Deep learning-based expiration experiment triage and queuing optimization method and system

Abstract

The invention provides a deep learning-based expiration experiment triage and queuing optimization method and system, and relates to the technical field of medical management, and the method comprises the steps: collecting multi-dimensional physiological data, carrying out the time-frequency domain decomposition to extract features, generating a health index through a mixed expert model, and constructing a multi-objective optimization function in combination with the time cost, the diagnosis and treatment matching degree and the resource utilization rate, and generating an optimal allocation scheme through adaptive search. According to the method, the triage accuracy can be improved, the waiting time of the patient is shortened, the medical resource configuration is optimized, and the treatment efficiency and satisfaction are improved.

Description

Technical Field

[0001] The present invention relates to the field of medical management technology, and in particular to a method and system for breathalyzer triage and queuing optimization based on deep learning. Background Art

[0002] With the increasing demand for medical care, medical institutions are facing problems such as long patient wait times and inefficient triage. Traditional medical queuing systems, often based on a first-come, first-served basis or simple triage criteria, are unable to effectively identify a patient's actual health status and urgency. This results in low triage accuracy, making it difficult to optimize overall medical outcomes, leading to irrational allocation of medical resources and low patient satisfaction. In recent years, breath analysis technology has gained increasing attention as a non-invasive and convenient method for measuring physiological status. It can assess health status by detecting specific components in a person's exhaled breath. Furthermore, the increasing application of deep learning technology in the medical field offers new possibilities for breath data analysis and patient triage. Summary of the Invention

[0003] The embodiments of the present invention provide a deep learning-based breath test triage and queuing optimization method and system, which can solve the problems in the prior art.

[0004] A first aspect of an embodiment of the present invention provides a deep learning-based breath test triage and queuing optimization method, comprising: Collect multi-dimensional physiological data of the target subject from the breath experiment site; Performing time-frequency decomposition on the multidimensional physiological data to extract a time-domain feature sequence and a frequency feature spectrum; extracting non-stationary feature vectors from the time-domain feature sequence, performing cluster analysis on the frequency feature spectrum to form a feature combination; inputting the non-stationary feature and the feature combination into a pre-built hybrid expert model, fusing the prediction results of multiple expert sub-networks to generate a health index of the target object; Obtaining the total number of target subjects and the number of clinics, triaging the target subjects according to the health index, and obtaining an initial allocation plan; A multi-objective function is constructed to calculate the time cost based on the target patient's waiting time, and the diagnosis and treatment matching degree and resource utilization rate are calculated based on the doctor's expertise and consultation rate. Based on the initial allocation plan, multiple adjustment plans that meet the multi-objective constraints are generated and their state transition probabilities are calculated. Based on the multi-objective function, a roulette wheel method is used to update the search direction, and the allocation plan is adjusted by adaptive step size. It is repeated until the multi-objective function value is less than the optimization threshold, and the optimal allocation plan is output. A triage number and an estimated consultation time are generated according to the optimal allocation plan, and the triage result is sent to the terminal device of the target object.

[0005] Performing time-frequency domain decomposition on the multidimensional physiological data to extract a time-domain feature sequence and a frequency feature spectrum; extracting non-stationary feature vectors in the time-domain feature sequence, performing cluster analysis on the frequency feature spectrum, and forming a feature combination, including: Performing empirical mode decomposition on the multidimensional physiological data to obtain a plurality of intrinsic mode functions, and performing time series reconstruction on the plurality of intrinsic mode functions to obtain a time domain feature sequence; extracting instantaneous frequency features from the multidimensional physiological data through Hilbert transform to obtain a frequency feature spectrum; Based on a preset time window, a sliding analysis is performed on the time domain feature sequence to calculate the local energy density within the time window; a convolution operation is performed on the time domain feature sequence with wavelet basis functions of different scales to obtain detail coefficients of multiple scales, a weighted sum is taken of the squares of the detail coefficients, and the sum is multiplied by the local energy density to obtain a non-stationary feature; An initial neighborhood radius is set, and the initial neighborhood radius is linearly reduced as the number of iterations increases. The frequency feature spectrum is clustered based on the current neighborhood radius and the preset learning rate to obtain multiple feature combinations and their cluster centers. When the distances between the cluster centers are all greater than the clustering threshold, the current feature combination is taken as the optimal feature combination.

[0006] The non-stationary features and the feature combination are input into a pre-built hybrid expert model, and the prediction results of multiple expert sub-networks are integrated to generate a health index of the target object, including: For each expert sub-network, the non-stationary features and the feature combination are subjected to multi-layer convolution and pooling operations to extract hierarchical features and generate preliminary prediction results. The preliminary prediction results are error-compensated using a residual learning mechanism to obtain residual correction terms. The preliminary prediction results and the residual correction terms are weightedly fused to obtain corrected prediction values ​​for each expert sub-network. The corrected prediction values ​​of each expert sub-network are input into the gated recurrent unit according to a preset time step, and a time series prediction result is generated by controlling the forgetting and updating of historical information; the prediction error between the time series prediction result and the true value is calculated, and a fusion weight is assigned to each expert sub-network through iterative optimization. When the error change in multiple consecutive iterations is less than the error threshold, the iteration is stopped to obtain the optimal fusion weight; The optimal fusion weight is weightedly combined with the corrected prediction values ​​of each expert sub-network to obtain the health index of the target object.

[0007] Construct a multi-objective function, calculate the time cost based on the target patient's waiting time, and calculate the diagnosis and treatment matching degree and resource utilization rate based on the doctor's expertise and consultation rate, including: Obtain the number of people queuing in the department, the number of doctors on duty, and the service capacity per unit time. Divide the number of people queuing in the department by the product of the number of doctors on duty and the service capacity per unit time to obtain the basic waiting time. Substitute the basic waiting time into the exponential decay function to obtain the time cost. Calculate the semantic similarity between the target subject's self-reported symptom keywords and the doctor's professional expertise keywords; obtain the average treatment time of the doctor's historical treatment data, and use the ratio of the average treatment time to the standard treatment time as the experience matching degree; and obtain the treatment matching degree by weighted summation of the semantic similarity and the experience matching degree; The number of patients the doctor has seen is divided by the maximum number of patients the doctor can see to obtain the consultation rate, and the average value of the consultation rate is used as the resource utilization rate; the time cost, the diagnosis and treatment matching degree and the resource utilization rate are weighted and summed to obtain a multi-objective function.

[0008] Based on the initial allocation plan, multiple adjustment plans that meet the multi-objective constraints are generated, and their state transition probabilities are calculated; combined with the multi-objective function, the search direction is updated using a roulette wheel method, the allocation plan is adjusted by an adaptive step size, and it is iterated repeatedly until the multi-objective function value is less than the optimization threshold, and the optimal allocation plan is output, including: Obtaining an initial allocation plan for the target objects, generating multiple adjustment paths for each target object while satisfying multi-objective constraints, and calculating the multi-objective function value on each adjustment path; calculating the product of the initial pheromone concentration of each adjustment path and the inverse of the multi-objective function value to obtain a state transition probability; Calculate the multi-objective function difference between the current adjustment path and the initial allocation plan, and dynamically update the search direction on the current adjustment path using a roulette wheel method in combination with the state transition probability; determine the product of the maximum adjustment step length, the complementary power term of the iteration progress, and the multi-objective function difference as the adaptive step length, and adjust the allocation plan of the target object based on the updated search direction and the adaptive step length; Repeat the above iterative process until the multi-objective function values ​​of all target objects under the current allocation scheme are less than the optimization threshold, and then output the current allocation scheme as the optimal allocation scheme.

[0009] Calculate the difference between the multi-objective function of the current adjustment path and the initial allocation plan, combine the state transition probability, and use the roulette method to dynamically update the search direction on the current adjustment path, including: Obtain the current adjustment path and its state transition probability, calculate the complement of the ratio of the current number of iterations to the maximum number of iterations, multiply the power product of the complement and the state transition probability by the probability threshold, and obtain the adaptive selection probability; Calculating a multi-objective function difference between the current adjustment path and the initial allocation solution, and when the multi-objective function difference is less than the adaptive selection probability, selecting a search direction with the maximum local pheromone concentration on the adjustment path; When the difference of the multi-objective function is greater than the adaptive selection probability, the complementary value of the pheromone volatility coefficient of the current adjustment path is multiplied by the local pheromone concentration of each search direction on the current adjustment path, and normalized to obtain the current selection probability of each search direction; according to the numbering order of each search direction, the current selection probabilities are accumulated one by one to obtain the cumulative selection probability of the search direction; the difference between the random number and the cumulative selection probability is calculated, and the search direction with the smallest difference is selected as the final search direction.

