Deep learning-based breath test triage and queuing optimization method and system

Abstract

The application provides a deep learning-based breath test triage and queuing optimization method and system, relating to the technical field of medical management, comprising collecting multi-dimensional physiological data, performing time-frequency domain decomposition to extract features, generating a health index using a hybrid expert model, constructing a multi-objective optimization function by combining time cost, diagnosis and treatment matching degree, and resource utilization rate, and generating an optimal allocation scheme through adaptive search. This method can improve triage accuracy, shorten patient waiting time, optimize medical resource allocation, and improve treatment efficiency and satisfaction.

Description

Technical Field

[0001] This invention relates to the field of medical management technology, and in particular to a method and system for triage and queuing optimization based on deep learning-based breath test. Background Technology

[0002] With the increasing demand for healthcare, medical institutions face problems such as long patient wait times and low triage efficiency. Traditional medical queuing systems are usually based on a first-come, first-served principle or simple triage criteria, which cannot effectively identify the actual health status and urgency of patients, resulting in low triage accuracy, difficulty in optimizing overall medical outcomes, and leading to unreasonable allocation of medical resources and low patient satisfaction. In recent years, breath analysis technology has gradually gained attention as a non-invasive and convenient means of physiological state detection, assessing health status by detecting specific components in exhaled breath. At the same time, the application of deep learning technology in the medical field is becoming increasingly widespread, providing new possibilities for breath data analysis and patient triage. Summary of the Invention

[0003] This invention provides a method and system for triage and queuing optimization based on deep learning in breath experiments, which can solve the problems in the prior art.

[0004] A first aspect of the present invention provides a deep learning-based method for triage and queuing optimization in breath experiments, comprising:

[0005] Multidimensional physiological data of target subjects were collected from breath test sites;

[0006] The multidimensional physiological data is decomposed in the time-frequency domain to extract time-domain feature sequences and frequency feature spectra; non-stationary feature vectors are extracted from the time-domain feature sequences, and cluster analysis is performed on the frequency feature spectra to form feature combinations; the non-stationary features and the feature combinations are input into a pre-constructed hybrid expert model, and the prediction results of multiple expert sub-networks are fused to generate the health index of the target object.

[0007] Obtain the total number of target individuals and the number of consultation rooms, and triage the target individuals according to the health index to obtain an initial allocation plan;

[0008] A multi-objective function is constructed. Time cost is calculated based on the waiting time of the target object. Treatment matching degree and resource utilization are calculated by combining the doctor's specialty and consultation rate. Based on the initial allocation scheme, multiple adjustment schemes satisfying the multi-objective constraints are generated, and their state transition probabilities are calculated. Combining the multi-objective function, the search direction is updated using a roulette wheel approach. The allocation scheme is adjusted through adaptive step size, and this process is repeated iteratively until the multi-objective function value is less than the optimization threshold. The optimal allocation scheme is then output.

[0009] Based on the optimal allocation scheme, a triage number and an estimated appointment time are generated, and the triage results are sent to the target's terminal device.

[0010] The multidimensional physiological data are decomposed in the time-frequency domain to extract time-domain feature sequences and frequency feature spectra; non-stationary feature vectors are extracted from the time-domain feature sequences, and cluster analysis is performed on the frequency feature spectra to form feature combinations, including:

[0011] Empirical mode decomposition is performed on the multidimensional physiological data to obtain multiple intrinsic mode functions (IMFs). Temporal reconstruction is then performed on the multiple IMFs to obtain a temporal feature sequence. Instantaneous frequency features are extracted from the multidimensional physiological data using Hilbert transform to obtain a frequency feature spectrum.

[0012] 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; the time-domain feature sequence is convolved with wavelet basis functions of different scales to obtain detail coefficients at multiple scales; the squares of the detail coefficients are weighted and summed, and then multiplied by the local energy density to obtain non-stationary features;

[0013] 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 a preset learning rate to obtain multiple feature combinations and their cluster centers. When the distance between the cluster centers is greater than the clustering threshold, the current feature combination is taken as the optimal feature combination.

[0014] The non-stationary features and the feature combination are input into a pre-constructed hybrid expert model, and the prediction results of multiple expert sub-networks are fused to generate a health index for the target object, including:

[0015] For each expert subnetwork, 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 then compensated for errors using a residual learning mechanism to obtain residual correction terms. The preliminary prediction results and residual correction terms are then weighted and fused to obtain the corrected prediction values ​​for each expert subnetwork.

[0016] The corrected prediction values ​​of each expert subnetwork are input into the gated loop unit according to a preset time step. By controlling the forgetting and updating of historical information, a time series prediction result is generated. The prediction error between the time series prediction result and the true value is calculated. The fusion weight is assigned to each expert subnetwork through iterative optimization. The iteration stops when the error change in multiple consecutive iterations is less than the error threshold, and the optimal fusion weight is obtained.

[0017] The optimal fusion weights are weighted and combined with the corrected prediction values ​​of each expert subnetwork to obtain the health index of the target object.

[0018] Construct a multi-objective function to calculate time cost based on the waiting time of the target object, and calculate the treatment matching degree and resource utilization rate by combining the doctor's expertise and consultation rate, including:

[0019] The number of people queuing in a department, the number of doctors on duty, and the service capacity per unit time are obtained. The basic waiting time is obtained by dividing the number of people queuing in a department by the product of the number of doctors on duty and the service capacity per unit time. The time cost is obtained by substituting the basic waiting time into an exponential decay function.

[0020] Calculate the semantic similarity between keywords of the target subject's self-reported symptoms and keywords of the doctor's area of ​​expertise; obtain the average consultation time of the doctor's historical treatment data, and use the ratio of the average consultation time to the standard consultation time as the empirical matching degree; and obtain the treatment matching degree by weighted summation of the semantic similarity and the empirical matching degree.

[0021] The consultation rate is obtained by dividing the number of patients already seen by the doctor by the maximum number of patients a doctor can see. The average value of the consultation rate is taken 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.

[0022] Based on the initial allocation scheme, multiple adjustment schemes satisfying multi-objective constraints are generated, and their state transition probabilities are calculated. Combining the multi-objective function, the search direction is updated using a roulette wheel approach. The allocation scheme is adjusted using an adaptive step size, and this process is repeated iteratively until the multi-objective function value is less than the optimization threshold. The optimal allocation scheme is then output, including:

[0023] Obtain the initial allocation scheme for the target objects. For each target object, generate multiple adjustment paths under the condition of satisfying multi-objective constraints, and calculate the multi-objective function value on each adjustment path. Calculate the product of the initial pheromone concentration of each adjustment path and the reciprocal of the multi-objective function value to obtain the state transition probability.

[0024] Calculate the multi-objective function difference between the current adjustment path and the initial allocation scheme, and dynamically update the search direction on the current adjustment path using a roulette wheel approach, combined with the state transition probability. Determine the adaptive step size by multiplying the maximum adjustment step size by the power of the complement of the iteration progress and the multi-objective function difference. Adjust the allocation scheme of the target object according to the updated search direction and the adaptive step size.

[0025] 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.

