.

[0067] The statistical features of candidate decision trees include node type distribution, branch factor statistics, path length distribution, and leaf node purity analysis. Node type distribution reflects the proportion of different types of diagnostic nodes in the tree; branch factor statistics calculate the distribution of the number of child nodes for each internal node; and path length distribution analyzes the distance changes from the root node to each leaf node. Leaf node purity is measured by calculating the concentration of sample classes in each leaf node; high purity leaf nodes indicate good classification performance.

[0068] The integration of disease evolution patterns is achieved by introducing constraints from medical expertise, maintaining a standard pattern library of disease evolution, including typical developmental stages, key time points, and patterns of complication occurrence for various diseases. The calculation of the upper bound of information gain is based on the expected information gain under the theoretically optimal condition, considering the class distribution of the dataset and the complexity of the feature space. For a dataset containing five disease categories, the theoretically optimal upper bound of information gain is determined by calculating the maximum information gain under a uniform distribution.

[0069] The candidate decision tree selection process evaluates and ranks trees based on the upper bound of their information gain. The actual information gain of each candidate tree on the validation dataset is calculated and compared to the theoretical upper bound. Decision trees with information gain close to the upper bound indicate good learning ability, while those with information gain far below the upper bound may suffer from underfitting. The selection process prioritizes decision trees with high information gain and simple structure, seeking the optimal balance between accuracy and complexity.

[0070] The selection of the optimal decision tree is based on a comprehensive evaluation index that combines the goal of minimizing information content with structural constraints. A comprehensive score is calculated for all candidate decision trees, including multiple dimensions such as classification accuracy, structural complexity, temporal consistency, and medical rationality. The decision tree with the highest score is selected as the optimal decision tree, as it possesses the simplest structure and branching logic that best aligns with the disease's evolution while ensuring diagnostic accuracy.

[0071] Boundary optimization is performed on the branches of the optimal decision tree, aiming to refine the decision boundaries to improve classification accuracy. The optimization process adjusts the segmentation threshold of each internal node, reducing the classification error rate through fine-tuning. For segmentation nodes with continuous features, the system searches for the optimal segmentation point near the original threshold, maximizing the purity of the segmented child nodes. Boundary optimization employs a local search algorithm to optimize node segmentation conditions while maintaining the tree structure, generating the final treatment decision tree that includes the disease evolution path.

[0072] In this embodiment, by introducing temporal consistency constraints to weight the information of non-leaf nodes, the process of retaining or pruning nodes is made more consistent with the actual trajectory of disease progression. An adaptive threshold curve is constructed by combining the initial tree depth and sample distribution, which enables dynamic setting of pruning criteria for different nodes based on their importance, avoiding information loss while significantly reducing redundant nodes. Furthermore, by fusing the number of leaf nodes, path complexity, and tree depth through structural constraints, dynamic programming search is performed to select the tree model with the least information content and optimal structure from multiple candidate schemes, thereby enhancing the model's discriminative power and interpretability. Finally, boundary optimization is performed while retaining key nodes in disease evolution, so that the generated treatment decision tree has both a concise structure and can accurately reflect the disease evolution path, demonstrating higher practical value and promotion potential in assisting clinical treatment plan formulation and individualized diagnosis and treatment.

[0073] Figure 3 This is a simulation diagram illustrating the similarity of drug-sign association maps and the importance of sign indicators. The diagram clearly distinguishes eight signs (blood pressure, heart rate, body temperature, blood oxygen, blood glucose, respiratory rate, liver function, and kidney function) using scatter plots of different shapes and colors. The cluster centers for each indicator are larger, showing their typical distribution locations. The scatter plot distribution reveals that blood oxygen (0.78, 92), liver function (0.75, 88), and blood pressure (0.73, 85) exhibit both high similarity and high importance, indicating their significant value in predicting disease progression pathways.

[0074] This simulation analysis constructs a strip penalty function to weight and correct the initial distance values, calculating the corrected distance values ​​at multiple time scales, thereby accurately quantifying the importance of vital signs. The charts clearly illustrate the vital sign indicator selection mechanism based on information relevance, and the differentiated contributions of indicators from different regions to the construction of the treatment decision tree.