[0010] A second aspect of an embodiment of the present invention provides a deep learning-based breath test triage and queuing optimization system, comprising: The first unit is used to collect multi-dimensional physiological data of the target object from the breath experiment site; The second unit is configured to perform time-frequency domain decomposition on the multidimensional physiological data, extract a time-domain feature sequence and a frequency feature spectrum; extract non-stationary feature vectors from the time-domain feature sequence, perform cluster analysis on the frequency feature spectrum, and form a feature combination; input the non-stationary feature and the feature combination into a pre-built hybrid expert model, and fuse the prediction results of multiple expert sub-networks to generate a health index of the target object; The third unit is used to obtain the total number of target subjects and the number of clinics, triage the target subjects according to the health index, and obtain an initial allocation plan; The fourth unit is used to construct a multi-objective function, calculate the time cost based on the waiting time of the target object, and calculate the diagnosis and treatment matching degree and resource utilization rate based on the doctor's expertise and consultation rate; based on the initial allocation plan, generate multiple adjustment plans that meet the multi-objective constraints and calculate their state transition probabilities; based on the multi-objective function, use a roulette wheel method to update the search direction, adjust the allocation plan through adaptive step size, repeat the iteration until the multi-objective function value is less than the optimization threshold, and output the optimal allocation plan; The fifth unit is used to generate a triage number and an estimated consultation time according to the optimal allocation plan, and send the triage result to the terminal device of the target object.

[0011] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0012] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0013] The beneficial effects of this application are as follows: The present invention collects multidimensional physiological data from breath test sites, performs time-frequency domain decomposition to extract features, and combines a hybrid expert model to generate a health index, thereby achieving an accurate assessment of the patient's health status, improving the accuracy and scientific nature of triage, and avoiding the problems of strong subjectivity and low precision of traditional triage methods.

[0014] The present invention constructs a multi-objective function, taking into account the three dimensions of time cost, diagnosis and treatment matching, and resource utilization. By adjusting the allocation plan through adaptive step size, it optimizes the allocation of medical resources while meeting the patients' medical needs, significantly reducing the average waiting time of patients, and improving medical efficiency and patient satisfaction.

[0015] This invention organically combines deep learning technology with queuing optimization algorithms to form a complete closed-loop system from data collection, feature extraction to triage and queuing. It can not only perform accurate triage based on the patient's actual health status, but also generate personalized consultation numbers and estimated consultation times, thereby improving the intelligence level of medical services and the patient's medical experience, and has good practicality and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flowchart of a deep learning-based breath test triage and queuing optimization method according to an embodiment of the present invention; Figure 2 This is the system architecture diagram of the adaptive allocation solution optimization method based on multi-objective constraints. DETAILED DESCRIPTION

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0018] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0019] Figure 1 FIG. 1 is a flow chart of a method for triaging and queuing optimization based on deep learning of a breath test according to an embodiment of the present invention. Figure 1 As shown, the method includes: Collect multi-dimensional physiological data of the target subject from the breath experiment site; Performing time-frequency decomposition on the multidimensional physiological data to extract a time-domain feature sequence and a frequency feature spectrum; extracting non-stationary feature vectors from the time-domain feature sequence, performing cluster analysis on the frequency feature spectrum to form a feature combination; inputting the non-stationary feature and the feature combination into a pre-built hybrid expert model, fusing the prediction results of multiple expert sub-networks to generate a health index of the target object; Obtaining the total number of target subjects and the number of clinics, triaging the target subjects according to the health index, and obtaining an initial allocation plan; A multi-objective function is constructed to calculate the time cost based on the target patient's waiting time, and the diagnosis and treatment matching degree and resource utilization rate are calculated based on the doctor's expertise and consultation rate. Based on the initial allocation plan, multiple adjustment plans that meet the multi-objective constraints are generated and their state transition probabilities are calculated. Based on the multi-objective function, a roulette wheel method is used to update the search direction, and the allocation plan is adjusted by adaptive step size. It is repeated until the multi-objective function value is less than the optimization threshold, and the optimal allocation plan is output. A triage number and an estimated consultation time are generated according to the optimal allocation plan, and the triage result is sent to the terminal device of the target object.

[0020] In an optional embodiment, the multidimensional physiological data is decomposed in the time-frequency domain to extract a time-domain feature sequence and a frequency feature spectrum; non-stationary feature vectors in the time-domain feature sequence are extracted, and cluster analysis is performed on the frequency feature spectrum to form a feature combination, including: Performing empirical mode decomposition on the multidimensional physiological data to obtain a plurality of intrinsic mode functions, and performing time series reconstruction on the plurality of intrinsic mode functions to obtain a time domain feature sequence; extracting instantaneous frequency features from the multidimensional physiological data through Hilbert transform to obtain a frequency feature spectrum; Based on a preset time window, a sliding analysis is performed on the time domain feature sequence to calculate the local energy density within the time window; a convolution operation is performed on the time domain feature sequence with wavelet basis functions of different scales to obtain detail coefficients of multiple scales, a weighted sum is taken of the squares of the detail coefficients, and the sum is multiplied by the local energy density to obtain a non-stationary feature; An initial neighborhood radius is set, and the initial neighborhood radius is linearly reduced as the number of iterations increases. The frequency feature spectrum is clustered based on the current neighborhood radius and the preset learning rate to obtain multiple feature combinations and their cluster centers. When the distances between the cluster centers are all greater than the clustering threshold, the current feature combination is taken as the optimal feature combination.

[0021] The present invention provides a multidimensional physiological data processing method that effectively processes physiological data through time-frequency domain decomposition and feature extraction analysis. In practice, the method acquires physiological data from multiple dimensions, including electrocardiogram (ECG), respiration, and blood pressure, which often exhibit complex nonlinear and nonstationary characteristics.

[0022] Empirical mode decomposition (EMD) is performed on the acquired multidimensional physiological data, decomposing complex physiological signals into multiple intrinsic mode functions (IMFs). For example, for ECG signals acquired at a frequency of 1000 Hz, EMD yields 8-12 IMFs, each representing a signal component at a different frequency scale. These IMFs are then reconstructed in time series. Specifically, the IMFs containing the primary physiological information (typically the third to seventh functions) are selected and reconstructed through a weighted superposition process to create a time-domain feature sequence. Weight coefficients are determined based on the energy contribution of each IMF. Typically, IMFs with higher energy contributions are given greater weights. For example, the weight of the fourth mode function might be set to 0.35, the fifth to 0.25, and the others adjusted accordingly.

[0023] Multidimensional physiological data is processed through a Hilbert transform to extract instantaneous frequency features. Specifically, a Hilbert transform is performed on each dimension of physiological data to obtain an analytical representation of the signal, from which the instantaneous amplitude and phase information are extracted. The instantaneous frequency is calculated using the time derivative of the instantaneous phase to form a frequency signature spectrum. For example, after processing the heart rate variability signal, a frequency signature spectrum in the range of 0.04Hz-0.4Hz can be obtained, reflecting the activity of the sympathetic and parasympathetic nervous systems.

[0024] A sliding analysis is performed on the time-domain feature sequence based on a preset time window to calculate the local energy density. In practice, the window length is set to 2 seconds, with a sliding step of 0.5 seconds. For each window position, the sum of the squares of the sampling points within the window is calculated and divided by the window length to obtain the local energy density value for that time period. For example, at a heart rate of 75 beats / minute, a 2-second window contains approximately 2.5 cardiac cycles, which can fully capture the local variation characteristics of the ECG signal.