[0026] Calculate the multi-objective function difference between the current adjustment path and the initial allocation scheme. Combined with the state transition probability, dynamically update the search direction on the current adjustment path using a roulette wheel approach, including:

[0027] Obtain the current adjustment path and its state transition probability, calculate the complement of the ratio of the current iteration number to the maximum iteration number, and add the product of the complement and the power of the state transition probability to the probability threshold to obtain the adaptive selection probability;

[0028] Calculate the multi-objective function difference between the current adjustment path and the initial allocation scheme. When the multi-objective function difference is less than the adaptive selection probability, select the search direction with the highest local pheromone concentration on the adjustment path.

[0029] When the difference in the multi-objective function is greater than the adaptive selection probability, the complement of the pheromone evaporation 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 probability is 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.

[0030] A second aspect of the present invention provides a deep learning-based breath test triage and queuing optimization system, comprising:

[0031] The first unit is used to collect multidimensional physiological data of the target subjects from the breath test site;

[0032] The second unit is used to perform time-frequency domain decomposition on the multidimensional physiological data, extract time-domain feature sequences and frequency feature spectra; extract non-stationary feature vectors from the time-domain feature sequences, perform cluster analysis on the frequency feature spectra to form feature combinations; input the non-stationary features and the feature combinations into a pre-constructed hybrid expert model, and fuse the prediction results of multiple expert sub-networks to generate the health index of the target object;

[0033] The third unit is used to obtain the total number of target objects and the number of clinics, and to triage the target objects according to the health index to obtain an initial allocation plan;

[0034] The fourth unit is used to construct a multi-objective function. It calculates the time cost based on the waiting time of the target object, and calculates the treatment matching degree and resource utilization rate by combining the doctor's expertise and consultation rate. Based on the initial allocation scheme, it generates multiple adjustment schemes that satisfy the multi-objective constraints and calculates their state transition probabilities. Combining the multi-objective function, it updates the search direction using a roulette wheel approach, adjusts the allocation scheme through adaptive step size, and iterates repeatedly until the multi-objective function value is less than the optimization threshold, outputting the optimal allocation scheme.

[0035] The fifth unit is used to generate a triage number and an estimated appointment time according to the optimal allocation scheme, and send the triage results to the target's terminal device.

[0036] A third aspect of the present invention,

[0037] An electronic device is provided, comprising:

[0038] processor;

[0039] Memory used to store processor-executable instructions;

[0040] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0041] Fourth aspect of the embodiments of the present invention,

[0042] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0043] The beneficial effects of this application are as follows:

[0044] This invention collects multidimensional physiological data from breath test sites, extracts features through time-frequency domain decomposition, and generates a health index using a hybrid expert model. This enables precise assessment of patients' health status, improves the accuracy and scientific nature of triage, and avoids the problems of strong subjectivity and low precision in traditional triage methods.

[0045] This invention constructs a multi-objective function that considers three dimensions: time cost, diagnosis and treatment matching degree, and resource utilization rate. By adjusting the allocation scheme with an adaptive step size, it optimizes the allocation of medical resources while meeting the needs of patients, significantly reducing the average waiting time for patients and improving medical efficiency and patient satisfaction.

[0046] This invention organically combines deep learning technology with queuing optimization algorithms to form a complete closed-loop system from data collection and feature extraction to triage and queuing. It can not only accurately triage patients based on their actual health conditions, but also generate personalized appointment numbers and estimated appointment times, thereby improving the intelligence level of medical services and the patient's medical experience. It has good practicality and promotional value. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the deep learning-based expiratory test triage and queuing optimization method according to an embodiment of the present invention.

[0048] Figure 2 This is a system architecture diagram of an adaptive allocation scheme optimization method based on multi-objective constraints. Detailed Implementation

[0049] 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.

[0050] 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.

[0051] Figure 1 This is a flowchart illustrating the deep learning-based expiratory breath test triage and queuing optimization method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0052] Multidimensional physiological data of target subjects were collected from breath test sites;

[0053] The multidimensional physiological data is decomposed in the time-frequency domain to extract time-domain feature sequences and frequency feature spectra; non-stationary feature vectors are extracted from the time-domain feature sequences, and cluster analysis is performed on the frequency feature spectra to form feature combinations; the non-stationary features and the feature combinations are input into a pre-constructed hybrid expert model, and the prediction results of multiple expert sub-networks are fused to generate the health index of the target object.

[0054] Obtain the total number of target individuals and the number of consultation rooms, and triage the target individuals according to the health index to obtain an initial allocation plan;

[0055] A multi-objective function is constructed. Time cost is calculated based on the waiting time of the target object. Treatment matching degree and resource utilization are calculated by combining the doctor's specialty and consultation rate. Based on the initial allocation scheme, multiple adjustment schemes satisfying the multi-objective constraints are generated, and their state transition probabilities are calculated. Combining the multi-objective function, the search direction is updated using a roulette wheel approach. The allocation scheme is adjusted through adaptive step size, and this process is repeated iteratively until the multi-objective function value is less than the optimization threshold. The optimal allocation scheme is then output.

[0056] Based on the optimal allocation scheme, a triage number and an estimated appointment time are generated, and the triage results are sent to the target's terminal device.

[0057] In one optional implementation, the multidimensional physiological data is decomposed in the time-frequency domain to extract time-domain feature sequences and frequency feature spectra; non-stationary feature vectors are extracted from the time-domain feature sequences, and cluster analysis is performed on the frequency feature spectra to form feature combinations, including:

[0058] Empirical mode decomposition is performed on the multidimensional physiological data to obtain multiple intrinsic mode functions (IMFs). Temporal reconstruction is then performed on the multiple IMFs to obtain a temporal feature sequence. Instantaneous frequency features are extracted from the multidimensional physiological data using Hilbert transform to obtain a frequency feature spectrum.

[0059] 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; the time-domain feature sequence is convolved with wavelet basis functions of different scales to obtain detail coefficients at multiple scales; the squares of the detail coefficients are weighted and summed, and then multiplied by the local energy density to obtain non-stationary features;

[0060] 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 a preset learning rate to obtain multiple feature combinations and their cluster centers. When the distance between the cluster centers is greater than the clustering threshold, the current feature combination is taken as the optimal feature combination.

[0061] This invention provides a multidimensional physiological data processing method that achieves effective processing of physiological data through time-frequency domain decomposition and feature extraction analysis. Specifically, this method acquires physiological data encompassing multiple dimensions such as electrocardiogram, respiration, and blood pressure, which typically exhibit complex nonlinear and non-stationary characteristics.

[0062] Empirical Mode Decomposition (EMD) is performed on the acquired multidimensional physiological data to decompose complex physiological signals into multiple intrinsic mode functions (EMFs). Taking electrocardiogram (ECG) signals as an example, ECG signals acquired at a frequency of 1000 Hz can yield 8-12 EMFs after EMD, each representing a signal component at a different frequency scale. Temporal reconstruction is then performed on these EMFs. Specifically, the EMFs containing the main physiological information (usually the 3rd to 7th functions) are selected, and a time-domain feature sequence is reconstructed through weighted superposition. The weight coefficients are determined based on the energy contribution of each EMF. Typically, EMFs with higher energy contributions are assigned larger weights; for example, the weight of the 4th mode function can be set to 0.35, the 5th to 0.25, and others adjusted accordingly.