[0075] In one optional implementation, the patient's disease progression trajectory is predicted based on a treatment decision tree, and the optimal referral timing is identified based on the treatment effectiveness coefficient and the specialty strengths of candidate medical institutions. The target medical institutions include: The pre-acquired set of vital signs indicators is decomposed temporally and features are extracted. The results are then input into the treatment decision tree to calculate the state transition probability and transition time of each node. A multidimensional state space is constructed based on the state transition probability and transition time. In the multidimensional state space, Monte Carlo sampling is performed on the state transition sequence in combination with the treatment effectiveness coefficient to generate a weighted set of disease progression trajectories; Historical treatment cases of candidate medical institutions are extracted from the collaborative database of medical institutions. These historical treatment cases are clustered according to disease type and treatment stage. The treatment success rate and disease duration of each category are calculated to generate a specialty matrix. The disease progression trajectory set is mapped to a specialty matrix, the trajectory-specialty matching degree is calculated, and the candidate medical institutions are dynamically sorted based on the matching degree. At the same time, a referral trigger threshold is set according to the treatment effectiveness coefficient. When the treatment effectiveness coefficient is lower than the referral trigger threshold and there are candidate medical institutions with a matching degree higher than the preset matching threshold, the current moment is determined as the best time for referral, and the medical institution with the highest matching degree is selected as the target medical institution.

[0076] In one specific embodiment, the medical system first acquires a set of vital signs indicators of the patient, including but not limited to key physiological parameters such as blood pressure, heart rate, blood glucose, body temperature, and blood oxygen saturation. The system performs time-series decomposition on these indicators, dividing the continuously monitored data into different time scales (hourly, daily, weekly), and extracts statistical features (mean, standard deviation, fluctuation trend) and waveform features (peak value, trough value, periodic changes). For example, for a patient with type 2 diabetes, the system can extract their blood glucose fluctuation pattern over 90 days, including a morning average blood glucose of 8.2 mmol / L with a standard deviation of 1.3 mmol / L, an afternoon peak average of 10.5 mmol / L, and a nighttime minimum of 6.1 mmol / L.

[0077] The extracted features are input into a pre-built treatment decision tree. This decision tree consists of multiple nodes, each representing a disease state. Taking the aforementioned diabetic patient as an example, the decision tree includes nodes such as "good blood sugar control," "significant blood sugar fluctuations," and "early diabetic nephropathy." The system calculates the state transition probabilities between each node; for example, the transition probability from "significant blood sugar fluctuations" to "early diabetic nephropathy" is 0.23, with an estimated transition time of 18 months. Based on these calculations, the system constructs a multi-dimensional state space, where each dimension represents a direction of disease progression.

[0078] In the constructed multidimensional state space, a treatment effectiveness coefficient is introduced to evaluate the effect of the current treatment plan. The treatment effectiveness coefficient is determined based on the patient's response to treatment, ranging from 0 to 1, where 0 represents complete ineffectiveness and 1 represents complete effectiveness. For example, the treatment effectiveness coefficient for this diabetic patient's oral hypoglycemic medication is 0.65, indicating that the treatment plan has some effect on blood sugar control but has not reached the ideal state. The system uses the Monte Carlo method to generate 10,000 possible disease progression trajectories based on state transition probabilities and the treatment effectiveness coefficient. Each trajectory includes a state sequence and its occurrence time points, and is assigned weights according to the treatment effectiveness coefficient. Trajectories with higher weights represent disease progression paths more likely to occur under the current treatment effect.

[0079] To evaluate the expertise of candidate medical institutions, a collaborative database of medical institutions was accessed, and historical treatment cases from each institution were extracted. These cases were clustered by disease type (e.g., cardiovascular disease, endocrine disease, neurological disease) and treatment stage (early intervention, disease control, complication management). For each category, treatment success rate and average duration of illness were calculated. Treatment success rate was defined as the proportion of cases achieving the expected clinical goal, and duration of illness was the average number of days from admission to achieving the treatment goal. For example, Medical Institution A had a success rate of 87% and an average duration of illness of 45 days in early intervention for diabetic nephropathy, while Medical Institution B had a success rate of 92% and an average duration of illness of 38 days in the same area. This data was organized into a specialty matrix, where rows represent disease type and treatment stage, columns represent different medical institutions, and cell values ​​are the overall score.

[0080] The generated disease progression trajectories are mapped to a specialty matrix, and the trajectory-specialty matching degree is calculated. The matching degree calculation considers three factors: the treatment success rate of the medical institution in predicting the disease state, the degree of matching between the medical institution's average disease duration and the patient's expected disease duration, and the similarity between the medical institution's historical cases and the current patient. For each predicted trajectory, a matching degree score is calculated with each candidate medical institution, ranging from 0 to 100. For example, for the possible development state of "early diabetic nephropathy," medical institution A has a matching degree of 78 points, while medical institution B has a matching degree of 85 points.