[0025] In order to accurately extract the non-stationary features in the time domain feature sequence, it is necessary to select an appropriate wavelet basis function. The selection of the wavelet basis function is crucial for capturing specific features in the signal. In this embodiment, a variety of wavelet basis functions can be used, such as Haar wavelet, Daubechies wavelet, Coiflet wavelet, or Symlet wavelet. Based on the characteristics of the exhalation signal, the Daubechies-4 (db4) wavelet can be selected as the basic analysis tool because it exhibits good time-frequency localization characteristics when analyzing respiratory-related signals.

[0026] The selected wavelet basis function is expanded and translated at different scales to form a set of wavelet basis functions of different scales. In practical applications, five different scale levels can be selected, corresponding to the analysis of different frequency components in the signal. For example, for an exhalation signal sampled at 100 Hz, these five scales can correspond to frequency components of approximately 50 Hz, 25 Hz, 12.5 Hz, 6.25 Hz, and 3.125 Hz. At each scale level, the time-domain feature sequence is convolved with the wavelet basis function of the corresponding scale. Convolution essentially calculates the correlation between the signal and the wavelet function, thereby detecting the signal's varying characteristics at different times and frequencies. For a time-domain signal of length 1000 and a db4 wavelet basis function of length 4, the convolution operation produces approximately 997 output points, forming the detail coefficient sequence at that scale. By performing convolution operations on five different scales, five sets of detail coefficient sequences are obtained. These detail coefficients reflect the varying characteristics of the original signal across different frequency ranges. Each set of detail coefficients is approximately the same length as the original signal, with a total of five sets of approximately 997 data points. To quantify the energy distribution of the signal at each scale, the detail coefficient sequence at each scale is squared. This squaring operation converts the detail coefficients into an energy representation, highlighting important changes in the signal. This step results in five energy sequences, each with approximately 997 data points.

[0027] Based on the importance of different scales, weight coefficients are assigned to the energy sequences at each scale. In breath analysis, mid-frequency components (such as the second and third scales) typically contain more physiological information and can therefore be given higher weights. For example, the weights for the five scales can be set to [0.1, 0.25, 0.3, 0.25, 0.1], indicating that the intermediate scales are more important. The weighted energy sequences at each scale are summed to obtain a multi-scale energy representation. This step combines the five weighted energy sequences into a comprehensive energy sequence of approximately 997 data points. The local energy density of the time-domain feature sequence is calculated. The local energy density reflects the energy concentration of the original signal within a short time window and can be calculated using a sliding window method. Specifically, a sliding window of length 20 (corresponding to 0.2 seconds at a 100Hz sampling rate) is used to calculate the average of the sum of squares of the signal within the window to obtain the local energy density sequence. For an original signal with 1000 data points, this step yields approximately 981 local energy density values. Multiplying the multiscale energy representation with the local energy density yields non-stationary features. This step integrates the signal's time-frequency characteristics and local energy distribution, effectively capturing the signal's non-stationary components. The multiplication operation requires resizing the two sequences to the same length, typically the minimum of the two, which in this case is approximately 981 data points.

[0028] To validate the effectiveness of the extracted non-stationary features, a representative set of exhaled breath samples was analyzed. This set of samples included exhaled breath data from 10 healthy subjects and 10 patients with respiratory diseases. The healthy subjects' exhaled flow rate exhibited a relatively stable bell-shaped curve, with the extracted non-stationary eigenvalues ​​ranging from 0.05 to 0.15. In contrast, the patients' exhaled flow curves exhibited significant irregular fluctuations, with the extracted non-stationary eigenvalues ​​ranging from 0.2 to 0.5, significantly higher than those in the healthy subjects. This demonstrates that the extracted non-stationary features can effectively distinguish exhaled patterns in different physiological states.

[0029] Non-stationary features are further processed to extract statistical indicators (mean, standard deviation, kurtosis, skewness, etc.) or time-domain features (peak location, duration, etc.) to form higher-dimensional feature vectors. For example, for each subject's non-stationary feature sequence, five statistical features were calculated: mean (healthy group: 0.1±0.03, patient group: 0.35±0.08), standard deviation (healthy group: 0.02±0.005, patient group: 0.08±0.02), maximum (healthy group: 0.15±0.04, patient group: 0.5±0.1), minimum (healthy group: 0.05±0.01, patient group: 0.2±0.05), and kurtosis (healthy group: 2.8±0.3, patient group: 4.5±0.7). These features form a five-dimensional feature vector for subsequent health status assessment.

[0030] A self-organizing map neural network method was used to cluster the frequency feature spectrum. The initial neighborhood radius was set to 4 and linearly decreased with iterations. The specific attenuation formula was: current neighborhood radius = initial neighborhood radius × (1 - current number of iterations / total number of iterations). The learning rate was initially set to 0.1 and also decayed with the number of iterations. During implementation, the total number of iterations was set to 1000. When the 800th iteration was reached, the learning rate dropped to approximately 0.02, and the neighborhood radius was reduced to 0.8. The frequency feature spectrum was clustered based on the current neighborhood radius and learning rate, and the cluster centers were dynamically updated. One hundred groups of sample data were selected for clustering, and five cluster centers were initially randomly generated. After iterative adjustment, a stable feature cluster combination was eventually formed. The Euclidean distance between cluster centers was calculated, and the clustering threshold was set to 0.15. When the distance between all cluster centers was greater than this threshold, the current clustering result was determined to be the optimal feature combination.

[0031] In a real-world application, this method was used to process multidimensional physiological signals, including electrocardiogram (ECG), respiration, and blood pressure data, with a sampling frequency of 1000 Hz and a recording duration of 5 minutes. After time-frequency domain decomposition and feature extraction, a 12-dimensional non-stationary feature vector and three primary feature clusters were obtained. The resulting feature combination was able to effectively distinguish between different physiological states.

[0032] In an optional embodiment, the non-stationary features and the feature combination are input into a pre-built hybrid expert model, and the prediction results of multiple expert sub-networks are integrated to generate the health index of the target object, including: For each expert sub-network, the non-stationary features and the feature combination are subjected to multi-layer convolution and pooling operations to extract hierarchical features and generate preliminary prediction results. The preliminary prediction results are error-compensated using a residual learning mechanism to obtain residual correction terms. The preliminary prediction results and the residual correction terms are weightedly fused to obtain corrected prediction values ​​for each expert sub-network. The corrected prediction values ​​of each expert sub-network are input into the gated recurrent unit according to a preset time step, and a time series prediction result is generated by controlling the forgetting and updating of historical information; the prediction error between the time series prediction result and the true value is calculated, and a fusion weight is assigned to each expert sub-network through iterative optimization. When the error change in multiple consecutive iterations is less than the error threshold, the iteration is stopped to obtain the optimal fusion weight; The optimal fusion weight is weightedly combined with the corrected prediction values ​​of each expert sub-network to obtain the health index of the target object.

[0033] This embodiment describes a method for inputting non-stationary features and feature combinations into a pre-built hybrid expert model to generate a target object health index. This method achieves an accurate assessment of the target object's health status by fusing the prediction results of multiple expert sub-networks.

[0034] The collected raw data of the target object is subjected to feature extraction and preprocessing to obtain non-stationary features and feature combinations. These features are then input into a pre-built mixture of experts model. This mixture of experts model contains multiple parallel expert sub-networks, each of which focuses on capturing different patterns and features in the data.

[0035] For each expert sub-network, non-stationary features and feature combinations are input and processed through multiple layers of convolution and pooling operations. Specifically, the first convolutional layer uses 32 3×3 convolution kernels to extract features from the input data. The ReLU activation function is then applied to increase nonlinearity, and then a 2×2 max pooling layer is used to reduce the feature dimension. The second convolutional layer uses 64 3×3 convolution kernels to further extract high-level features, also applying the ReLU activation function and a 2×2 max pooling layer. This layered convolution and pooling operation effectively extracts local and global features of the input data and generates preliminary prediction results.