[0063] Instantaneous frequency features are extracted from multidimensional physiological data using Hilbert transform. Specifically, Hilbert transform is performed on each dimension of the physiological data to obtain an analytical representation of the signal, from which instantaneous amplitude and phase information are extracted. The instantaneous frequency is calculated using the time derivative of the instantaneous phase, forming a frequency characteristic spectrum. For example, after processing heart rate variability signals, a frequency characteristic spectrum in the range of 0.04Hz-0.4Hz can be obtained, reflecting the activity of the sympathetic and parasympathetic nervous systems.

[0064] 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 and the sliding step size is 0.5 seconds. For each window position, the sum of 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, with a heart rate of 75 beats / minute, a 2-second window contains approximately 2.5 cardiac cycles, which can effectively capture the local variation characteristics of the electrocardiogram signal.

[0065] To accurately extract non-stationary features from a time-domain feature sequence, it is necessary to select an appropriate wavelet basis function. The choice of wavelet basis function is crucial for capturing specific features in the signal. In this embodiment, various wavelet basis functions can be used, such as the Haar wavelet, Daubechies wavelet, Coiflet wavelet, or Symlet wavelet. Based on the characteristics of the expiratory 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.

[0066] The selected wavelet basis functions are expanded and translated at different scales to form a set of wavelet basis functions at different scales. In practical applications, five different scale levels can be selected, each corresponding to the analysis of different frequency components in the signal. For example, for a 100Hz sampled exhalation signal, these five scales can correspond to frequency components of approximately 50Hz, 25Hz, 12.5Hz, 6.25Hz, and 3.125Hz. For each scale level, the time-domain feature sequence is convolved with the wavelet basis function of the corresponding scale. The convolution operation essentially calculates the correlation between the signal and the wavelet function, thereby detecting the signal's variation 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 will produce approximately 997 output points, constituting the detail coefficient sequence at that scale. By performing convolution operations at five different scales, five sets of detail coefficient sequences are obtained. These detail coefficients reflect the variation characteristics of the original signal in different frequency ranges. The length of each set of detail coefficients is similar to that of the original signal, totaling five sets of approximately 997 data points. To quantify the energy distribution of the signal at each scale, the sequence of detail coefficients at each scale is squared. The squaring operation converts the detail coefficients into an energy representation, highlighting important variations in the signal. This step yields five energy sequences, each with approximately 997 data points.

[0067] Weighting coefficients are assigned to the energy sequences at each scale based on their importance. In breath analysis, mid-frequency components (such as scales 2 and 3) typically contain more physiological information and can therefore be assigned 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 have higher importance. The weighted energy sequences at each scale are summed to obtain a multi-scale energy representation. This step merges the five weighted energy sequences into a single comprehensive energy sequence of approximately 997 data points. The local energy density of the time-domain feature sequence is calculated. Local energy density reflects the energy concentration of the original signal within a short time window and can be calculated using the sliding window method. Specifically, a sliding window of length 20 (corresponding to 0.2 seconds at a sampling rate of 100Hz) is used, and the average of the sum of squares of the signal is calculated within the window to obtain the local energy density sequence. For an original signal of 1000 data points, this step yields approximately 981 local energy density values. Multiplying the multi-scale energy representation by the local energy density yields the non-stationary characteristics. This step integrates the time-frequency properties of the signal and the local energy distribution, effectively capturing the non-stationary components in the signal. The multiplication operation requires adjusting the two sequences to the same length, typically taking the shorter of the two, which in this example is approximately 981 data points.

[0068] To verify the effectiveness of the extracted non-stationary features, a typical set of expiratory samples was analyzed. This set of samples included expiratory data from 10 healthy subjects and 10 patients with respiratory diseases. The expiratory flow rate of the healthy subjects exhibited a relatively stable bell-shaped curve, with extracted non-stationary feature values ​​ranging from 0.05 to 0.15; while the expiratory flow rate curves of the patients showed significant irregular fluctuations, with extracted non-stationary feature values ​​ranging from 0.2 to 0.5, significantly higher than those of the healthy group. This indicates that the extracted non-stationary features can effectively distinguish expiratory patterns under different physiological states.

[0069] Non-stationary features are further processed to extract statistical indicators (mean, standard deviation, kurtosis, skewness, etc.) or time-domain features (peak position, duration, etc.) to form a higher-dimensional feature vector. For example, for each subject's non-stationary feature sequence, five statistical features are 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 value (healthy group: 0.15±0.04, patient group: 0.5±0.1), minimum value (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 constitute a 5-dimensional feature vector for subsequent health status assessment.

[0070] When performing clustering analysis on the frequency feature spectrum, a self-organizing map neural network method is used. The initial neighborhood radius is set to 4, decreasing linearly with each iteration. The decay formula is: current neighborhood radius = initial neighborhood radius × (1 - current iteration number / total iteration number). The initial learning rate is preset to 0.1, also decreasing with the number of iterations. In practice, the total number of iterations is set to 1000. When the 800th iteration is reached, the learning rate drops to approximately 0.02, and the neighborhood radius decreases to 0.8. Clustering of the frequency feature spectrum is performed based on the current neighborhood radius and learning rate, with cluster centers dynamically updated. 100 sets of sample data are selected for clustering. Initially, 5 cluster centers are randomly generated. After iterative adjustments, stable feature cluster combinations are formed. The Euclidean distance between cluster centers is calculated, and a clustering threshold of 0.15 is set. When the distance between all cluster centers is greater than this threshold, the current clustering result is considered the optimal feature combination.

[0071] In a practical application, this method was used to process multidimensional physiological signals containing 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 main feature clusters were obtained. The final feature combination can effectively distinguish different physiological states.

[0072] In one optional implementation, 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 fused to generate a health index for the target object, including:

[0073] For each expert subnetwork, 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 then compensated for errors using a residual learning mechanism to obtain residual correction terms. The preliminary prediction results and residual correction terms are then weighted and fused to obtain the corrected prediction values ​​for each expert subnetwork.

[0074] The corrected prediction values ​​of each expert subnetwork are input into the gated loop unit according to a preset time step. By controlling the forgetting and updating of historical information, a time series prediction result is generated. The prediction error between the time series prediction result and the true value is calculated. The fusion weight is assigned to each expert subnetwork through iterative optimization. The iteration stops when the error change in multiple consecutive iterations is less than the error threshold, and the optimal fusion weight is obtained.

[0075] The optimal fusion weights are weighted and combined with the corrected prediction values ​​of each expert subnetwork to obtain the health index of the target object.

[0076] This embodiment describes in detail a method for generating a health index for a target object by inputting non-stationary features and feature combinations into a pre-constructed hybrid expert model. This method achieves an accurate assessment of the target object's health status by fusing the prediction results of multiple expert sub-networks.

[0077] Feature extraction and preprocessing are performed on the collected raw data of the target object to obtain non-stationary features and feature combinations. These features are then fed into a pre-built hybrid expert model. This hybrid expert model contains multiple parallel expert sub-networks, each focused on capturing different patterns and features in the data.