[0081] The referral trigger threshold is set based on the current treatment effectiveness coefficient. The lower the treatment effectiveness coefficient, the higher the referral trigger threshold. For example, when the treatment effectiveness coefficient is 0.65, the referral trigger threshold is set to 80. The system continuously monitors the patient's condition and updates the treatment effectiveness coefficient. When the coefficient drops to 0.52 (below the safe threshold of 0.6) and there is a matching institution B with a matching score of 85 (above the preset matching threshold of 80), the system determines that this is the optimal time for referral and identifies institution B as the target referral institution. The system generates a referral recommendation report, which includes an analysis of the patient's current condition, a prediction of disease progression, a description of the recommended medical institution and its specialties, and an analysis of the rationality of the referral timing.

[0082] In this embodiment, high-precision prediction of patient disease progression trends can be achieved, and intelligent assessment of treatment effectiveness and medical institution capabilities can be combined to dynamically identify the optimal timing and destination for referral. By integrating temporal sign analysis, state transition modeling, and Monte Carlo sampling, the individualized disease progression path of each patient can be fully reconstructed, improving the reliability of predictions. The introduction of a specialty matrix and its matching analysis with disease trajectories enables precise alignment of medical resources with disease evolution. By setting referral thresholds using treatment effectiveness coefficients, real-time perception of declining clinical efficacy is achieved, ensuring timely and evidence-based referral decisions. The overall solution strengthens intelligent collaboration in cross-institutional referral processes, significantly improving resource utilization efficiency and patient prognosis.

[0083] A second aspect of this invention provides a cross-institutional referral intelligent routing system based on medical intelligent agent middleware, the system comprising: The first unit is used to receive referral application information through the medical intelligent agent middleware. The referral application information includes patient medical record information and referral department information, wherein the patient medical record information includes vital sign data and medication records. The second unit is used to establish a monitoring window for vital signs data, to time-series map the rate of change of vital signs indicators with the time points and doses in the medication records, to identify the response delay time and peak level of vital signs indicators to different drug doses, to construct a drug response sensitivity curve, and to calculate the treatment effectiveness coefficient based on the drug response sensitivity curve. The third unit is used to obtain the diagnosis and treatment data of candidate medical institutions from the collaborative database of medical institutions, segment the drug response sensitivity curve according to the fluctuation cycle, extract the features of each segment of the curve to obtain the time-series feature vector, construct the drug-sign association map, use the dynamic time warping algorithm to calculate the similarity between the drug-sign association map and the historical association map in the diagnosis and treatment data, determine the set of sign indicators based on the similarity matching results, and generate a treatment decision tree containing the disease evolution path; The fourth unit is used to predict the patient's disease progression trajectory based on the treatment decision tree, identify the best referral time based on the treatment effectiveness coefficient and the specialty of candidate medical institutions, determine the target medical institution, and complete the data exchange by establishing a data transmission channel through the medical intelligent agent middleware.

[0084] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0085] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0086] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0087] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cross-institutional referral intelligent routing method based on medical intelligent agent middleware, characterized in that, include: The referral application information is received through the medical intelligent agent middleware. The referral application information includes patient medical record information and referral department information, wherein the patient medical record information includes vital sign data and medication records. Establish a monitoring window for vital signs data, map the rate of change of vital signs indicators to the time points and dosages in the medication records, identify the response delay time and peak level of vital signs indicators to different drug doses, construct a drug response sensitivity curve, and calculate the treatment effectiveness coefficient based on the drug response sensitivity curve; The system obtains diagnosis and treatment data of candidate medical institutions from the collaborative database of medical institutions, segments the drug response sensitivity curve according to the fluctuation cycle, extracts features from each segment of the curve to obtain time-series feature vectors, constructs a drug-sign association map, uses the dynamic time warping algorithm to calculate the similarity between the drug-sign association map and the historical association map in the diagnosis and treatment data, determines the set of sign indicators based on the similarity matching results, and generates a treatment decision tree containing the disease evolution path. Based on the treatment decision tree, the patient's condition trajectory is predicted. Based on the treatment effectiveness coefficient and the specialty of candidate medical institutions, the best time for referral is identified, the target medical institution is determined, and the data exchange is completed by establishing a data transmission channel through the medical intelligent agent middleware.