[0036] To improve prediction accuracy, a residual learning mechanism is used to compensate for errors in the initial prediction results. The residual learning mechanism is implemented through a bypass branch consisting of a two-layer fully connected network. The first layer contains 128 neurons, and the output dimension of the second layer is the same as the initial prediction result. This bypass branch learns the residual between the predicted value and the true value to generate a residual correction term. The preliminary prediction result and the residual correction term are weighted and fused with the weight coefficients set to 1 and 1.5 to obtain the corrected prediction value of each expert sub-network. Specifically, when the preliminary health index prediction of a target object is 80.5 and the residual correction term is 1.8, the corrected health index will be 83.2, which is closer to the true value. For a certain target object, the corrected prediction values ​​of the three expert sub-networks are 83.2, 81.9, and 82.5, respectively.

[0037] The corrected predictions from each expert sub-network are fed into a gated recurrent unit (GRU) consisting of 64 hidden units at a preset time step length (e.g., 10 time points). The GRU uses update and reset gates to control the forgetting and updating of historical information. The update gate determines how much previous state information is retained, while the reset gate determines how new input information is combined with the previous state. This mechanism enables the model to capture long-term dependencies in time series data and generate time series forecasts that take into account historical trends.

[0038] The prediction error between the time series forecast and the true value is calculated, using mean squared error as the evaluation metric. A fusion weight is assigned to each expert subnetwork through an iterative optimization process. During this iteration, the system uses the Adam optimization algorithm, with a learning rate of 0.001 and a batch size of 32. When the error change over five consecutive iterations is less than the preset error threshold of 0.0001, iteration terminates and the optimal fusion weight is obtained. For example, for the three expert subnetworks described above, the optimal fusion weights obtained after iterative optimization are 0.4, 0.25, and 0.35, respectively. The optimal fusion weight is weighted and combined with the corrected predictions of each expert subnetwork to obtain the health index of the target object. The specific calculation is: 0.4 × 83.2 + 0.25 × 81.9 + 0.35 × 82.5 = 82.66, which means the health index of the target object is 82.66.

[0039] By fusing the prediction results of multiple expert sub-networks and combining residual learning and gated recurrent units, this method can effectively handle the non-stationary characteristics and temporal dependencies of the data, providing more accurate health assessment for the target object.

[0040] In an optional embodiment, a multi-objective function is constructed to calculate the time cost based on the target patient's waiting time, and to calculate the diagnosis and treatment matching degree and resource utilization rate based on the doctor's expertise and consultation rate, including: Obtain the number of people queuing in the department, the number of doctors on duty, and the service capacity per unit time. Divide the number of people queuing in the department by the product of the number of doctors on duty and the service capacity per unit time to obtain the basic waiting time. Substitute the basic waiting time into the exponential decay function to obtain the time cost. Calculate the semantic similarity between the target subject's self-reported symptom keywords and the doctor's professional expertise keywords; obtain the average treatment time of the doctor's historical treatment data, and use the ratio of the average treatment time to the standard treatment time as the experience matching degree; and obtain the treatment matching degree by weighted summation of the semantic similarity and the experience matching degree; The number of patients the doctor has seen is divided by the maximum number of patients the doctor can see to obtain the consultation rate, and the average value of the consultation rate is used as the resource utilization rate; the time cost, the diagnosis and treatment matching degree and the resource utilization rate are weighted and summed to obtain a multi-objective function.

[0041] In this embodiment, a medical resource allocation system based on multi-objective queuing optimization is provided. A multi-objective function is constructed that comprehensively considers waiting time cost, diagnosis and treatment matching, and resource utilization to achieve reasonable allocation of medical resources.

[0042] Obtain the current queue number, number of doctors on duty, and each doctor's service capacity per unit time for each department. For the respiratory department, for example, the current queue number is 50, the number of doctors on duty is 5, and the average service capacity per doctor per unit time is 3 people / hour. The basic waiting time is calculated as the number of people in the queue divided by the product of the number of doctors on duty and the service capacity per unit time, that is, 50 ÷ (5 × 3) = 3.33 hours.

[0043] In order to more accurately reflect the nonlinear effect of waiting time on patient satisfaction, an exponential decay function is used to calculate the time cost. When the basic waiting time is t, the time cost can be expressed as 1-e (-αt) , where α is the attenuation coefficient, set to 0.5 based on the actual hospital situation. Taking the respiratory department mentioned above as an example, substituting the basic waiting time of 3.33 hours into the attenuation coefficient α = 0.5, the calculated time cost is 0.81, which represents the patient's dissatisfaction caused by waiting.

[0044] To calculate the degree of diagnosis and treatment matching, it is necessary to obtain the target subject's self-reported symptoms. The target subject's symptom description is collected through a digital questionnaire. The questionnaire is designed to include open-ended questions related to gastrointestinal symptoms, such as "Please describe your stomach discomfort symptoms", as well as structured questions such as "Do you have acid reflux?", "Do you have abdominal pain?", "How long do the symptoms last?", etc. The target subject fills in this information through the interactive terminal of the breath test site. For example, a target subject may describe: "In the past month, I have often experienced dull pain in the upper abdomen. The symptoms worsen after meals. I occasionally have acid reflux and heartburn. The pain at night causes poor sleep. I have taken antacids on my own but the effect is not obvious." The collected self-reported symptom text is preprocessed. Preprocessing includes text segmentation, stop word removal, and standardization. The segmentation process uses a segmentation algorithm optimized in the medical field to split continuous text into independent words or phrases. For example, the symptom description above can be segmented into terms such as "one month," "upper abdomen," "dull pain," "after meals," "worsened symptoms," "acid reflux," "heartburn," "nighttime," "pain," "poor sleep," "antacid medication," and "ineffective." The stop word removal step filters out common words like "的," "和," and "会," which contribute little to symptom recognition. Normalization unifies different expressions of the same symptom into standard medical terms, such as normalizing "heartburn" to "heartburn" and "dull pain in the upper abdomen" to "upper abdominal pain."

[0045] Symptom keywords potentially associated with Helicobacter pylori infection are extracted from the preprocessed text. A symptom dictionary constructed from a knowledge base of gastrointestinal diseases contains approximately 150 common gastrointestinal symptom words and phrases, with a particular focus on symptoms associated with H. pylori infection, such as "upper abdominal pain," "heartburn," "acid reflux," "bloating," "indigestion," "nausea," "vomiting," and "loss of appetite." These keywords are identified from the subject's self-reported descriptions. For the example above, the extracted keywords include "upper abdominal pain," "worsened after meals," "acid reflux," "heartburn," "night pain," and "ineffective antacid medication." Each extracted keyword is assigned a weight reflecting its importance in the diagnosis and treatment of H. pylori infection and related gastric diseases. Weighting is based on multiple factors: the symptom's relevance to H. pylori infection (symptoms with high relevance receive higher weights), the symptom's typicality (symptoms with high typicality receive higher weights), the symptom's duration (symptoms with long duration receive higher weights), and the symptom's frequency (symptoms with high frequency receive higher weights). For example, "upper abdominal pain" as a typical symptom of H. pylori infection may be assigned a weight of 0.85, "acid reflux" a weight of 0.75, "heartburn" a weight of 0.7, "night pain" a weight of 0.8, and "antacids are not effective" a weight of 0.65.

[0046] A database of physician expertise specifically focused on digestive system diseases was constructed, documenting each gastroenterologist's specialty, research interests, clinical experience, and specialized diseases, with a specific focus on Helicobacter pylori-related diseases. This information was provided by the hospital's gastroenterology department and updated regularly. Keyword expertise was extracted from physician profiles, such as "Helicobacter pylori infection," "peptic ulcer," "chronic gastritis," and "reflux esophagitis." For example, the expertise of Doctor A in the gastroenterology department included "Helicobacter pylori," "chronic gastritis," "upper abdominal pain," and "breath test"; Doctor B's expertise included "peptic ulcer," "acid reflux," "heartburn," and "antibiotic treatment"; and Doctor C's expertise included "gastroesophageal reflux," "gastrointestinal motility," "functional dyspepsia," and "abdominal bloating." Weights were assigned to each physician's expertise to reflect the degree to which the keyword represents the physician's professional expertise, particularly their expertise in the diagnosis and treatment of Helicobacter pylori-related diseases. Weighting is based on the doctor's experience in Helicobacter pylori research (more experience, more weight), the number of relevant papers published (more number, more weight), the number of Helicobacter pylori cases successfully treated (more cases, more weight), and the accuracy of breath test interpretation (more accuracy, more weight). For example, for Doctor A, who has eight years of experience in Helicobacter pylori research, the keyword weight for "Helicobacter pylori" is 0.95, "chronic gastritis" is 0.9, "upper abdominal pain" is 0.85, and "breath test" is 0.92.