[0078] For each expert subnetwork, 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 convolutional kernels to extract features from the input data, then applies the ReLU activation function to increase non-linearity, and finally reduces the feature dimensionality through a 2×2 max pooling layer. The second convolutional layer uses 64 3×3 convolutional kernels to further extract high-level features, also applying the ReLU activation function and a 2×2 max pooling layer. Through this hierarchical convolution and pooling operation, local and global features of the input data can be effectively extracted, generating preliminary prediction results.

[0079] To improve prediction accuracy, a residual learning mechanism is employed to compensate for errors in the initial prediction results. This mechanism is implemented through a bypass branch 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. This bypass branch learns the residual between the predicted and true values, generating a residual correction term. The initial prediction and the residual correction term are then weighted and fused, with weight coefficients set to 1 and 1.5, to obtain the corrected prediction values ​​for each expert subnetwork. Specifically, when the initial health index prediction for a target object is 80.5, and the residual correction term is 1.8, the corrected health index will be 83.2, closer to the true value. For a given target object, the corrected prediction values ​​for the three expert subnetworks are 83.2, 81.9, and 82.5, respectively.

[0080] The corrected predictions from each expert subnetwork are input into a gated recurrent unit (GRU) containing 64 hidden units at preset time steps (e.g., 10 time points). The GRU controls the forgetting and updating of historical information through update and reset gates. The update gate determines how much previous state information is retained, and the reset gate determines how new input information is combined with previous states. This mechanism enables the model to capture long-term dependencies in time series data and generate time series predictions that take historical trends into account.

[0081] The prediction error between the time-series prediction results and the actual values ​​is calculated, and the mean squared error is used as the evaluation metric. Fusion weights are assigned to each expert subnetwork through an iterative optimization process. During the iteration, the system uses the Adam optimization algorithm with a learning rate of 0.001 and a batch size of 32. The iteration stops when the error change over five consecutive iterations is less than the preset error threshold of 0.0001, yielding the optimal fusion weights. For example, for the three expert subnetworks mentioned above, the optimal fusion weights obtained after iterative optimization are 0.4, 0.25, and 0.35, respectively. The optimal fusion weights are then weighted and combined with the corrected prediction values ​​of each expert subnetwork to obtain the health index of the target object. Specifically, the calculation is: 0.4 × 83.2 + 0.25 × 81.9 + 0.35 × 82.5 = 82.66, meaning the health index of the target object is 82.66.

[0082] By fusing the prediction results of multiple expert subnetworks and combining residual learning and gated recurrent units, this method can effectively handle the non-stationary nature and temporal dependencies of data, providing a more accurate health assessment for the target object.

[0083] In one alternative implementation, a multi-objective function is constructed to calculate the time cost based on the waiting time of the target object, and to calculate the treatment matching degree and resource utilization rate by combining the doctor's expertise and the consultation rate, including:

[0084] The number of people queuing in a department, the number of doctors on duty, and the service capacity per unit time are obtained. The basic waiting time is obtained by dividing the number of people queuing in a department by the product of the number of doctors on duty and the service capacity per unit time. The time cost is obtained by substituting the basic waiting time into an exponential decay function.

[0085] Calculate the semantic similarity between keywords of the target subject's self-reported symptoms and keywords of the doctor's area of ​​expertise; obtain the average consultation time of the doctor's historical treatment data, and use the ratio of the average consultation time to the standard consultation time as the empirical matching degree; and obtain the treatment matching degree by weighted summation of the semantic similarity and the empirical matching degree.

[0086] The consultation rate is obtained by dividing the number of patients already seen by the doctor by the maximum number of patients a doctor can see. The average value of the consultation rate is taken 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.

[0087] This embodiment provides a medical resource allocation system based on multi-objective queuing optimization. A multi-objective function is constructed that comprehensively considers waiting time cost, treatment matching degree, and resource utilization rate to achieve the rational allocation of medical resources.

[0088] Obtain the current number of people in the queue, the number of doctors on duty, and the service capacity per doctor per unit time for each department. Taking the respiratory department as an example, the current number of people in the queue 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 by dividing the number of people in the queue by the product of the number of doctors on duty and the service capacity per unit time, i.e., 50 ÷ (5 × 3) = 3.33 hours.

[0089] To more accurately reflect the non-linear impact of waiting time on patient satisfaction, an exponential decay function is used to calculate the time cost. When the baseline waiting time is t, the time cost can be expressed as 1 - e^(-t / t). (-αt) Where α is the attenuation coefficient, set to 0.5 based on the actual situation of the hospital. Taking the respiratory department 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, representing the patient's dissatisfaction caused by waiting.

[0090] To calculate the matching degree of diagnosis and treatment, it is necessary to obtain the self-reported symptoms of the target subjects. Symptom descriptions of the target subjects are collected through a digital consultation form. The consultation form design includes open-ended questions related to gastrointestinal symptoms, such as "Please describe your stomach discomfort symptoms," as well as structured questions, such as "Do you experience acid reflux?", "Do you experience abdominal pain?", and "Duration of symptoms?". The target subjects fill in this information through the interactive terminal of the breath test site. For example, a target subject might describe: "For the past month, I have frequently experienced dull pain in my upper abdomen, which worsens after meals. I occasionally experience acid reflux and heartburn. The pain at night leads to poor sleep. I have tried taking antacids, but the effect was not significant." The collected self-reported symptom text undergoes preprocessing. Preprocessing includes text segmentation, stop word removal, and standardization. The segmentation process uses a medically optimized segmentation algorithm to divide continuous text into independent words or phrases. For example, the above symptom descriptions can be segmented into words such as "one month", "upper abdomen", "dull pain", "after meals", "symptoms worsening", "acid reflux", "heartburn", "at night", "pain", "poor sleep", "antacids", and "not effective". The stop word removal step filters out commonly used words that contribute little to symptom identification, such as "of", "and", and "will". Standardization unifies the same symptoms expressed differently into standard medical terminology, such as standardizing "heartburn" to "heartburn" and "dull pain in the upper abdomen" to "upper abdominal pain".

[0091] Keywords related to symptoms potentially associated with Helicobacter pylori infection were extracted from the preprocessed text. A symptom dictionary constructed using a gastrointestinal disease knowledge base, containing approximately 150 common gastrointestinal symptom words and phrases, was used, with particular attention to symptoms associated with Helicobacter pylori infection such as "upper abdominal pain," "heartburn," "acid reflux," "bloating," "indigestion," "nausea," "vomiting," and "loss of appetite." These keywords were identified from the self-reported descriptions of the target subjects. For the example above, the extracted keywords included "upper abdominal pain," "worse after meals," "acid reflux," "heartburn," "night pain," and "ineffective use of antacids." Weights were assigned to each extracted keyword to reflect its importance in the diagnosis of Helicobacter pylori infection and the treatment of related gastrointestinal diseases. The weighting was based on several factors: the relevance of the symptom to Helicobacter pylori infection (symptoms with high relevance have higher weights), the typicality of the symptom (typical symptoms have higher weights), the duration of the symptom (symptoms with longer duration have higher weights), and the frequency of the symptom (symptoms with higher frequency have higher weights). For example, "upper abdominal pain" as a typical symptom of Helicobacter 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 "ineffective use of antacids" a weight of 0.65.