2. The method according to claim 1, characterized in that, A monitoring window for vital signs data is established, and the rate of change of vital signs indicators is time-series mapped with the dosing time points and dosages in the medication records. The response delay time and peak level of vital signs indicators to different drug doses are identified, and drug response sensitivity curves are constructed, including: Wavelet denoising was performed on the vital signs data to obtain smoothed vital signs data. The initial monitoring window was divided according to the interval between adjacent dosing time points in the medication record. The window size was dynamically adjusted based on the autocorrelation characteristics of the vital signs data to obtain the windowed vital signs sequence. The windowed vital sign sequence is calculated by time difference to obtain the vital sign change rate sequence. The dosage in the medication record is normalized to obtain the dosage sequence. The difference in probability distribution between the vital sign change rate sequence and the dosage sequence is calculated as the transmission loss. The time-series mapping relationship is obtained by minimizing the transmission loss. In the time-frequency decomposition of the time-series mapping relationship, the continuous trajectory with the highest energy density is extracted as the feature ridge. The starting point of the feature ridge is marked as the response delay time, and the maximum value point of the ridge is marked as the peak response level. A feature point set is constructed based on the response delay time and peak response level. The kernel function bandwidth is dynamically adjusted based on the local distribution pattern of the feature point set. The kernel function is then used to interpolate and fit the feature point set to construct a drug response sensitivity curve.

3. The method according to claim 1, characterized in that, The calculation of the treatment effectiveness coefficient based on the drug response sensitivity curve includes: The drug response sensitivity curve is segmented according to the curvature change point to obtain the characteristic segment reflecting the drug action process. The mapping relationship between the change trend of vital signs and the drug dosage in each characteristic segment is calculated to generate dose response characteristics. A compensation coefficient is constructed, which varies with the sensitivity of vital signs to drug dosage. The compensation coefficient is multiplied by the dose response characteristics to generate a compensated response value. Based on the compensated response values, an impulse response sequence is constructed, which reflects the response pattern of vital signs to a single dose. The superposition effect of the impulse response sequence is analyzed to obtain drug accumulation characteristics. The drug accumulation characteristics are combined with the peak values ​​of the impulse response sequence to obtain the therapeutic efficacy coefficient.

4. The method according to claim 1, characterized in that, The drug response sensitivity curve is segmented according to its fluctuation period to extract features from each segment, resulting in a time-series feature vector. A drug-sign association map is then constructed, including: Wavelet transform is performed on the drug response sensitivity curve to obtain periodic features. Based on the periodic features, fluctuation time windows are identified, and the drug response sensitivity curve is divided into multiple subsequences according to the fluctuation time windows. Extract the temporal feature vector of the subsequence, and use principal component analysis to reduce the dimensionality of the temporal feature vector to obtain the dimensionality-reduced feature vector; An initial heterogeneous graph network is constructed based on the dimensionality-reduced feature vector, which includes drug nodes and vital sign nodes. The correlation value between the drug nodes and vital sign nodes is calculated using the Pearson correlation coefficient, and the correlation value is used as the initial weight between the drug nodes and vital sign nodes. The initial weights are dynamically adjusted using a graph attention mechanism to obtain dynamic weights. The delay response value corresponding to the dynamic weights is calculated by combining time coding. A dynamic heterogeneous graph network is constructed based on the dynamic weights and delay response value. A multi-head attention mechanism is used to extract node interaction features from the dynamic heterogeneous graph network. A drug-sign association map is constructed based on the node interaction features.

5. The method according to claim 1, characterized in that, The dynamic time warping algorithm is used to calculate the similarity between the drug-sign association map and the historical association map in the diagnosis and treatment data. Based on the similarity matching results, the set of sign indicators is determined, and a treatment decision tree containing the disease evolution path is generated, including: The drug-sign association map is constructed as a time series matrix. A time series matching window is set, and a dynamic time warping algorithm is executed within the time series matching window to calculate the initial distance value between the time series matrix and the historical association map in the diagnosis and treatment data. Construct a strip penalty function, and perform weighted correction on the initial distance value according to the strip penalty function to obtain the corrected distance value. Calculate the corrected distance values ​​at different time scales and perform weighted combination to obtain the map similarity value. Historical cases are clustered and grouped according to the map similarity value. The pattern of vital signs change is extracted from the clusters to obtain the vital sign feature sequence. The information correlation between features in the vital sign feature sequence is calculated. The importance of the vital sign feature sequence is quantitatively scored based on the information correlation. Features with importance quantitative scores higher than the preset score are selected to construct a set of vital sign indicators. Decision tree nodes are constructed based on the set of vital signs indicators. In the decision tree nodes, indicator judgment intervals, drug response probabilities, and time window parameters are set to generate an initial decision tree. The initial decision tree is pruned and optimized based on the principle of minimum information content. Structural constraint terms are introduced to limit the branch size of the initial decision tree to obtain the treatment decision tree of the disease evolution path.