[0047] To calculate semantic similarity, a word embedding model specifically trained for digestive diseases was used to convert symptom keywords and physician-specific keywords into high-dimensional vectors. This word embedding model was trained on a large-scale digestive disease literature corpus, containing approximately 500,000 digestive disease-related terms. Each term was mapped into a 250-dimensional vector space. For example, "upper abdominal pain" might be represented as a 250-dimensional vector [0.15, -0.08, 0.21, ..., 0.05]. The cosine similarity between the target object's symptom keyword vector and the physician-specific keyword vector was calculated. For example, the cosine similarity between "upper abdominal pain" and "Helicobacter pylori" was 0.82, the cosine similarity with "peptic ulcer" was 0.78, and the cosine similarity with "gastroesophageal reflux" was 0.45, indicating that "upper abdominal pain" is semantically closest to "Helicobacter pylori." The cosine similarity between each pair of symptom keywords and the doctor's professional keywords is multiplied by the corresponding weight. The weighted similarities of all pairs are then summed and normalized by dividing by the sum of the weights. For example, for the target subject and Doctor A, the calculation process includes: the similarity between "upper abdominal pain" (weight 0.85) and "Helicobacter pylori" (weight 0.95) is 0.82, resulting in a weighted value of 0.85 × 0.95 × 0.82 = 0.6627; the similarity between "acid reflux" (weight 0.75) and "Helicobacter pylori" (weight 0.95) is 0.68, resulting in a weighted value of 0.75 × 0.95 × 0.68 = 0.4845. Similarly, the weighted similarities of all pairs are calculated, summed, and then divided by the sum of the weighted products, resulting in a comprehensive semantic similarity of 0.78 between the target subject and Doctor A. Semantic similarity is calculated for the target subject and all available gastroenterologists and ranked according to similarity. For example, the semantic similarity between the target object and Doctor A is 0.78, with Doctor B it is 0.71, and with Doctor C it is 0.52, indicating that the target object's symptoms are most consistent with Doctor A's expertise, especially in the diagnosis and treatment of Helicobacter pylori.

[0048] To further improve the matching accuracy of Helicobacter pylori breath tests, the system has enhanced semantic understanding related to the breath test process. When identifying key information such as "breath test performed," "taking C13 medication," and "waiting for test time" in the target subject's description, it prioritizes matching with physicians experienced in breath test interpretation and follow-up treatment. For example, if the target subject mentions "previously had a breath test with inconclusive results" or "tested 30 minutes after taking C13 tablets," the system will increase the semantic weight associated with keywords such as "breath test," "C13 urea," and "Helicobacter pylori test."

[0049] Considering the doctor's historical diagnosis and treatment data, the average treatment time is calculated. For example, if Doctor A's average treatment time for patients with similar symptoms is 15 minutes, while the standard treatment time for this type of symptom is 20 minutes, the experience matching degree is 15 ÷ 20 = 0.75, indicating that the doctor's treatment efficiency is high. The semantic similarity and experience matching degrees are weighted and summed, with weights of 0.6 and 0.4, respectively, resulting in a treatment matching degree of 0.85 × 0.6 + 0.75 × 0.4 = 0.81, which comprehensively reflects the degree of doctor-patient matching.

[0050] For resource utilization, the system calculates the doctor's current visit rate: the number of patients seen divided by the maximum number of patients seen. For example, for Doctor A, whose maximum number of patients seen that day was 30 and who had seen 12, the visit rate is 12 ÷ 30 = 0.4. The average visit rate of all doctors in the department is calculated as the overall resource utilization rate. For example, if the visit rates of five doctors in the internal medicine department are 0.4, 0.5, 0.3, 0.6, and 0.4, respectively, the average resource utilization rate is 0.44.

[0051] A multi-objective optimization function was constructed by weighting time cost, diagnosis and treatment matching, and resource utilization. The weights w1 = 0.4, w2 = 0.4, and w3 = 0.2 were set, and the multi-objective function value was calculated as (1 - 0.81) × 0.4 + 0.81 × 0.4 + 0.44 × 0.2 = 0.464. The time cost was converted to a time-effectiveness indicator by subtracting the original time cost from 1. Based on the multi-objective function value, the optimal doctor was recommended for each patient in the queue.

[0052] During implementation, a distributed computing architecture was adopted to handle large-scale concurrent requests. Patient data and physician information were stored in a distributed database, and a caching mechanism was used to speed up data access. The semantic similarity calculation module adopted a pre-computation strategy, storing the similarity between common symptoms and professional keywords in a lookup table to reduce real-time computing overhead. Queuing status and resource utilization were updated every 5 minutes to ensure that decisions were based on the latest data. In addition, an adaptive weight adjustment mechanism was designed. By analyzing historical queuing data and patient satisfaction feedback, the weights of time cost, diagnosis and treatment matching, and resource utilization were dynamically adjusted. For example, during peak treatment periods, the weight of resource utilization was increased to improve overall throughput; during non-peak periods, the weight of diagnosis and treatment matching was increased to enhance medical quality.

[0053] Figure 2This is a diagram of the architecture of an adaptive allocation optimization method based on multi-objective constraints. In one optional implementation, based on the initial allocation, multiple adjustment plans that satisfy the multi-objective constraints are generated, and their state transition probabilities are calculated. A roulette wheel approach is used to update the search direction, combining the multi-objective function. The allocation plan is adjusted using an adaptive step size, and iterations are repeated until the multi-objective function value is less than the optimization threshold. The optimal allocation plan is then output, including: Obtaining an initial allocation plan for the target objects, generating multiple adjustment paths for each target object while satisfying multi-objective constraints, and calculating the multi-objective function value on each adjustment path; calculating the product of the initial pheromone concentration of each adjustment path and the inverse of the multi-objective function value to obtain a state transition probability; Calculate the multi-objective function difference between the current adjustment path and the initial allocation plan, and dynamically update the search direction on the current adjustment path using a roulette wheel method in combination with the state transition probability; determine the product of the maximum adjustment step length, the complementary power term of the iteration progress, and the multi-objective function difference as the adaptive step length, and adjust the allocation plan of the target object based on the updated search direction and the adaptive step length; Repeat the above iterative process until the multi-objective function values ​​of all target objects under the current allocation scheme are less than the optimization threshold, and then output the current allocation scheme as the optimal allocation scheme.

[0054] The present invention provides an adaptive allocation solution optimization method based on multi-objective constraints. This method can, while satisfying the multi-objective constraints, update the search direction through adaptive step size adjustment and a roulette wheel method, ultimately outputting the optimal allocation solution. The specific implementation steps of this method are described in detail below.

[0055] When implementing this method, an initial allocation plan for target patients is required. This initial allocation plan is generated based on the target patient's health index. A higher health index value indicates a worse health condition and a higher priority. For example, at a breathalyzer site, there are 20 target patients awaiting triage, with health indices of [0.35, 0.82, 0.56, 0.74, 0.41, 0.63, 0.29, 0.77, 0.52, 0.68, 0.44, 0.89, 0.37, 0.71, 0.49, 0.65, 0.58, 0.46, 0.83, 0.39]. The site has five available clinics, numbered 1 to 5. Target patients are initially ranked based on their health index, with those with higher health indices receiving higher priority. The sorted target object sequence is [0.89, 0.83, 0.82, 0.77, 0.74, 0.71, 0.68, 0.65, 0.63, 0.58, 0.56, 0.52, 0.49, 0.46, 0.44, 0.41, 0.39, 0.37, 0.35, 0.29], and the corresponding target object numbers are [12, 19, 2, 8, 4, 14, 10, 16, 6, 17, 3, 9, 15, 18, 11, 5, 20, 13, 1, 7]. The target objects are initially assigned to each clinic, considering the number of clinics and the number of target objects. In this example, an average of 4 target objects are assigned to each of the 5 clinics. The system assigns the target object numbered 12 to Clinic 1, the target object numbered 19 to Clinic 2, and so on in order of priority. The initial allocation plan is: Clinic 1: [12, 6, 15, 7], Clinic 2: [19, 16, 9, 13], Clinic 3: [2,17, 18, 1], Clinic 4: [8, 3, 11, 20], Clinic 5: [4, 14, 10, 5].