[0092] A specialized database of physicians' expertise in digestive system diseases will be constructed, recording each gastroenterologist's area of ​​expertise, research direction, clinical experience, and diseases they specialize in treating, with particular emphasis on their experience in diagnosing and treating Helicobacter pylori-related diseases. This information will be provided by the hospital's gastroenterology department and updated regularly. Professional keywords will be extracted from physician profiles, such as "Helicobacter pylori infection," "peptic ulcer," "chronic gastritis," and "reflux esophagitis." For example, Physician A's professional keywords might include "Helicobacter pylori," "chronic gastritis," "upper abdominal pain," and "breath test"; Physician B's keywords might include "peptic ulcer," "acid reflux," "heartburn," and "antibiotic treatment"; and Physician C's keywords might include "gastroesophageal reflux," "gastrointestinal motility," "functional dyspepsia," and "abdominal distension." Weights will be assigned to these professional keywords to reflect the degree to which they represent a physician's professional expertise, especially in the diagnosis and treatment of Helicobacter pylori-related diseases. The weighting is based on the physician's experience in Helicobacter pylori research (more experience, higher weight), the number of related publications (more publications, higher weight), the number of successfully treated Helicobacter pylori infection cases (more cases, higher weight), and the accuracy of breath test interpretation (higher accuracy, higher weight). For example, for Physician A, who has 8 years of experience in Helicobacter pylori research, the keyword "Helicobacter pylori" has a weight of 0.95, "chronic gastritis" has a weight of 0.9, "upper abdominal pain" has a weight of 0.85, and "breath test" has a weight of 0.92.

[0093] To calculate semantic similarity, a word vector model specifically trained for digestive system diseases was used to convert symptom keywords and physician-specific keywords into high-dimensional vectors. This word vector model was trained on a large-scale corpus of digestive system disease literature, containing approximately 500,000 related terms, each mapped to 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, with "peptic ulcer" 0.78, and with "gastroesophageal reflux" 0.45, indicating that "upper abdominal pain" is semantically closest to "Helicobacter pylori". The cosine similarity of each pair of symptom keywords and doctor's professional keywords is multiplied by the corresponding weight. Then, the weighted similarity of all pairs is summed, and finally, the result is normalized by dividing by the sum of the weights. For example, for the aforementioned target object 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; and so on, calculating the weighted similarity of all pairs, summing them, and dividing by the sum of the weighted products to obtain a final semantic similarity of 0.78 between the target object and Doctor A. Semantic similarity is then calculated between the target object and all available gastroenterologists, and the results are ranked according to the similarity score. 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 closely matched with Doctor A's expertise, especially in the diagnosis and treatment of Helicobacter pylori.

[0094] To further improve the matching accuracy related to Helicobacter pylori breath tests, semantic understanding related to the breath test procedure has been enhanced. When the system identifies key information such as "have undergone a breath test," "have taken thiol urea," and "waiting time for testing" in the target's description, it prioritizes matching with doctors experienced in breath test interpretation and subsequent treatment. For example, if the target mentions "have previously undergone a breath test but the result is uncertain" or "tested 30 minutes after taking thiol urea tablets," the system will increase the semantic weight associated with keywords such as "breath test," "thiol urea," and "Helicobacter pylori testing."

[0095] Considering the doctor's historical 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, then the empirical matching degree is 15 ÷ 20 = 0.75, indicating that the doctor's treatment efficiency is relatively high. The semantic similarity and empirical matching degree 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 suitability of the doctor-patient match.

[0096] Regarding resource utilization, the system calculates a doctor's current consultation rate, which is the number of patients already seen divided by the maximum number of patients they can see. For example, Dr. A's maximum daily consultation capacity is 30, and he has already seen 12 patients, so his consultation rate is 12 ÷ 30 = 0.4. The average consultation rate of all doctors in the department is then calculated as the overall resource utilization rate. For instance, if the consultation rates of the 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.

[0097] A multi-objective optimization function is constructed by weighting and summing time cost, treatment matching degree, and resource utilization rate. With weights set as w1=0.4, w2=0.4, and w3=0.2, the multi-objective function value is calculated as (1-0.81)×0.4+0.81×0.4+0.44×0.2=0.464. The time cost is converted to a time efficiency index by subtracting the original time cost value from 1. Based on the multi-objective function value, the optimal doctor selection is recommended for each queued patient.

[0098] In implementation, a distributed computing architecture is used to handle large-scale concurrent requests. Patient data and doctor information are stored in a distributed database, and a caching mechanism improves data access speed. The semantic similarity calculation module employs a pre-computation strategy, storing the similarity between common symptoms and professional keywords in a lookup table to reduce real-time computation overhead. Queue status and resource utilization are updated every 5 minutes to ensure decisions are based on the latest data. Furthermore, an adaptive weight adjustment mechanism is designed. By analyzing historical queuing data and patient satisfaction feedback, the weights of time cost, treatment matching degree, and resource utilization are dynamically adjusted. For example, during peak hours, the weight of resource utilization is increased to improve overall throughput; during off-peak hours, the weight of treatment matching degree is increased to improve the quality of medical care.

[0099] Figure 2This is a system architecture diagram of an adaptive allocation scheme optimization method based on multi-objective constraints. In one optional implementation, based on the initial allocation scheme, multiple adjustment schemes satisfying 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 approach, and the allocation scheme is adjusted through an adaptive step size. This process is repeated iteratively until the multi-objective function value is less than the optimization threshold, and the optimal allocation scheme is output, including:

[0100] Obtain the initial allocation scheme for the target objects. For each target object, generate multiple adjustment paths under the condition of satisfying multi-objective constraints, and calculate the multi-objective function value on each adjustment path. Calculate the product of the initial pheromone concentration of each adjustment path and the reciprocal of the multi-objective function value to obtain the state transition probability.

[0101] Calculate the multi-objective function difference between the current adjustment path and the initial allocation scheme, and dynamically update the search direction on the current adjustment path using a roulette wheel approach, combined with the state transition probability. Determine the adaptive step size by multiplying the maximum adjustment step size by the power of the complement of the iteration progress and the multi-objective function difference. Adjust the allocation scheme of the target object according to the updated search direction and the adaptive step size.

[0102] 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.

[0103] This invention provides an adaptive allocation scheme optimization method based on multi-objective constraints. This method, while satisfying multi-objective constraints, updates the search direction through adaptive step size adjustment and roulette wheel betting, ultimately outputting the optimal allocation scheme. The specific implementation steps of this method will be described in detail below.