6. The method according to claim 5, characterized in that, Based on the principle of minimum information content, the initial decision tree is pruned and optimized. Structural constraints are introduced to limit the branch size of the initial decision tree, resulting in a treatment decision tree for the disease evolution path, including: The classification results, node transition time, and subtree information are extracted from the non-leaf nodes of the initial decision tree to construct a node information vector containing temporal consistency constraints. The node information vector is then weighted and combined based on the temporal consistency constraints to generate the node information quantity. The information change value of the non-leaf node under the principle of minimum information is calculated, and an adaptive threshold curve is constructed based on the depth distribution and sample distribution of the initial decision tree. The information change value is projected onto the adaptive threshold curve for comparison. When the projection result is lower than the adaptive threshold curve, pruning is performed to obtain the pruned decision tree. Extract the number of leaf nodes, tree depth, and path complexity of the pruning decision tree, and combine them to construct structural constraints. Based on the structural constraints, perform dynamic programming search on the pruning decision tree to generate multiple candidate decision trees. Feature statistics are performed on the nodes of the candidate decision trees, and the upper bound of information gain is calculated in combination with the disease evolution pattern. The candidate decision trees are then screened according to the upper bound of information gain, and the decision tree with the smallest information gain is selected as the optimal decision tree. Boundary optimization is performed on the branches of the optimal decision tree to generate a treatment decision tree containing the disease evolution path.

7. The method according to claim 1, characterized in that, Based on the treatment decision tree to predict the patient's disease progression trajectory, and considering the treatment effectiveness coefficient and the specialty strengths of candidate medical institutions, the optimal referral time is identified, and target medical institutions are determined, including: The pre-acquired set of vital signs indicators is decomposed temporally and features are extracted. The results are then input into the treatment decision tree to calculate the state transition probability and transition time of each node. A multidimensional state space is constructed based on the state transition probability and transition time. In the multidimensional state space, Monte Carlo sampling is performed on the state transition sequence in combination with the treatment effectiveness coefficient to generate a weighted set of disease progression trajectories; Historical treatment cases of candidate medical institutions are extracted from the collaborative database of medical institutions. These historical treatment cases are clustered according to disease type and treatment stage. The treatment success rate and disease duration of each category are calculated to generate a specialty matrix. The disease progression trajectory set is mapped to a specialty matrix, the trajectory-specialty matching degree is calculated, and the candidate medical institutions are dynamically sorted based on the matching degree. At the same time, a referral trigger threshold is set according to the treatment effectiveness coefficient. When the treatment effectiveness coefficient is lower than the referral trigger threshold and there are candidate medical institutions with a matching degree higher than the preset matching threshold, the current moment is determined as the best time for referral, and the medical institution with the highest matching degree is selected as the target medical institution.

8. A cross-institutional referral intelligent routing system based on medical intelligent agent middleware, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to receive referral application information through the medical intelligent agent middleware. The referral application information includes patient medical record information and referral department information, wherein the patient medical record information includes vital sign data and medication records. The second unit is used to establish a monitoring window for vital signs data, to time-series map the rate of change of vital signs indicators with the time points and doses in the medication records, to identify the response delay time and peak level of vital signs indicators to different drug doses, to construct a drug response sensitivity curve, and to calculate the treatment effectiveness coefficient based on the drug response sensitivity curve. The third unit is used to obtain the diagnosis and treatment data of candidate medical institutions from the collaborative database of medical institutions, segment the drug response sensitivity curve according to the fluctuation cycle, extract the features of each segment of the curve to obtain the time-series feature vector, construct the drug-sign association map, use the dynamic time warping algorithm to calculate the similarity between the drug-sign association map and the historical association map in the diagnosis and treatment data, determine the set of sign indicators based on the similarity matching results, and generate a treatment decision tree containing the disease evolution path; The fourth unit is used to predict the patient's disease progression trajectory based on the treatment decision tree, identify the best referral time based on the treatment effectiveness coefficient and the specialty of candidate medical institutions, determine the target medical institution, and complete the data exchange by establishing a data transmission channel through the medical intelligent agent middleware.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images