[0056] For each target object, multiple possible adjustment paths are generated. The adjustment path represents a possible solution for adjusting the target object from the current clinic to another clinic. For example, for the target object numbered 12 initially assigned to Clinic 1, the possible adjustment paths include: adjustment to Clinic 2, adjustment to Clinic 3, adjustment to Clinic 4, and adjustment to Clinic 5. The system generates all possible adjustment paths for each target object. In this example, each target object has 4 possible adjustment paths, a total of 20 target objects, and a total of 80 adjustment paths. Check whether each adjustment path meets the multi-objective constraints. For the path of adjusting the target object numbered 12 from Clinic 1 to Clinic 2, check whether the number of target objects in Clinic 2 after adjustment meets the resource utilization constraint, check whether the matching degree between the target object and the doctor in Clinic 2 meets the diagnosis and treatment matching constraint, and check whether the waiting time of the target object after adjustment meets the waiting time constraint. If any constraint is not met, the adjustment path is marked as invalid.

[0057] Calculate the multi-objective function value for each valid adjustment path. The calculation method is the same as for evaluating the initial allocation plan, but based on the adjusted allocation plan. For example, after adjusting target patient number 12 from Clinic 1 to Clinic 2, Clinic 1 becomes [6, 15, 7] and Clinic 2 becomes [19, 12, 16, 9, 13]. Recalculate the waiting time for each target patient. For example, target patient number 12 is second in Clinic 2 and has a waiting time of 15 minutes. Recalculate the overall treatment matching degree. For example, the matching degree of target patient number 12 with Clinic 2 is 0.82, which is higher than the matching degree of 0.45 with Clinic 1. Recalculate resource utilization. For example, after the adjustment, Clinic 1 has 3 target patients and Clinic 2 has 5 target patients, resulting in a slight decrease in resource utilization.

[0058] The state transition probability is calculated by multiplying the initial pheromone concentration of each adjusted path by the inverse of the multi-objective function value. Assuming that the initial pheromone concentration of paths A, B, and C is 1.0, and the multi-objective function values ​​are 0.35, 0.42, and 0.28, respectively (smaller values ​​are preferred), the state transition probabilities are: 1.0 × (1 / 0.35) = 2.86, 1.0 × (1 / 0.42) = 2.38, and 1.0 × (1 / 0.28) = 3.57. After normalization, the state transition probabilities are 0.324 for path A, 0.270 for path B, and 0.406 for path C. The difference in the multi-objective function between the current adjusted path and the initial allocation plan is calculated. Taking path C as an example, the multi-objective function value of the initial plan is 0.65, while that of path C is 0.28, resulting in a difference of 0.37. Based on the state transition probabilities, the system dynamically updates the search direction on the current adjusted path using a roulette wheel method. In implementation, a random number between 0 and 1 is generated, such as 0.36, which falls within the probability interval of search direction 6 on path C. Therefore, search direction 6 on path C is selected as the search direction for the current iteration. The adaptive step size is determined by multiplying the maximum adjustment step size by the power term of the iteration progress's complement and the difference between the multi-objective functions. Assuming the maximum adjustment step size is 1.0 and the current iteration is the first (10 iterations are planned), the iteration progress's complement is (1 - 0.1) = 0.9, and the power term of the complement is 2, or 0.81. The adaptive step size is calculated as 1.0 × 0.81 × 0.37 = 0.3, representing the magnitude of the adjustment on path C.

[0059] The target object allocation scheme is adjusted based on the updated search direction and adaptive step size. During the adjustment process, the mutual influence of multiple target objects is simultaneously considered. For example, when the step size is 3, three target objects are selected for adjustment: ID-12 is adjusted from Clinic 1 to Clinic 2, ID-8 is adjusted from Clinic 4 to Clinic 3, and ID-15 is adjusted from Clinic 1 to Clinic 5. The multi-objective function values ​​after these adjustment combinations are calculated. If the adjusted function value is better than the original value, the adjustment scheme is accepted; otherwise, new adjustment combinations are generated until an improved solution is found or the number of attempts reaches a limit. A strategy combining greedy search and simulated annealing is used to select adjustment schemes. In most cases, an adjustment scheme that improves the multi-objective function value is selected; however, with a certain probability, an adjustment scheme that temporarily does not improve or even slightly worsens the value is accepted to escape the local optimal solution. The acceptance probability is related to the current iteration number and the degree of function value deterioration. For example, in the 50th iteration, if an adjustment plan causes the multi-objective function value to deteriorate from 2.34 to 2.36, with a deterioration of 0.02, the system will accept the plan with a probability of 0.15. As the number of iterations increases, the acceptance probability gradually decreases, and the optimization process gradually converges.

[0060] After each adjustment, the target patient's waiting time is updated. The waiting time depends on the target patient's ranking position in the clinic and the estimated visit duration of the previous target patient. The visit duration for each target patient is estimated based on historical data. For example, the estimated visit duration for target patient ID-19, whose health index is 0.82, is 20 minutes, while the estimated visit duration for target patient ID-3, whose health index is 0.56, is 15 minutes. After the adjustment, if target patient ID-12 is ranked after ID-19, its waiting time will be 20 minutes. The diagnosis and treatment match is recalculated after each adjustment. For example, if the symptom feature vector of target patient ID-12 is [0.8, 0.6, 0.3, 0.7, 0.2] and the professional expertise vector of the doctor in clinic 2 is [0.75, 0.65, 0.4, 0.8, 0.3], the calculated match is 0.82. After the adjustment, the diagnosis and treatment match is updated for all affected target patients. Resource utilization is recalculated after each adjustment. For example, before the adjustment, the number of target patients in each clinic was [4, 4, 4, 4, 4], with a standard deviation of 0 and the highest resource utilization; after the adjustment, it became [3, 5, 5, 3, 4], with a standard deviation of approximately 0.89 and a slight decrease in resource utilization. The multi-objective function value is calculated based on the updated waiting time, diagnosis and treatment matching, and resource utilization. An optimization threshold is set as the iteration termination condition. The optimization threshold is the expected minimum value of the multi-objective function value and indicates the degree of satisfaction achieved by the allocation scheme. For example, setting the optimization threshold to 1.5 means that when the overall multi-objective function value is less than 1.5, the allocation scheme is considered sufficiently optimized. It is also possible to further require that the individual multi-objective function values ​​of all target patients be less than the optimization threshold to ensure the balance of the allocation scheme.

[0061] The iterative process of updating the search direction, adjusting the step size, adjusting the allocation scheme, and calculating the multi-objective function value is repeated. In each iteration, the system records the current allocation scheme and the corresponding multi-objective function value. For example, the allocation plan after the first iteration is: Clinic 1: [6, 15, 7], Clinic 2: [19, 12, 16, 9, 13], Clinic 3: [2, 8, 17, 18, 1], Clinic 4: [3, 11, 20], Clinic 5: [4, 14, 10, 5], and the corresponding multi-objective function value is 2.25. The allocation plan after the 50th iteration is: Clinic 1: [6, 1, 20, 5], Clinic 2: [19, 12, 16, 13], Clinic 3: [2, 8, 17, 3], Clinic 4: [4, 11, 10, 7], Clinic 5: [14, 9, 18, 15], and the corresponding multi-objective function value is 1.78. Determine whether the current iteration meets the termination condition. Termination conditions include: the individual multi-objective function values ​​for all target objects are less than the optimization threshold; or a certain number of consecutive iterations fail to significantly improve the multi-objective function value, such as an improvement less than 0.001 for 30 consecutive iterations; or the maximum number of iterations is reached, such as 500. If any of these conditions are met, the system stops iterating; otherwise, it continues to the next iteration. For example, if, after the 217th iteration, the individual multi-objective function values ​​for all target objects are less than the optimization threshold of 1.5, the system terminates the iteration process. From all recorded allocation solutions, the solution with the smallest multi-objective function value is selected and output as the optimal allocation solution. For example, the allocation plan for the 198th iteration is: Clinic 1: [7, 1, 5], Clinic 2: [19, 12, 9, 13], Clinic 3: [2, 8, 17, 3, 20], Clinic 4: [4, 11, 16], Clinic 5: [14, 10, 18, 15]. The corresponding multi-objective function value is 1.42, which is the minimum value among all iterations, and is therefore selected as the optimal allocation plan.