[0104] When implementing this method, an initial allocation plan for the target subjects needs to be obtained. This initial allocation plan is generated based on the target subjects' health index; a higher health index indicates a worse health condition and a higher priority. For example, at a breath test site, there are 20 target subjects waiting for 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 5 available examination rooms, numbered 1 to 5. The target subjects are initially ranked according to their health indices, with those having higher health indices receiving higher priority. The sorted sequence of target objects 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], with corresponding target object numbers of [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 allocated to each consultation room, considering the number of consultation rooms and the number of target objects. In this example, an average of 4 target objects are allocated to each of the 5 consultation rooms. The system assigns target object number 12 to clinic 1, target object number 19 to clinic 2, and so on, according to priority. The initial allocation scheme 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].

[0105] For each target object, multiple possible adjustment paths are generated. An adjustment path represents a possible scheme to move the target object from the current clinic to another clinic. For example, for target object number 12 initially assigned to clinic 1, possible adjustment paths include: moving to clinic 2, moving to clinic 3, moving to clinic 4, and moving 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, for a total of 20 target objects and 80 adjustment paths. Each adjustment path is checked to ensure it meets multi-objective constraints. For the path of target object number 12 moving from clinic 1 to clinic 2, it is checked whether the number of target objects in clinic 2 after adjustment meets the resource utilization constraint, whether the matching degree between the target object and the doctor in clinic 2 meets the treatment matching constraint, and 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.

[0106] Calculate the multi-objective function value for each valid adjustment path. The calculation method is the same as evaluating the initial allocation scheme, but based on the adjusted allocation scheme. For example, after moving target object 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 object; for example, target object number 12 is second in clinic 2, with a waiting time of 15 minutes. Recalculate the overall treatment matching degree; for example, target object number 12 has a matching degree of 0.82 with clinic 2, which is higher than the matching degree of 0.45 with clinic 1. Recalculate the resource utilization rate; for example, after the adjustment, clinic 1 has 3 target objects, and clinic 2 has 5 target objects, resulting in a slight decrease in resource utilization.

[0107] The state transition probability is obtained by multiplying the initial pheromone concentration of each adjustment path by the reciprocal of the multi-objective function value. Assuming 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 (lower values ​​are better), 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 between the multi-objective function of the current adjustment path and the initial allocation scheme is calculated. Taking path C as an example, the multi-objective function value of the initial scheme 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 adjustment path using a roulette wheel approach. In practice, a random number between 0 and 1 is generated, such as 0.36, which falls within the probability interval of the search direction numbered 6 on path C. Therefore, the search direction numbered 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 of the complement of the iteration progress and the difference of the multi-objective function. Assuming the maximum adjustment step size is 1.0, and the current iteration is the first iteration (10 iterations are planned), then the complement of the iteration progress is (1-0.1)=0.9, and the power of the complement is 2, i.e., 0.81. The adaptive step size is calculated as 1.0×0.81×0.37=0.3, representing the adjustment magnitude on path C.

[0108] The allocation scheme of target objects is adjusted based on the updated search direction and adaptive step size. During the adjustment process, the mutual influence of multiple target objects is considered simultaneously. For example, when the step size is 3, three target objects are selected for adjustment: ID-12 is moved from clinic 1 to clinic 2, ID-8 from clinic 4 to clinic 3, and ID-15 from clinic 1 to clinic 5. The multi-objective function value after these adjustments is 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 scheme is found or the maximum number of attempts is reached. A strategy combining greedy search and simulated annealing is used to select the adjustment scheme. In most cases, the adjustment scheme that improves the multi-objective function value is selected; however, there is a certain probability that an adjustment scheme that does not improve or even slightly worsens the problem will be accepted to escape local optima. 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 scheme causes the multi-objective function value to deteriorate from 2.34 to 2.36, with a deterioration of 0.02, the system accepts the scheme with a probability of 0.15. As the number of iterations increases, the acceptance probability gradually decreases, and the optimization process gradually converges.

[0109] Update the waiting time for each target object after each adjustment. The waiting time depends on the target object's ranking position in the clinic and the estimated consultation time of the target objects preceding it. Estimate the consultation time for each target object based on historical data. For example, target object ID-19 with a health index of 0.82 has an estimated consultation time of 20 minutes, and target object ID-3 with a health index of 0.56 has an estimated consultation time of 15 minutes. After adjustment, if target object ID-12 is ranked after ID-19, its waiting time will be 20 minutes. Recalculate the treatment matching degree after each adjustment. For example, target object ID-12's symptom feature vector is [0.8, 0.6, 0.3, 0.7, 0.2], and clinic 2 doctor's specialty vector is [0.75, 0.65, 0.4, 0.8, 0.3], resulting in a matching degree of 0.82. Update the treatment matching degree for all affected target objects after adjustment. Recalculate resource utilization after each adjustment. For example, before adjustment, the number of target subjects in each clinic was [4, 4, 4, 4, 4], with a standard deviation of 0, resulting in the highest resource utilization rate. After adjustment, it became [3, 5, 5, 3, 4], with a standard deviation of approximately 0.89, indicating a slight decrease in resource utilization. The multi-objective function value is calculated based on the updated waiting time, treatment matching degree, and resource utilization rate. An optimization threshold is set as the iteration termination condition. The optimization threshold is the expected minimum value of the multi-objective function, representing 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 below 1.5, the allocation scheme is considered sufficiently optimized. Further requirements can be made that the individual multi-objective function values ​​of all target subjects are less than the optimization threshold to ensure the balance of the allocation scheme.

[0110] The system iterates through the processes of updating the search direction, adjusting the step size, adjusting the allocation scheme, and calculating the multi-objective function values. In each iteration, the system records the current allocation scheme and the corresponding multi-objective function values. For example, the allocation scheme 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], with a corresponding multi-objective function value of 2.25; the allocation scheme 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], with a corresponding multi-objective function value of 1.78. Determine whether the current iteration meets the termination condition. Termination conditions include: all 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 of less than 0.001 for 30 consecutive iterations; or the maximum number of iterations is reached, such as 500 iterations. If any condition is met, the system stops iterating; otherwise, it continues to the next iteration. For example, after the 217th iteration, if all individual multi-objective function values ​​for all target objects are below the optimization threshold of 1.5, the system terminates the iteration process. The allocation scheme with the smallest multi-objective function value from all recorded allocation schemes is output as the optimal allocation scheme. For example, the allocation scheme in the 198th iteration: 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], has a corresponding multi-objective function value of 1.42, which is the minimum value among all iterations, and therefore it is selected as the optimal allocation scheme.

[0111] The optimal allocation scheme is translated into specific triage instructions, including the target object ID, assigned examination room, sorting position within the examination room, and estimated waiting time. For example, target object ID-12 is assigned to examination room 2, sorting position 2, with an estimated waiting time of 20 minutes. The triage instructions are sent to the corresponding target object and medical staff via the display terminal or mobile application at the breath test site, completing the entire intelligent triage process.

[0112] In practical applications, the weights and constraints of the multi-objective function can be adjusted according to different scenarios. For example, for latency-sensitive applications, the weight of service quality 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 scheme that satisfies multi-objective constraints in complex and ever-changing environments.

[0113] In one optional implementation, the multi-objective function difference between the current adjustment path and the initial allocation scheme is calculated, and the search direction on the current adjustment path is dynamically updated using a roulette wheel approach, taking into account the state transition probability.