[0062] The optimal allocation plan is converted into specific triage instructions, including the target patient ID, assigned clinic, ranking position within the clinic, and estimated wait time. For example, target patient ID-12 is assigned to clinic 2, ranking position 2, and estimated wait time 20 minutes. The triage instructions are sent to the corresponding target patient and medical staff via the display terminal at the breathalyzer site or the mobile app, completing the entire intelligent triage process.

[0063] In practical applications, the weights and constraints of the multi-objective function can be adjusted based on different scenarios. For example, for latency-sensitive applications, the weight of quality of service in the multi-objective function can be increased; for resource-constrained environments, resource utilization constraints can be strengthened. Through this adaptive optimization method, the system can find the optimal allocation solution that meets multiple objective constraints in complex and changing environments.

[0064] In an optional embodiment, the difference between the multi-objective function of the current adjustment path and the initial allocation solution is calculated, and the search direction on the current adjustment path is dynamically updated using a roulette wheel method in combination with the state transition probability, including: Obtain the current adjustment path and its state transition probability, calculate the complement of the ratio of the current number of iterations to the maximum number of iterations, multiply the power product of the complement and the state transition probability by the probability threshold, and obtain the adaptive selection probability; Calculating a multi-objective function difference between the current adjustment path and the initial allocation solution, and when the multi-objective function difference is less than the adaptive selection probability, selecting a search direction with the maximum local pheromone concentration on the adjustment path; When the difference of the multi-objective function is greater than the adaptive selection probability, the complementary value of the pheromone volatility coefficient of the current adjustment path is multiplied by the local pheromone concentration of each search direction on the current adjustment path, and normalized to obtain the current selection probability of each search direction; according to the numbering order of each search direction, the current selection probabilities are accumulated one by one to obtain the cumulative selection probability of the search direction; the difference between the random number and the cumulative selection probability is calculated, and the search direction with the smallest difference is selected as the final search direction.

[0065] The present invention provides a resource scheduling method based on an ant colony optimization algorithm, which improves scheduling efficiency by dynamically updating search directions. The technical content of the present invention is described in detail below.

[0066] Get the current adjustment path and its state transition probability. Assume that the current adjustment path is P and the state transition probability is p trans , the current number of iterations is iter current , the maximum number of iterations is iter max The system calculates the ratio of the current number of iterations to the maximum number of iterations as ratio = iter current / iter max , and then calculate the complement of the ratio comp ratio = 1 - ratio. Add the power product of the complement value and the state transition probability to the preset probability threshold to obtain the adaptive selection probability p adaptive = comp ratio ^p trans + threshold. For example, when the current number of iterations is 50, the maximum number of iterations is 200, the state transition probability is 0.7, and the probability threshold is 0.1, the ratio complement is 1-50 / 200=0.75, and the adaptive selection probability pa daptive It is 0.75^0.7+0.1=0.72.

[0067] The multi-objective function difference diff of the current adjustment path P and the initial allocation scheme P0 is calculated. Assuming that the multi-objective function value of the initial allocation scheme is f(P0) = 2.3 and the multi-objective function value of the current adjustment path is f(P) = 2.1, the multi-objective function difference is diff = |f(P) - f(P0)| = |78 - 85| = 0.2. It is judged whether the multi-objective function difference norm is less than the adaptive selection probability p adaptive In the above example, since 0.2 < 0.72, the search direction with the maximum local pheromone concentration on the current adjustment path is selected. Assuming that there are three selectable search directions D1, D2 and D3 on the current adjustment path, with local pheromone concentrations of 0.45, 0.35 and 0.20 respectively, the system selects the search direction D1 with the maximum local pheromone concentration as the final search direction.

[0068] When the normalized multi-objective function difference is greater than the adaptive selection probability, a roulette wheel-based method is used to select the search direction. In another case, the multi-objective function difference is 0.85 and the adaptive selection probability is 0.72, since 0.85 > 0.72, the selection probability of each search direction needs to be further calculated. The pheromone evaporation coefficient rho of the current adjustment path is obtained, for example, rho = 0.3, and the complement value comp rho = 1 - rho = 0.7 is calculated. This complement value is multiplied by the local pheromone concentration of each search direction on the current adjustment path to obtain the original selection probability of the three search directions: prob D1 = 0.7 x 0.45 = 0.315, prob D2 = 0.7 x 0.35 = 0.245, prob D3 = 0.7 x 0.2 = 0.14. The original selection probabilities are normalized to obtain the current selection probability of each search direction: norm probD1 = 0.315 / (0.315 + 0.245 + 0.14) = 0.45, norm probD2 = 0.245 / (0.315 + 0.245 + 0.14) = 0.35, norm probD3 = 0.14 / (0.315 + 0.245 + 0.14) = 0.2. According to the numbering order of the search directions, the system accumulates the current selection probabilities one by one to obtain the cumulative selection probability of each search direction: cum probD1 = 0.45, cum probD2 = 0.45 + 0.35 = 0.8, cum probD3 = 0.45 + 0.35 + 0.2 = 1. A random number rand between 0 and 1 is generated, for example, rand = 0.6. The difference between the random number and the cumulative selection probability of each search direction is calculated: diff D1=|0.6-0.45|=0.15, diff D2 =|0.6-0.8|=0.2, diff D3 =|0.6-1|=0.4. The system selects the search direction with the smallest difference as the final search direction, which is D1 in this example.

[0069] Through this method, the system achieves the ability to dynamically balance global exploration and local exploitation. In the initial iterations, the probability of adaptive selection is high, and the system tends to select search directions based on local pheromone concentrations, which favors global exploration. As the number of iterations increases, the probability of adaptive selection decreases, and the system tends to select search directions using a roulette-wheel approach, enhancing local exploitation capabilities. By dynamically adjusting the search strategy, the method of the present invention achieves a good balance between global exploration and local exploitation, avoiding regression into local optimal solutions while simultaneously accelerating convergence.

[0070] The embodiment of the present invention is based on a deep learning breath test triage and queuing optimization system, including: The first unit is used to collect multi-dimensional physiological data of the target object from the breath experiment site; The second unit is configured to perform time-frequency domain decomposition on the multidimensional physiological data, extract a time-domain feature sequence and a frequency feature spectrum; extract non-stationary feature vectors from the time-domain feature sequence, perform cluster analysis on the frequency feature spectrum, and form a feature combination; input the non-stationary feature and the feature combination into a pre-built hybrid expert model, and fuse the prediction results of multiple expert sub-networks to generate a health index of the target object; The third unit is used to obtain the total number of target subjects and the number of clinics, triage the target subjects according to the health index, and obtain an initial allocation plan; The fourth unit is used to construct a multi-objective function, calculate the time cost based on the waiting time of the target object, and calculate the diagnosis and treatment matching degree and resource utilization rate based on the doctor's expertise and consultation rate; based on the initial allocation plan, generate multiple adjustment plans that meet the multi-objective constraints and calculate their state transition probabilities; based on the multi-objective function, use a roulette wheel method to update the search direction, adjust the allocation plan through adaptive step size, repeat the iteration until the multi-objective function value is less than the optimization threshold, and output the optimal allocation plan; The fifth unit is used to generate a triage number and an estimated consultation time according to the optimal allocation plan, and send the triage result to the terminal device of the target object.