[0114] Obtain the current adjustment path and its state transition probability, calculate the complement of the ratio of the current iteration number to the maximum iteration number, and add the product of the complement and the power of the state transition probability to the probability threshold to obtain the adaptive selection probability;

[0115] Calculate the multi-objective function difference between the current adjustment path and the initial allocation scheme. When the multi-objective function difference is less than the adaptive selection probability, select the search direction with the highest local pheromone concentration on the adjustment path.

[0116] When the difference in the multi-objective function is greater than the adaptive selection probability, the complement of the pheromone evaporation 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 probability is 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.

[0117] This invention provides a resource scheduling method based on ant colony optimization algorithm, which improves scheduling efficiency by dynamically updating the search direction. The technical content of implementing this invention is described in detail below.

[0118] Obtain the current adjustment path and its state transition probability. Assume the current adjustment path is P, and the state transition probability is p. trans The current iteration number is iter current The maximum number of iterations is iter max The system calculates the ratio of the current iteration count to the maximum iteration count as ratio = iter current / iter max Then calculate the complement of that ratio, comp. ratio = 1 - ratio. Add the product of the complement and the power of the state transition probability to the preset probability threshold to obtain the adaptive selection probability p. adaptive = comp ratio ^p trans + threshold. For example, if the current iteration count is 50, the maximum iteration count 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 daptiveThe result is 0.75^0.7 + 0.1 = 0.72.

[0119] Calculate the multi-objective function difference (diff) between the current adjusted path P and the initial allocation scheme P0. Assume the multi-objective function value of the initial allocation scheme is f(P0) = 2.3, and the multi-objective function value of the current adjusted path is f(P) = 2.1. Then the multi-objective function difference is diff = |f(P) - f(P0)| = |78 - 85| = 0.2. Determine whether the multi-objective function difference norm is less than the adaptive selection probability p. adaptive In the example above, since 0.2 < 0.72, the search direction with the highest local pheromone concentration on the adjustment path is selected. Assuming there are three possible 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 highest local pheromone concentration as the final search direction.

[0120] When the difference in the standardized multi-objective function is greater than the adaptive selection probability, a roulette wheel-based method is used to select the search direction. In another case, the difference in the multi-objective function is 0.85, and the adaptive selection probability is 0.72. Since 0.85 > 0.72, further calculation of the selection probability for each search direction is required. The pheromone evaporation coefficient rho of the current adjustment path is obtained; for example, rho = 0.3, its complement comp is calculated. rho =1-rho=0.7. Multiply this complement by the local pheromone concentration in each search direction on the current adjustment path to obtain the original selection probability of the three search directions: prob D1 =0.7 × 0.45 = 0.315, prob D2 =0.7 × 0.35 = 0.245, prob D3 =0.7 × 0.2 = 0.14. Normalizing these original selection probabilities yields the current selection probability for 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. Following the numbering order of the search directions, the system accumulates the current selection probability for each direction to obtain the cumulative selection probability: cum probD1 =0.45, cum probD2 =0.45+0.35=0.8, cum probD3=0.45 + 0.35 + 0.2 = 1. Generate a random number rand between 0 and 1, for example, rand = 0.6. Calculate the difference between this random number and the cumulative selection probability for each search direction: 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.

[0121] Through the above method, the system achieves a dynamic balance between global exploration and local exploitation. In the early stages of iteration, the adaptive selection probability is relatively high, and the system tends to choose the search direction based on local pheromone concentration, which is beneficial for global exploration. As the number of iterations increases, the adaptive selection probability decreases, and the system tends to choose the search direction through a roulette wheel approach, enhancing local exploitation capabilities. By dynamically adjusting the search strategy, the method of this invention can achieve a good balance between global exploration and local exploitation, avoiding getting trapped in local optima while improving convergence speed.

[0122] This invention provides an embodiment of a deep learning-based expiratory breath test triage and queuing optimization system, comprising:

[0123] The first unit is used to collect multidimensional physiological data of the target subjects from the breath test site;

[0124] The second unit is used to perform time-frequency domain decomposition on the multidimensional physiological data, extract time-domain feature sequences and frequency feature spectra; extract non-stationary feature vectors from the time-domain feature sequences, perform cluster analysis on the frequency feature spectra to form feature combinations; input the non-stationary features and the feature combinations into a pre-constructed hybrid expert model, and fuse the prediction results of multiple expert sub-networks to generate the health index of the target object;

[0125] The third unit is used to obtain the total number of target objects and the number of clinics, and to triage the target objects according to the health index to obtain an initial allocation plan;

[0126] The fourth unit is used to construct a multi-objective function. It calculates the time cost based on the waiting time of the target object, and calculates the treatment matching degree and resource utilization rate by combining the doctor's expertise and consultation rate. Based on the initial allocation scheme, it generates multiple adjustment schemes that satisfy the multi-objective constraints and calculates their state transition probabilities. Combining the multi-objective function, it updates the search direction using a roulette wheel approach, adjusts the allocation scheme through adaptive step size, and iterates repeatedly until the multi-objective function value is less than the optimization threshold, outputting the optimal allocation scheme.

[0127] The fifth unit is used to generate a triage number and an estimated appointment time according to the optimal allocation scheme, and send the triage results to the target's terminal device.

[0128] A third aspect of the present invention provides an electronic device, comprising:

[0129] processor;

[0130] Memory used to store processor-executable instructions;

[0131] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0132] 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.

[0133] 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.

[0134] 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 method for breath test triage and queue optimization based on deep learning, characterized in that, The method comprises the following steps: Collecting multi-dimensional physiological data of a target object from an exhalation test site; Performing time-frequency domain decomposition on the multi-dimensional physiological data to extract time domain feature sequences and frequency feature spectra; extracting non-stationary feature vectors in the time domain feature sequences, and performing cluster analysis on the frequency feature spectra to form feature combinations; inputting the non-stationary features and the feature combinations into a pre-constructed hybrid expert model to fuse the prediction results of multiple expert sub-networks to generate a health index of the target object; Obtaining the total number of target objects and the number of clinics, and performing triage on the target objects according to the health index to obtain an initial allocation scheme; Constructing a multi-objective function, calculating the time cost based on the waiting time of the target objects, calculating the diagnosis and treatment matching degree and resource utilization rate in combination with the professional expertise of doctors and the visit rate; based on the initial allocation scheme, generating multiple adjustment schemes that satisfy multi-objective constraints and calculating their state transition probabilities; combining the multi-objective function, updating the search direction in a roulette wheel manner, adjusting the allocation scheme through an adaptive step size, and repeating iteration until the multi-objective function value is less than an optimization threshold, and outputting an optimal allocation scheme; Generating a triage number and an estimated visit time according to the optimal allocation scheme, and sending the triage result to a terminal device of the target object.