[0071] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0072] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0073] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements 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 deep learning-based breath test triage and queuing optimization method, characterized by: include: Collect multi-dimensional physiological data of the target subject from the breath experiment site; Performing time-frequency decomposition on the multidimensional physiological data to extract a time-domain feature sequence and a frequency feature spectrum; extracting non-stationary feature vectors from the time-domain feature sequence, performing cluster analysis on the frequency feature spectrum to form a feature combination; inputting the non-stationary feature and the feature combination into a pre-built hybrid expert model, fusing the prediction results of multiple expert sub-networks to generate a health index of the target object; Obtaining the total number of target subjects and the number of clinics, triaging the target subjects according to the health index, and obtaining an initial allocation plan; A multi-objective function is constructed to calculate the time cost based on the target patient's waiting time, and the diagnosis and treatment matching degree and resource utilization rate are calculated based on the doctor's expertise and consultation rate. Based on the initial allocation plan, multiple adjustment plans that meet the multi-objective constraints are generated and their state transition probabilities are calculated. Based on the multi-objective function, a roulette wheel method is used to update the search direction, and the allocation plan is adjusted by adaptive step size. It is repeated until the multi-objective function value is less than the optimization threshold, and the optimal allocation plan is output. A triage number and an estimated consultation time are generated according to the optimal allocation plan, and the triage result is sent to the terminal device of the target object.

2. The method according to claim 1, characterized in that Performing time-frequency domain decomposition on the multidimensional physiological data to extract a time-domain feature sequence and a frequency feature spectrum; extracting non-stationary feature vectors in the time-domain feature sequence, performing cluster analysis on the frequency feature spectrum, and forming a feature combination, including: Performing empirical mode decomposition on the multidimensional physiological data to obtain a plurality of intrinsic mode functions, and performing time series reconstruction on the plurality of intrinsic mode functions to obtain a time domain feature sequence; extracting instantaneous frequency features from the multidimensional physiological data through Hilbert transform to obtain a frequency feature spectrum; Based on a preset time window, a sliding analysis is performed on the time domain feature sequence to calculate the local energy density within the time window; a convolution operation is performed on the time domain feature sequence with wavelet basis functions of different scales to obtain detail coefficients of multiple scales, a weighted sum is taken of the squares of the detail coefficients, and the sum is multiplied by the local energy density to obtain a non-stationary feature; An initial neighborhood radius is set, and the initial neighborhood radius is linearly reduced as the number of iterations increases. The frequency feature spectrum is clustered based on the current neighborhood radius and the preset learning rate to obtain multiple feature combinations and their cluster centers. When the distances between the cluster centers are all greater than the clustering threshold, the current feature combination is taken as the optimal feature combination.

3. The method according to claim 1, characterized in that The non-stationary features and the feature combination are input into a pre-built hybrid expert model, and the prediction results of multiple expert sub-networks are integrated to generate a health index of the target object, including: For each expert sub-network, the non-stationary features and the feature combination are subjected to multi-layer convolution and pooling operations to extract hierarchical features and generate preliminary prediction results. The preliminary prediction results are error-compensated using a residual learning mechanism to obtain residual correction terms. The preliminary prediction results and the residual correction terms are weightedly fused to obtain corrected prediction values ​​for each expert sub-network. The corrected prediction values ​​of each expert sub-network are input into the gated recurrent unit according to a preset time step, and a time series prediction result is generated by controlling the forgetting and updating of historical information; the prediction error between the time series prediction result and the true value is calculated, and a fusion weight is assigned to each expert sub-network through iterative optimization. When the error change in multiple consecutive iterations is less than the error threshold, the iteration is stopped to obtain the optimal fusion weight; The optimal fusion weight is weightedly combined with the corrected prediction values ​​of each expert sub-network to obtain the health index of the target object.

4. The method according to claim 1, wherein Construct a multi-objective function, calculate the time cost based on the target patient's waiting time, and calculate the diagnosis and treatment matching degree and resource utilization rate based on the doctor's expertise and consultation rate, including: Obtain the number of people queuing in the department, the number of doctors on duty, and the service capacity per unit time. Divide the number of people queuing in the department by the product of the number of doctors on duty and the service capacity per unit time to obtain the basic waiting time. Substitute the basic waiting time into the exponential decay function to obtain the time cost. Calculate the semantic similarity between the target subject's self-reported symptom keywords and the doctor's professional expertise keywords; obtain the average treatment time of the doctor's historical treatment data, and use the ratio of the average treatment time to the standard treatment time as the experience matching degree; and obtain the treatment matching degree by weighted summation of the semantic similarity and the experience matching degree; The number of patients the doctor has seen is divided by the maximum number of patients the doctor can see to obtain the consultation rate, and the average value of the consultation rate is used as the resource utilization rate; the time cost, the diagnosis and treatment matching degree and the resource utilization rate are weighted and summed to obtain a multi-objective function.

5. The method according to claim 1, wherein Based on the initial allocation plan, multiple adjustment plans that meet the multi-objective constraints are generated, and their state transition probabilities are calculated; combined with the multi-objective function, the search direction is updated using a roulette wheel method, the allocation plan is adjusted by an adaptive step size, and it is iterated repeatedly until the multi-objective function value is less than the optimization threshold, and the optimal allocation plan is output, including: Obtaining an initial allocation plan for the target objects, generating multiple adjustment paths for each target object while satisfying multi-objective constraints, and calculating the multi-objective function value on each adjustment path; calculating the product of the initial pheromone concentration of each adjustment path and the inverse of the multi-objective function value to obtain a state transition probability; Calculate the multi-objective function difference between the current adjustment path and the initial allocation plan, and dynamically update the search direction on the current adjustment path using a roulette wheel method in combination with the state transition probability; determine the product of the maximum adjustment step length, the complementary power term of the iteration progress, and the multi-objective function difference as the adaptive step length, and adjust the allocation plan of the target object based on the updated search direction and the adaptive step length; Repeat the above iterative process until the multi-objective function values ​​of all target objects under the current allocation scheme are less than the optimization threshold, and then output the current allocation scheme as the optimal allocation scheme.

6. The method according to claim 5, characterized in that Calculate the difference between the multi-objective function of the current adjustment path and the initial allocation plan, combine the state transition probability, and use the roulette method to dynamically update the search direction on the current adjustment path, including: Obtain the current adjustment path and its state transition probability, calculate the complement of the ratio of the current number of iterations to the maximum number of iterations, multiply the power product of the complement and the state transition probability by the probability threshold, and obtain the adaptive selection probability; Calculating a multi-objective function difference between the current adjustment path and the initial allocation solution, and when the multi-objective function difference is less than the adaptive selection probability, selecting a search direction with the maximum local pheromone concentration on the adjustment path; When the difference of the multi-objective function is greater than the adaptive selection probability, the complementary value of the pheromone volatility coefficient of the current adjustment path is multiplied by the local pheromone concentration of each search direction on the current adjustment path, and normalized to obtain the current selection probability of each search direction; according to the numbering order of each search direction, the current selection probabilities are accumulated one by one to obtain the cumulative selection probability of the search direction; the difference between the random number and the cumulative selection probability is calculated, and the search direction with the smallest difference is selected as the final search direction.

7. A deep learning-based breath test triage and queuing optimization system for implementing the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to collect multi-dimensional physiological data of the target object from the breath experiment site; The second unit is configured to perform time-frequency domain decomposition on the multidimensional physiological data, extract a time-domain feature sequence and a frequency feature spectrum; extract non-stationary feature vectors from the time-domain feature sequence, perform cluster analysis on the frequency feature spectrum, and form a feature combination; input the non-stationary feature and the feature combination into a pre-built hybrid expert model, and fuse the prediction results of multiple expert sub-networks to generate a health index of the target object; The third unit is used to obtain the total number of target subjects and the number of clinics, triage the target subjects according to the health index, and obtain an initial allocation plan; The fourth unit is used to construct a multi-objective function, calculate the time cost based on the waiting time of the target object, and calculate the diagnosis and treatment matching degree and resource utilization rate based on the doctor's expertise and consultation rate; based on the initial allocation plan, generate multiple adjustment plans that meet the multi-objective constraints and calculate their state transition probabilities; based on the multi-objective function, use a roulette wheel method to update the search direction, adjust the allocation plan through adaptive step size, repeat the iteration until the multi-objective function value is less than the optimization threshold, and output the optimal allocation plan; The fifth unit is used to generate a triage number and an estimated consultation time according to the optimal allocation plan, and send the triage result to the terminal device of the target object.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images