2. The method of claim 1, wherein, Performing time-frequency domain decomposition on the multi-dimensional physiological data to extract time domain feature sequences and frequency feature spectra; extracting non-stationary feature vectors in the time domain feature sequences, and performing cluster analysis on the frequency feature spectra to form feature combinations, comprising: Performing empirical mode decomposition on the multi-dimensional physiological data to obtain multiple intrinsic mode functions, and performing time series reconstruction on the multiple intrinsic mode functions to obtain time domain feature sequences; extracting instantaneous frequency features of the multi-dimensional physiological data through Hilbert transform to obtain frequency feature spectra; Based on a preset time window, performing sliding analysis on the time domain feature sequences to calculate the local energy density within the time window; performing convolution operation on the time domain feature sequences and wavelet basis functions of different scales to obtain detail coefficients of multiple scales, performing weighted summation on the squares of the detail coefficients, and multiplying the local energy density to obtain non-stationary features; Setting an initial neighborhood radius, linearly reducing the initial neighborhood radius as the number of iterations increases, clustering the frequency feature spectra based on the current neighborhood radius and a preset learning rate to obtain multiple feature combinations and their cluster centers, and when the distances between the cluster centers are all greater than a clustering threshold, taking the current feature combination as an optimal feature combination.

3. The method of claim 1, wherein, Inputting the non-stationary features and the feature combinations into a pre-constructed hybrid expert model to fuse the prediction results of multiple expert sub-networks to generate a health index of the target object, comprising: For each expert sub-network, performing multi-layer convolution and pooling operations on the non-stationary features and the feature combinations to extract hierarchical features and generate preliminary prediction results, using a residual learning mechanism to compensate for errors in the preliminary prediction results to obtain a residual correction term, and performing weighted fusion on the preliminary prediction results and the residual correction term to obtain a corrected prediction value of each expert sub-network; The correction prediction values of the expert sub-networks are input into a preset time step input gate cycle unit to generate a time sequence prediction result by controlling the forgetting and updating of historical information; a prediction error between the time sequence prediction result and a true value is calculated, and a fusion weight is assigned to each expert sub-network through iterative optimization, and the iteration is stopped when the error change of continuous multiple rounds of iteration is less than an error threshold, and an optimal fusion weight is obtained; The optimal fusion weight is combined with the correction prediction values of the expert sub-networks to obtain a health index of the target object.

4. The method of claim 1, wherein, A multi-objective function is constructed, and a time cost is calculated based on the waiting time of the target object, and a diagnosis and treatment matching degree and a resource utilization rate are calculated in combination with the professional expertise of the doctor and the treatment rate, including: The number of queued departments, the number of on-duty doctors, and the service capacity per unit time are obtained, the basic waiting time is obtained by dividing the number of queued departments by the product of the number of on-duty doctors and the service capacity per unit time, and the time cost is obtained by inputting the basic waiting time into an exponential decay function; The semantic similarity between the key words of the symptoms of the target object and the key words of the professional expertise of the doctor is calculated, the average diagnosis and treatment time of the historical diagnosis and treatment data of the doctor is obtained, and the ratio of the average diagnosis and treatment time to the standard diagnosis and treatment time is taken as the experience matching degree; and the diagnosis and treatment matching degree is obtained by weighted sum of the semantic similarity and the experience matching degree. The treatment rate is obtained by dividing the number of patients already treated by the doctor by the maximum number of patients treated by the doctor, and the average value of the treatment rate is taken as the resource utilization rate; and the multi-objective function is obtained by weighted sum of the time cost, the diagnosis and treatment matching degree, and the resource utilization rate.

5. The method of claim 1, wherein, Based on the initial allocation scheme, a plurality of adjustment schemes satisfying the multi-objective constraints are generated, and the state transition probability thereof is calculated; in combination with the multi-objective function, the search direction is updated in a roulette wheel manner, the allocation scheme is adjusted by an adaptive step size, and the iteration is repeated until the multi-objective function value is less than an optimization threshold, and the optimal allocation scheme is output, including: An initial allocation scheme of a target object is obtained, and for each target object, a plurality of adjustment paths are generated under the condition of satisfying the multi-objective constraints, and the multi-objective function value on each adjustment path is calculated; the product of the initial pheromone concentration of each adjustment path and the reciprocal of the multi-objective function value is calculated to obtain the state transition probability; The multi-objective function difference between the current adjustment path and the initial allocation scheme is calculated, the search direction on the current adjustment path is dynamically updated in a roulette wheel manner in combination with the state transition probability, and the allocation scheme of the target object is adjusted according to the updated search direction and the adaptive step size; The above iteration process is repeated until the multi-objective function value of all target objects under the current allocation scheme is less than the optimization threshold, and the current allocation scheme is taken as the optimal allocation scheme and output.

6. The method of claim 5, wherein, The multi-objective function difference between the current adjustment path and the initial allocation scheme is calculated, the search direction on the current adjustment path is dynamically updated in a roulette wheel manner in combination with the state transition probability, including: The current adjustment path and its state transition probability are acquired, a complement of the ratio of the current iteration number to the maximum iteration number is calculated, a product of the complement and a power of the state transition probability is added to a probability threshold to obtain an adaptive selection probability; A difference of a multi-objective function between the current adjustment path and the initial allocation scheme is calculated, and when the difference of the multi-objective function is less than the adaptive selection probability, a search direction with the maximum local pheromone concentration on the adjustment path is selected; When the difference of the multi-objective function is greater than the adaptive selection probability, a complement of the pheromone evaporation coefficient of the current adjustment path is multiplied by the local pheromone concentration of each search direction on the current adjustment path, and a normalization processing is performed to obtain a current selection probability of each search direction; the current selection probability is sequentially accumulated to obtain a cumulative selection probability of the search direction according to the number order of the search direction; a difference between a random number and the cumulative selection probability is calculated, and a search direction with the minimum difference is selected as a final search direction.

7. A deep learning-based breath test triage and queue optimization system for implementing the method of any one of claims 1-6, characterized in that, The method comprises: A first unit is configured to collect multi-dimensional physiological data of a target object from an exhalation test site; A second unit is configured to perform time-frequency domain decomposition on the multi-dimensional physiological data to extract a time domain feature sequence and a frequency domain feature spectrum; extract a non-stationary feature vector in the time domain feature sequence, perform clustering analysis on the frequency domain feature spectrum to form a feature combination; and input the non-stationary feature and the feature combination into a pre-constructed hybrid expert model to fuse prediction results of multiple expert sub-networks to generate a health index of the target object; A third unit is configured to obtain a total number of target objects and a number of clinics, and perform triage on the target objects according to the health index to obtain an initial allocation scheme; A fourth unit is configured to construct a multi-objective function, calculate a time cost based on a target object waiting time, calculate a diagnosis and treatment matching degree and a resource utilization rate in combination with a doctor's specialty and a visit rate; generate multiple adjustment schemes that satisfy multi-objective constraints based on the initial allocation scheme, and calculate state transition probabilities thereof; update search directions in a roulette wheel manner in combination with the multi-objective function, adjust the allocation scheme through an adaptive step, and repeat iteration until a multi-objective function value is less than an optimization threshold, and output an optimal allocation scheme; A fifth unit is configured to generate a triage number and a predicted visit time according to the optimal allocation scheme, and send a triage result to a terminal device of the target object.

8. An electronic device, comprising: The method comprises: A processor; A memory for storing processor-executable instructions; The processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 6.

Images